The guide follows the real job process of a traffic-management business – from the initial enquiry through to invoicing. It is intended as a working document, not as a technology overview.
Every figure-based example in this guide is labelled as a calculation model and rests on assumptions that are stated openly. None of them is a promise of the savings you will achieve. The regulations and deadlines cited reflect the position as of July 2026; they replace neither a review of the individual case nor legal advice.
Traffic safety is a business of high responsibility, tight time windows and many interfaces. Traffic control plans have to match the actual situation on site. Materials, vehicles and crews have to reach the location on time. Orders, conditions, inspection runs, setup records, photo documentation and proofs of performance all have to be complete – and, if it comes to it, still findable years later.
Today, much of this work is handled with a mix of email, phone calls, spreadsheets, paper forms, messenger chats and experience. That works as long as experienced staff are available and the number of parallel projects stays manageable. As the volume of jobs grows, however, so does the risk of lost information, duplicated effort, delays and undocumented deviations. The strain rarely falls on the crews; it falls almost always on the back office, on planning and on dispatching.
Artificial intelligence does not take over this work on its own. It replaces neither an official traffic order nor the professional responsibility of the named person accountable for traffic safety. What it can do is speed up large parts of the information processing that precedes that responsibility.
The greatest benefit does not come from a single AI chatbot. It comes when AI is connected to the job data, documents, maps, material stocks, deployment plans and mobile applications you already have. The sensible starting point is therefore not the question of which AI model to use, but a concrete bottleneck in the job process.
Those who miss this distinction buy tools and keep their problems. Those who grasp it change the information situation first – and then find that a considerable share of the benefit arises before any elaborate technology is involved.
In 2025, incoming orders in the German main construction sector reached a nominal €113.0 billion, 6.8 percent higher in real terms than the year before. Civil engineering exceeded its previous record from 2024 by a further 6.2 percent in real terms. More construction sites mean more traffic safety – but also more parallel projects with the same number of staff.
Information is rarely lost within a department. It is lost at the transition between enquiry and job, between order and work instruction, between crew and invoicing. That is exactly where AI pays off.
An AI system may neither set traffic arrangements nor change them on its own authority. It may prepare, check, flag and summarize. Selecting the standard plan, setting clearances and granting approval remain with the qualified responsible person.
Unambiguous job numbers, traceable plan versions and a central document store matter more than the choice of model. AI does not turn poor documentation into good documentation. It only makes contradictions visible faster – or adopts them unnoticed.
Since February 2025 the EU AI Act has required an appropriate level of AI literacy among staff. The transparency obligations take effect on 2 August 2026. The extensive high-risk obligations were postponed by the Digital Omnibus on AI to December 2027 and August 2028 respectively. An assistant that summarizes emails is, in any case, to be judged differently from a system that intervenes in traffic control.
In a safety-critical trade, completeness, traceability and responsiveness count at least as much as saved processing time. A complete setup record is valuable even when it saves not a single minute.
Begin with a limited, reversible use case in which an error endangers no one and a person approves the result: enquiry analysis, document search, preparing inspection reports, or voice input for site notes. Chapter 22 sets out a 90-day sequence for this, with a clear option to stop.
In the coming years Germany will see more road works, not fewer. Bridges have to be renewed, roads resurfaced, fibre-optic and utility lines laid, cycle paths extended and municipal infrastructure modernized.
Source: Federal Statistical Office, press release No. 061 of 25 February 2026 (main construction sector 2025). destatis.de
For traffic-management companies this development means broadly stable demand. Rail-line renewal and the expansion of digital infrastructure were the main drivers of the large orders in 2025 – both fields that generate road works on a lasting basis. At the same time, the demands on these businesses are rising.
In 2025, 2,832 people died in road traffic accidents in Germany – 62 more than the year before. Around 49,200 people were seriously injured and about 322,000 slightly. The number of serious injuries thus reached its lowest level since 1991, while the number of fatalities rose again. The police recorded around 2.52 million accidents in total.
Traffic safety is therefore not an organizational side task that can be costed in passing. It is an essential part of road-safety work – and any gap in planning, setup or inspection can have consequences that go far beyond a deduction on the invoice.
Source: Federal Statistical Office, press release No. 238 of July 2026, final results for road accidents 2025. destatis.de
At first glance, traffic safety is a physical business: signs, delineator beacons, barrier boards, portable traffic signals, road markings, mobile safety barriers and vehicles. In daily operations, however, it is usually the quality of the information that determines whether a project can be carried out economically and in line with the rules.
None of these questions is answered by materials. All of them are answered by information – which has to be gathered from documents, conversations, local knowledge and experience, usually under time pressure and usually more than once, because the result is not recorded in a single place.
This is precisely where AI can contribute. It processes text, images, tables, voice input and structured data. That makes it suitable for processes in which information has to be captured, checked, assigned, condensed and documented – in other words, for the greater part of the case work in a traffic-management business.
A look at the adoption rate helps put things in context. According to the ICT survey of the Federal Statistical Office, in 2025 around 26 percent of all companies with ten or more employees used AI technologies: 23 percent of small companies with 10 to 49 employees, 36 percent of medium-sized ones with 50 to 249, and 57 percent of large ones with 250 or more. For companies with 20 or more employees, the industry association Bitkom now measures an adoption rate of 41 percent, up from 17 percent a year earlier.
The trades tell a different story. According to a Bitkom survey of 504 businesses in 2025, only 4 percent of trades companies were using AI, and a further 9 percent were planning to. For businesses working in a construction- and trades-related environment, the gap to the rest of the economy is therefore considerable – and the lead is correspondingly large for anyone who makes an early but cleanly managed start.
Sources: Federal Statistical Office, ICT survey 2025 (destatis.de) · Bitkom e. V., press release of 11 March 2026 and study "Digitalization of the trades 2025" (bitkom.org)
The term "AI" is used for very different technologies. Five categories are relevant for traffic safety. They differ not only technically, but above all in one question: how much trust their results deserve.
Language models process tenders, orders, emails, records and regulatory texts. They summarize content, extract information, compare documents and draft text. In practice, this means analysing an enquiry, summarizing an official traffic order, spotting missing details, preparing queries, drafting a work instruction or an inspection report, and reviewing a bill of quantities.
Language models do not, however, deliver automatically reliable professional decisions. They can misinterpret content, use out-of-date information or invent details that sound plausible. What makes this especially treacherous is that errors sound just as assured as correct results. Professionally relevant outputs must therefore be reviewed and approved – and the system should show what it based its answer on.
Computer vision refers to the automated analysis of images and video. In the traffic-safety environment it can recognize traffic signs in site photos, assess the condition of delineator beacons or barrier boards, flag damaged or toppled elements, sort image series by location and construction phase, roughly classify traffic volumes, count vehicles, cyclists or pedestrians, and highlight differences between the target and actual state.
Image recognition does not replace a competent inspection. Its value lies in being an additional check stage and in pre-sorting large volumes of images – there, where today no one reviews every photo any more, simply because there are too many.
Forecasting models analyse historical and current data to estimate the expected workload, material and staff requirements, likely deployment duration, traffic strength in particular time windows, disruption probability, maintenance needs of equipment or possible supply shortages.
Their quality depends entirely on the available data. If deployment times, material movements and disruptions are not captured in a structured way, even a good model delivers no reliable forecast. This use case therefore rarely belongs at the beginning.
The closer an AI application sits to a safety-relevant decision, the tighter the human control must be – and the later it should appear in the roll-out plan.
Optimization methods support decisions in which several conditions apply at once. In traffic safety this includes route planning for inspection runs, assigning crews to sites, providing materials, dispatching vehicles, planning setup and takedown sequences, minimizing empty runs, and accounting for qualifications and working hours.
Unlike language models, optimization methods invent nothing. They calculate precisely – but they calculate only with what they know. What an experienced dispatcher knows about a difficult access point, an unreliable client or a slow crew is not in the model. Optimization therefore provides proposals; dispatching keeps the authority to decide.
AI agents combine language models with software functions. They do not just read; they act. An agent can read a new enquiry, recognize the site address, identify missing details, create a case and prepare a draft query.
A more far-reaching agent could process an entire inbox:
Such systems should not act without oversight. The more an agent does itself, the more precisely it must be defined what it may not do. Changes to jobs, permits, plans or invoice-relevant data require defined approvals. A workable rule: the agent may prepare everything and finalize nothing.
As soon as an agent processes documents from the inbox, it processes content that comes from outside. A manipulated PDF or email can contain hidden instructions that the system mistakes for a task – such as ignoring certain conditions or forwarding data. This risk is called prompt injection, and it is not a theoretical problem.
The countermeasures are organizational and technical: content from external documents is treated as data, not as instructions. Outgoing actions – sending messages, exporting data, finalizing cases – require human confirmation. And every action is logged.
Language models read and write. Computer vision sees and sorts. Forecasting models estimate. Optimization calculates. Agents act. For every category the same question applies: what happens if the result is wrong – and who notices?
AI does not change the professional and legal foundations of traffic safety. It only changes how quickly the information needed to comply with those foundations becomes available. Anyone who confuses the two builds a compliance problem instead of relief.
Under § 45 of the German Road Traffic Regulations (StVO), road traffic authorities may restrict, prohibit or divert the use of roads or road sections for reasons of traffic safety or order. This applies in particular to work in the road space. Before starting work that affects road traffic, contractors must obtain the corresponding orders and implement them.
For AI systems this establishes a clear boundary: they may not make unauthorized changes to ordered traffic arrangements. If a system detects a possible deviation or a supposed improvement, this must give rise to a review or coordination process – never an automatic adjustment. The order is an administrative act, not a recommendation.
The "Guidelines for the traffic-related safeguarding of road works" in their 2021 edition (RSA 21, FGSV 370) are the central technical basis for the traffic-related safeguarding of road works. They cover work zones in urban areas, on rural roads and on motorways, and contain standard plans for work zones of shorter and longer duration.
One point that regularly causes misunderstandings in practice: the RSA 21 do not automatically apply identically nationwide. They are made binding through introductory decrees of the federal states. Some states have introduced them unchanged; others have made separate provisions that must be observed during planning and execution.
A standard-plan assistant that knows only the RSA 21 is incomplete. It must know in which federal state – and, where relevant, under which road owner – the work zone lies, and which state regulation applies there. Otherwise it delivers proposals that are formally correct but practically wrong. In every case, selecting and adapting the plan remains the task of qualified staff.
The Technical Rule for Workplaces ASR A5.2 covers requirements for workplaces and traffic routes on construction sites in the border area to road traffic. It addresses the protection of workers and gives concrete form to the requirements of the Workplaces Ordinance. The applicable edition dates from December 2018 (Joint Ministerial Gazette, GMBl 2018, p. 1160) and was last amended formally in 2022 (GMBl 2022, p. 252).
RSA 21 and ASR A5.2 pursue different protective aims. The RSA 21 address the traffic-related safeguarding and management of traffic. The ASR A5.2 address the safety and health of workers. Since the RSA 21, occupational-safety provisions have been deliberately excluded from the RSA – so the two works interlock rather than overlap.
On this point, the Federal Highway Research Institute has published a "Handbook on the interaction of ASR A5.2 and RSA in planning road works in the border area to road traffic" (2020 edition, published in the Transport Gazette No. 4/2021). A digital planning or checking system should reflect both perspectives. A formally suitable standard plan is not enough if the working and safety spaces for the workforce are missing.
Sources: BAuA, ASR A5.2 (baua.de) · BASt, Handbook ASR A5.2 / RSA (bast.de) · FGSV Verlag, RSA 21 and MVAS 99 (fgsv-verlag.de)
The "Guidance sheet on the framework conditions for the expertise required to safeguard road works" (MVAS 99, FGSV 371) sets out which knowledge those responsible for traffic safety must demonstrate. It contains training programmes for different affected groups – from the ordering authority through the client to the executing company.
The RSA 21 refer to it explicitly: under Part A, Chapter 1.4, paragraph 3, those responsible for traffic safety must demonstrate the required expertise in accordance with the MVAS. For work with minor traffic effects, the ordering authority may permit exceptions and should require the naming of a deputy with the same qualifications. The responsible person is to be named in the official traffic order; the qualification proof is demanded in almost all public tenders and checked for currency.
AI competence does not replace this expertise – not even remotely. Conversely, an experienced responsible person should understand how AI results come about, which errors typically occur and when a proposal must be discarded. This second direction is easily overlooked: a responsible person who trusts a system blindly, because they do not know its limits, is a greater risk than a business with no AI at all.
| Regulation | What it means for the business |
|---|---|
| ZTV-SA (FGSV 369) | Additional technical contract conditions and guidelines for safeguarding work at road works. Governs contract and execution quality and requires proof of expertise. |
| ATV DIN 18299 / 18329 | General technical contract conditions of the VOB/C; DIN 18329 covers traffic-safeguarding work and is the basis for the specification and invoicing. |
| ArbSchG / ArbStättV | Risk assessment, instruction and documentation duties – for the traffic-safeguarding work itself, not only for the site being safeguarded. |
| GDPR / BDSG | Relevant as soon as photos, GPS data, voice recordings or employee data are processed. See Chapter 17. |
| AI Act (EU) 2024/1689 | AI literacy, transparency, risk classification. See Chapter 17. |
AI may support qualified staff. It may not obscure responsibilities, bypass permits or make safety-relevant decisions without human control. Anyone who writes this sentence into an internal policy and holds to it has already completed the greater part of AI governance in traffic safety.
To find sensible AI applications, you should not search for general AI ideas. The starting point is the actual job process – as it runs in the business, not as it appears in the quality manual.
| No. | Process step | Typical bottleneck |
|---|---|---|
| 1 | Enquiry or tender | inconsistent formats, incomplete details |
| 2 | Reviewing the documents | ties up experienced staff, result not documented |
| 3 | Site visit and measurement | findings stay in someone's head or on a phone |
| 4 | Planning the traffic arrangement | recurring review steps, no four-eyes check |
| 5 | Application for the official traffic order | data has to be gathered again |
| 6 | Coordination with authority and parties | results scattered across email threads |
| 7 | Costing and quotation | time pressure, inconsistent assumptions |
| 8 | Deployment and materials planning | double reservations, missing stock data |
| 9 | Setting up the traffic safety | outdated plan version with the crew |
| 10 | Acceptance or handover | proof incomplete |
| 11 | Regular inspection | records on paper, deadlines not monitored |
| 12 | Changes and disruptions | settled by phone, not documented |
| 13 | Takedown | additional services not captured |
| 14 | Final documentation | gathered together afterwards |
| 15 | Measurement, proof of performance, invoicing | deductions for lack of proof |
Information can be lost at every transition. The most critical points are those where the medium changes – because no one is responsible there:
AI can secure these transitions. The benefit usually lies less in spectacular features than in the consistent structuring of information at exactly those points where it evaporates today.
Traffic-management companies receive enquiries in very different forms – from a formal bill of quantities to a phone note. The first review ties up precisely the staff who are most urgently needed for other tasks.
The first review calls for judgment: what is this about? Which documents are missing? Can the business deliver the service economically and to professional standard? AI cannot answer these questions. But it can prepare everything needed to answer them – and that is the time-consuming part.
These fields do not produce a finished job; they produce a structured checklist. The difference is essential: the system delivers a basis for work, not a decision.
A classic OCR tool reads characters. A language model understands that "excavation approx. 35 m, footway closed" touches the fields length, mode of traffic affected and management requirement – and that the residual carriageway width is then missing. The real work is not the reading, but the recognition of the gap.
| Present | Still open |
|---|---|
|
|
The case handler thereby receives not a finished job, but a structured basis for the further work. They decide which points to query, whether the enquiry suits the business at all, and how urgent it is.
The following conditions are not exclusion criteria, but good indicators. If at least three of them apply, the use case is usually economic:
The system may not create quotations, set prices or reject an enquiry on its own. It prepares. The judgment of whether a job suits the business remains an entrepreneurial decision.
Planning a traffic arrangement is safety-relevant. An AI may therefore not decide on its own which traffic control plan is professionally admissible. It can, however, relieve the planner of the recurring steps that eat up most of the time.
The user describes the work zone: urban or rural, shorter or longer duration, carriageway width, number of lanes, existing footways and cycle paths, position of the work zone, planned traffic arrangement, target speed and particular local conditions. The system proposes suitable starting points from the body of rules – taking account of the introductory regulation applicable in the respective federal state.
The planner decides whether a standard plan is applicable or an individual traffic control plan is required. The assistant shortens the searching, not the thinking.
A checking assistant can verify whether details are missing from the plan or the accompanying documents. This is a mechanical task that people reliably do badly – especially on a Friday afternoon at the twelfth plan:
When changes occur, AI can compare different plan versions and summarize the differences in plain language: sign position moved, speed changed, pedestrian routing adjusted, construction phase extended, additional access added, lane width altered, authority condition added.
This function looks unremarkable but addresses one of the most frequent and most expensive problems in operations: that fitters or clients work with an outdated plan.
The decisive point of this model is Stage 3. As long as an approval does not exist as a separate, logged step, responsibility blurs – regardless of whether an AI is involved. Introducing an AI assistant is, in experience, a good occasion to define this step cleanly at last. Not seldom, that is the greater gain.
Put the checking assistant to work first on completed projects whose outcome you already know. Does it find the errors you found yourself back then? Does it find any you overlooked? Does it report things that are not errors? These three numbers say more about its suitability than any product demonstration.
Applications for an official traffic order contain many recurring details. At the same time, forms, processing routes and requirements differ considerably between authorities – sometimes even between two neighbouring municipalities. It is precisely this combination of repetition and variance that makes a good AI use case.
Filling in a form faster saves minutes. The real gain comes from the fact that the data no longer has to be gathered anew from emails, plans and master data for every application. Instead, the system draws on a central case – and the case knows which documents depend on it.
If, for example, the construction period changes, the assistant can immediately show what is affected:
This is the kind of knowledge that today usually lives in a single head in the back office – and that vanishes when that person is on leave.
The return journey is at least as valuable as the outward one. An issued order is a dense document that has to be turned into concrete tasks in the business – and from which individual conditions regularly fall out.
From this arises a comprehensible work summary for planning, dispatching and the crew – in the language spoken on site, not the language of the official notice.
The authority adds five points to the order that were not envisaged that way in the application:
Each of these points has operational consequences: night premiums, additional material for the pedestrian routing, a daily inspection appointment, a phone call and a programme change at the signals. The assistant flags the points as execution-relevant and creates tasks for the responsible staff. Two of them – night work and the daily inspection – also affect the costing.
The original order remains the binding document. The summary is a working aid and does not replace it. In case of doubt, the official notice applies, not the summary – and every summary must link back to the original.
If orders are captured in a structured way over the years, something arises along the way that hardly any business has today: an analysable history of which authority typically imposes which conditions, how long it takes to process on average, and which documents it regularly requests afterwards.
That is not a spectacular AI application. But it is the basis for costing lead times realistically instead of guessing them – and, for new staff, the difference between "go and ask Mr Meier" and a dependable answer.
In most traffic-management companies, dispatching is the operational bottleneck. It has to take account of short-notice changes, staff availability, vehicle capacities, material stocks, qualifications and site locations all at once – and without the time to look everything up.
From the job data and the approved plan, a resource requirement can be calculated: traffic signs by type and number, delineator beacons, barrier boards, warning lamps, base plates, portable traffic signals, mobile safety barriers, marking material, mounting devices, trailers, work vehicles, required staff, necessary qualifications and the expected setup time.
The system can then check what, in practice, often no one checks in full:
Automatic optimization proposals do not fully capture operational experience. An experienced dispatcher knows the actual performance of individual crews, local particularities, the reliability of clients, the typical delays of certain authorities, difficult access points, material conditions and dependencies within construction sequences.
The right approach is therefore not to replace this knowledge, but to let it flow gradually into the system as structured data and rules. If a crew regularly needs 20 percent longer for a particular type of setup, that belongs in the planning values – not in one person's head.
Dispatching projects rarely fail on the technology. They fail because no one defined beforehand what was supposed to improve. The following metrics can be gathered in almost any business – often before introduction.
| Metric | What it shows |
|---|---|
| Share of setups completed on time | reliability towards client and authority |
| Short-notice re-plans per week | how disruption-prone the planning is |
| Empty runs | direct costs, often underestimated |
| Material shortfalls | quality of the stock data |
| Overtime | workload and planning quality |
| Unproductive waiting times | coordination losses at the interfaces |
| Utilization of vehicles and traffic signals | capital tied up |
| Number of unplanned extra trips | completeness of the preparation |
| Time from job approval to complete deployment plan | throughput time of dispatching |
AI-supported dispatching presupposes that material stocks, storage locations, qualifications and deployment times are available digitally and up to date at all. If this basis is missing, dispatching is not the right entry point – then a digital deployment folder or an enquiry analysis brings considerably more, and the data foundation grows along the way.
No dispatcher accepts a system that makes proposals they cannot follow. The assistant must therefore be able to justify why it suggests a particular assignment – for instance: crew B, because crew A is already committed the following day and the qualification for the traffic signals is held by only two staff.
A proposal without justification is ignored or overridden on principle. A proposal with justification is examined – and then either accepted or corrected. Both are a success: the correction is the most valuable feedback the system can receive, because it makes the missing knowledge visible.
A dispatching assistant that cannot explain itself will not be used. A dispatching assistant that cannot be contradicted becomes dangerous.
On site, fitters do not need all the project documents. They need the information relevant to their task – in a comprehensible and demonstrably current form. A 40-page PDF on a smartphone at the edge of the carriageway is no basis for work.
An AI assistant can answer questions about these documents – spoken or typed: which condition applies to the site access? When must the first inspection take place? What minimum width is envisaged for the footway? Who must be informed in the event of a delay? Which plan version is approved?
The answer must rest on the approved project documents and show the source used. An assistant that answers from general model knowledge is, at this point, not merely useless but dangerous.
In the site environment, typing is impractical: gloves, noise, time pressure, darkness. Staff can therefore capture information by voice. A realistic example:
The system turns this statement into a structured case: time, site location, type of deviation, element affected, reason, action taken, person informed, further approval required.
Experiential knowledge and deviations no longer stay hidden in phone calls. They become part of the project documentation – at the moment they arise, not three weeks later at invoice review, when no one remembers any more.
For the fitter, the process must be shorter than reaching for the phone. If it takes longer, it will not be used. That is not an acceptance problem but a sober judgment about the tool.
Extensive photo documentation does not automatically protect against disputes. What matters is whether images are unambiguously assigned, chronologically traceable and sufficient in content. Two hundred unsorted photos in a folder are not proof in court but a problem.
Images can be classified automatically – by advance warning, start of the safeguarded area, lane shift, longitudinal barrier, transverse barrier, pedestrian routing, cycle routing, site access, traffic signals, diversion signing, end area and overview shot.
More valuable than the classification is often the feedback to the member of staff while they are still on site:
A prompt after two seconds costs one extra press of the shutter. The same defect, discovered three weeks later at invoice review, costs a return trip – or the invoice item.
In a more advanced solution, plan information is linked with site images. The system flags possible deviations: a planned traffic sign was not recognized, an additional object is in the safety area, a delineator beacon has toppled, a barrier is interrupted, a lamp may have failed, a footway is blocked, the site access is not safeguarded as envisaged.
A flag does not mean a defect exists. It triggers a human review. A system that reports "no defect" is, incidentally, considerably more dangerous than one that warns too often – Chapter 28 explains why this application still calls for caution.
Site photos are seldom free of personal data. A shot of the safeguarded area can quickly show number plates, faces, properties, residents or one's own employees – even when no one meant to photograph them.
| Question | What it is about |
|---|---|
| Necessity | Is the shot really needed for the proof? Would a different framing suffice? |
| Legal basis | Performance of a contract, a legal obligation or a legitimate interest – with a documented weighing of interests. |
| Retention periods | How long? Warranty and limitation periods are the usual benchmark here. |
| Access rights | Who sees the images? Client, insurer, authority – on what basis in each case? |
| Masking | Automatic pixelation of faces and number plates is technically available and, in many cases, the simplest route. |
| Employee data | Is the documenting member of staff identifiable? Does this create performance monitoring? See Chapter 18. |
The German data protection authorities (Datenschutzkonferenz) recommend taking data protection requirements into account as early as the selection and introduction of an AI application – not only once the system is running. For systems that connect internal knowledge sources with language models, they additionally published guidance on retrieval-augmented-generation systems in 2025. The Federal Commissioner for Data Protection and Freedom of Information takes the same line.
Set a short photo policy before you roll out a camera function: which subjects are required, which are to be avoided, how long they are stored and who has access. Two pages are enough. This policy is at the same time the basis for the photo requirements in the digital deployment folder – so data protection here produces a direct operational benefit.
As soon as cameras permanently capture moving traffic – for counting, say, or for detecting congestion – the assessment shifts markedly. Here a data protection impact assessment regularly has to be examined, and the further question of classification under the AI Act arises. Such applications do not belong in a first pilot.
Sources: Datenschutzkonferenz, guidance on artificial intelligence and data protection and on RAG systems (datenschutzkonferenz-online.de) · BfDI, guidance on the use of AI (bfdi.bund.de)
Traffic safeguarding must not only be set up correctly; it must be inspected and maintained throughout its entire duration. This duty is the part of the business most often underestimated and most rarely proven cleanly.
A mobile inspection application can generate the check points depending on the project – instead of using the same generic list for every job:
With many ongoing jobs, not every inspection is equally urgent. A system can prioritize inspection runs – based on traffic strength, standing time, severe-weather conditions, previous disruptions, frequent third-party interference, complaint volume, complexity of the traffic routing, presence of traffic signals, night work zones, known accident clusters, the share of heavy traffic, and importance as a school or cycle route.
Prioritization means order, not selection. Ordered or contractually agreed inspection intervals must be kept – even when a model deems the job uncritical. A system that lets inspections lapse violates conditions and traffic-safeguarding duties.
For operational planning, inspection points can be bundled into routes, taking account of prescribed inspection times, the current traffic situation, the priority of the job, the member of staff's starting point, the expected inspection duration, the replacement material needed and vehicle capacity.
On completion, the system produces a draft report with the inspected job, time, observed state, documented defects, corrections carried out, replacement material needed, forwarded tasks, photos and the member of staff's name. The member of staff reviews and confirms – and it is precisely this confirmation that is the proof, not the draft.
During ongoing jobs, events occur – and regularly outside office hours. The quality of the handling is decided in the first few minutes: who takes the report, how completely, and what happens with it?
An AI assistant can extract the essential information from a phone call or a voice message and transfer it into a form that otherwise no one fills in at 11 p.m.:
The value lies less in the speed than in the completeness. A report that serves three weeks later as the basis for an insurance question or a liability dispute is only as good as what was captured that night.
On-call duty is where poor documentation takes its most expensive revenge and where it is hardest to insist on. Anyone who reduces capture here to a voice message lowers the barrier decisively – and gains data that would otherwise never exist.
A disruption report without a priority is a queue. A simple, three-tier scheme is enough in most businesses – provided it is set down in writing and not left to the mood of the day.
| Priority | Characteristics | Response |
|---|---|---|
| 1 · Acute danger | unsecured traffic space · failed traffic control · major accident damage · blocked emergency route · significant danger to pedestrians, cyclists or workers | immediate escalation to on-call duty and the responsible office; feedback mandatory |
| 2 · Time-critical disruption | damaged or missing elements · disruption without immediate acute danger · obstruction of an access · repeated complaint · foreseeable escalation | handling within a defined time window; the deadline is monitored |
| 3 · Note | minor deviation · improvement proposal · soiling · incomplete documentation · information for the next inspection appointment | added to the regular inspection or planning |
After approval, the system can send prepared messages: a report to the client, information to the authority, a work order to on-call duty, a note to dispatching, feedback to residents, documentation for the insurer or police.
Two words are decisive here: after approval. An automatically sent report to an authority that rests on a misinterpretation is harder to correct than a delayed phone call.
An AI may not declare a safety-relevant incident closed on its own. It may capture it, classify it, forward it and prepare the closure. Establishing that no danger remains is a professional decision with liability consequences – and needs a name behind it.
Who is to be informed, and when? What happens if that person cannot be reached? At what point is management brought in? When must the authority be notified, when the client, when both?
These questions are not technical. In practice, though, they are often answered only once a system asks them. That, too, is a legitimate benefit of introducing AI: it forces the clarification of responsibilities that were previously settled informally – that is, not at all.
Modern traffic control increasingly uses data from sensors, counting stations, vehicles, traffic control centres and digital platforms. Not all of these developments are directly usable for a medium-sized business – but the direction is clear.
With AutobahnOS, the federal motorway company (Autobahn GmbH des Bundes) is developing a nationally uniform digital platform for its traffic control centres. One of the specialist applications expressly concerns road-works management: it is intended to standardize business and permit processes and to improve the planning of all work zones on motorways. According to the company, all traffic control centres are to work with the system and all available applications by the end of 2027; individual applications are already in use, others are still being tested or developed.
Research, too, is working on AI-supported infrastructure analysis. The mFUND project TWIN4ROAD is investigating an infrastructure database for the road space – including signing, road markings, kerbstones and traffic signals – as well as an AI-based road-condition assessment on the basis of 3D, image and ground-radar data. The Smart-Check project is working on an automated, cross-disciplinary conformity check in the planning and permit procedure of infrastructure projects.
For traffic-management companies this means, in the medium term: services will be more strongly linked with digital road, construction and traffic data – and clients will increasingly expect that data can be handed over in a structured form, not as a bundle of PDFs.
On the A8 crossing of the Enz valley near Pforzheim, a dynamic congestion-warning system has been in operation since before construction began. Citing the road-accident statistics of the Pforzheim police headquarters, Autobahn GmbH reports that the system has prevented accidents and that, in a three-year comparison, serious accidents involving personal injury in particular have been reduced. According to the police in December 2025, there were no fatal accidents in the work zone in 2024 or 2025.
At the same time, the total number of accidents on the section rose from 594 (2024) to 668 (2025) – predominantly damage-only collisions. The message is therefore more nuanced than it usually appears in marketing material: data-based warning systems can markedly influence the severity of accidents in work zones. They do not prevent every accident.
Sources: Autobahn GmbH des Bundes, AutobahnOS and the A8 Enz-valley crossing project (autobahn.de) · Federal Ministry for Transport, mFUND projects TWIN4ROAD and Smart-Check (bmv.de) · Pforzheim police headquarters, information of December 2025
In many traffic-management companies, crucial knowledge is held in scattered form – and a considerable part of it only in the heads of experienced staff. That is not a tidiness problem but a risk: it is lost with every retirement and every resignation.
Technically, this usually uses a RAG method (retrieval-augmented generation). The language model draws on approved internal documents instead of answering solely from general model knowledge – and can supply the source location.
A knowledge system must not be built on an uncleaned collection of old PDF files, private notes, outdated plans and contradictory instructions. AI does not turn poor documentation into good documentation. It merely makes existing contradictions findable faster – or adopts them unnoticed. The old RSA 95 sitting next to the RSA 21 in the same folder is not a detail but an error waiting to happen.
A chatbot on its own changes the business only to a limited degree. It knows neither the current jobs nor the approved plans. The greater benefit arises when the AI is connected to the operational systems – and those need not be built anew for it.
| Layer | Contents |
|---|---|
| Customers and sales | CRM · enquiry inbox · tendering platforms · customer portal |
| Job management | project and job numbers · status · dates · contacts · scope of services |
| Planning | traffic control plans · maps · geodata · construction phases · approvals |
| Resources | material stock · storage locations · vehicles · trailers · traffic signals · staff · qualifications |
| Mobile execution | digital deployment folder · setup checklists · photo documentation · inspection reports · disruption reports · time recording |
| Documents | applications · orders · plan versions · proofs · records · invoicing documents |
| Analysis | metrics · deviations · disruptions · contribution margins · material utilization · post-costing |
AI forms an assistance layer above these systems. It does not replace them. Its task is a double translation: it turns unstructured information into structured data – and it makes existing structured data easier to reach, without anyone having to know the operating logic of four programs.
This framing matters more than it sounds. It means that introducing AI need not be a change of system. A business with working industry software does not have to replace it. It has to make it accessible.
The most important question to put to an AI provider is therefore not "Which model do you use?", but "How do you get at our job data – and how do we get it back out again?" Chapter 26 contains a full checklist for this.
A concrete example makes the difference between a chatbot and a connected assistance layer clearer than any architecture diagram.
Four sentences. For an experienced case handler, ten minutes of work – create the case, transfer the data, spot the gaps, draft a query. For a new member of staff, half an hour, with the risk of overlooking three points.
| Recognized and transferred | Flagged as missing |
|---|---|
|
|
The case handler reviews the data and transfers it into the case. Two things are worth noting. First: the system decided nothing. Second: it noticed the cycle traffic even though it was not mentioned in the email – because a footway closure with a pedestrian diversion regularly requires the cycle routing to be clarified as well.
Once the case exists, everything else can build on it. The query to the construction firm is prepared. As soon as the order arrives, it is assigned to the case and evaluated. The conditions become tasks. The deployment folder is generated from the approved plan. The photo documentation automatically carries the job number. The final file does not have to be gathered together, because it came into being along the way.
A chatbot answers questions. A connected assistance layer creates and maintains the case that everyone involved later relies on. The first saves minutes. The second changes how the business works.
Many AI projects fail not on the model but on the state of the data. That is an uncomfortable insight, because it means part of the work lies before the actual project – and that this part is not particularly exciting.
Before a larger AI roll-out, at least these foundations should be in place. None of them is an IT project – all are organizational decisions:
That is the other side of the coin – and it matters just as much. Anyone who waits until the data is right waits forever. A pilot can also start with a limited data foundation, if the use case is deliberately chosen to suit it.
| Well suited as an entry point | Less suited as an entry point |
|---|---|
|
|
The pattern is clear: the suitable entry points process documents that are there anyway. The unsuitable ones presuppose historical operating data that has yet to come into being. The trick is to start with the first and, in doing so, create the foundation for the second.
The AI Regulation (EU) 2024/1689 entered into force on 1 August 2024. Its obligations apply in stages. In spring 2026, this timetable was changed in essential parts by the Digital Omnibus on AI – anyone still planning around the original calendar is working with outdated assumptions.
On 7 May 2026, the Council and the European Parliament reached a political agreement in trilogue. Parliament gave its approval on 16 June 2026, and the Council formally adopted the legal act on 29 June 2026. The legislative procedure is thereby concluded; publication in the Official Journal was expected in July 2026, with entry into force on the third day thereafter.
| Set of obligations | Applicable from |
|---|---|
| Prohibited practices (Art. 5) | 2 February 2025 – unchanged |
| AI literacy of staff (Art. 4) | 2 February 2025 – unchanged |
| Obligations for general-purpose AI (GPAI) | 2 August 2025; enforcement from 2 August 2026 |
| Transparency obligations (Art. 50) | 2 August 2026 – unchanged |
| New ban on NCII- and CSAM-generating systems | 2 December 2026 |
| High-risk obligations, standalone systems (Annex III) | 2 December 2027 (previously: 2 August 2026) |
| High-risk obligations, product-embedded systems (Annex I) | 2 August 2028 (previously: 2 August 2027) |
The background to the postponement is the delayed availability of the harmonized standards that are meant to give concrete form to the high-risk obligations. The supervisory authorities expressly assume that companies will use the additional time for the preparation that is due in any case – risk management, data quality, technical documentation and governance.
The postponement concerns solely the full high-risk obligations. AI literacy, transparency and the bans continue to apply on their own calendar. For a medium-sized traffic-management business that operates no high-risk AI, the Omnibus has changed little in practice – the tasks remain the same.
Position as of 17 July 2026. Legal developments after this date are not taken into account. Sources: EUR-Lex, Regulation (EU) 2024/1689 (eur-lex.europa.eu) · European Commission, Digital Omnibus on AI, procedure 2025/0359(COD) (ec.europa.eu)
| Application | Classification (indicative, not legal advice) |
|---|---|
| Assistant that summarizes emails and orders | Regularly low risk; possibly transparency obligations |
| Internal knowledge system on approved documents | Regularly low risk; data protection is the more relevant lever here |
| Photo classification for documentation review | Low to limited risk; observe data protection |
| Systems for assessing employee performance | Annex III – employment context; high requirements. See also Chapter 18. |
| Safety component in the management of critical infrastructure, e.g. in traffic control | Annex III – high requirements; obtain professional and legal advice early |
The classification depends on the specific purpose, the technical role and the effects – not on the name of the product. A traffic-management company is, as a rule, a deployer, not a provider; but if it develops its own systems or has them developed under its own name, that can change.
The use cases 1 to 10 described in this ebook do not, as things stand today, fall for the most part into the high-risk area – as long as human approval is retained and no systems for assessing employee performance arise. That is precisely why the approval architecture from Chapter 6 is not only a safety but also a compliance decision.
Personal data occur in traffic safety in more places than is at first apparent: names and contact details, employee data, GPS positions, working hours, photos, video recordings, number plates, voice recordings, caller data, complaint data, and performance and behaviour data of staff.
What has to be examined is the purpose and legal basis, data minimization, retention periods, processing on a controller's behalf, a possible transfer to third countries, technical and organizational measures, access rights, logging, data-subject rights and the need for a data protection impact assessment.
Two points deserve particular attention. First, processing on a controller's behalf: if an AI provider processes documents containing employee or resident data, a contract under Article 28 GDPR is required – even when no one wants to "use" the data. Second, model training: it must be contractually excluded, or expressly regulated, that content entered is not used for training.
An AI system connected to the operational systems gains access to sensitive information: safety concepts, site plans, infrastructure information, access credentials, contact details, communication with authorities, contract data and internal costings. For infrastructure projects this is a relevant attack surface – site plans tell third parties precisely where and when a road lies open.
Public clients increasingly ask where, and by whom, data is processed. Anyone who can demonstrate processing within the EU and cleanly document the processing arrangement has, in a procurement, not a disadvantage but an argument. Data protection under the EU GDPR is, in this environment, not an obstacle but a mark of quality.
Two topics are regularly considered too late in AI projects – and then cause the greatest delays. Neither concerns technology; both concern people and responsibility.
Under § 87(1) no. 6 of the German Works Constitution Act (BetrVG), the works council has a right of co-determination in the introduction and use of technical devices that are intended or suited to monitor the conduct or performance of employees. The word "suited" is decisive: it does not matter whether monitoring is intended.
In traffic safety, this concerns a whole series of the applications described here:
| Application | Why it is relevant to co-determination |
|---|---|
| Vehicle GPS data | location and driving behaviour can be assigned to individual staff |
| Mobile time recording | working hours per person, often with a location reference |
| Photo documentation with staff identification | allows inferences about work result and pace |
| Inspection reports with names | number and quality of inspections per person become analysable |
| Dispatching data with actual times | comparison of crews and individuals becomes possible |
The practical advice is simple: involve the works council before the selection decision is made, not after. A works agreement that sets out which analyses expressly do not take place creates trust and speeds the project up. Conversely, a co-determination right that has been bypassed can block the use of a finished system.
Whoever sets up a work zone in the road space assumes the traffic-safeguarding duty (Verkehrssicherungspflicht) for that area. If they breach it and damage results, civil claims and, in serious cases, criminal consequences come into consideration. This responsibility is tied to persons – to the entrepreneur and to the named responsible person.
The sentence "the system did not report it" is no defence. It merely shifts the question: why was a system deployed whose limits were not known? Who reviewed the output? Where is that documented? An AI system cannot assume liability – it can only make it better or worse documented.
From this, however, the positive side also follows. A gapless, chronologically traceable documentation of setup, inspections and disruptions is, in a dispute, the most effective means of proving the proper state. It is precisely here that a benefit lies which no time-saving calculation captures – and which, in a single serious case, can justify the entire cost of introduction.
Not every company needs its own AI infrastructure. In principle there are three models. They are not mutually exclusive – most businesses pass through them one after another.
Suitable for: text drafts, summaries, internal research, first pilot applications, low integration requirements.
| Advantages | Disadvantages |
|---|---|
|
|
Suitable for: searching bodies of rules, internal work instructions, project-related document search, order and contract analysis, standardized assistance processes.
| Advantages | Disadvantages |
|---|---|
|
|
Suitable for: automated enquiry processing, job creation, dispatching, mobile documentation, inspection management, communication with customers and authorities, metrics and forecasts.
| Advantages | Disadvantages |
|---|---|
|
|
The right entry point follows not from the desired goal, but from today's starting position. In the left column, find the description that best fits your business.
| Starting position | Recommended entry point | Model |
|---|---|---|
| Few structured data, many documents | Company knowledge and document analysis | B |
| Many enquiries, high effort on initial screening | AI-supported enquiry analysis | A → B |
| Frequent errors at handovers | Digital deployment folder | B → C |
| High effort for proofs | Mobile photo and inspection documentation | C |
| Many parallel jobs | Dispatching support | C |
| High number of disruptions | Digital incident and escalation management | B → C |
| Good data foundation and established software | Integrated AI platform | C |
In practice, most businesses begin with Model A, because it is quickly available and costs little. That is sensible – as long as it is clear that Model A does not solve the actual bottleneck. A language model without access to the job data can compose text, but cannot say which plan version is approved.
The decisive step is therefore the transition from A to B: the moment the system accesses your own, approved documents and names source locations. From this point the use changes – a writing tool becomes a point of reference. Model C is then a question of integration, no longer of fundamental decision.
Model B lives on the quality of the documents fed into it. Before you release bodies of rules, work instructions and sample plans, it must be clear: who owns the document? Is it current? Which version applies? Who may see it? Answering these four questions for every document is the main effort – and at the same time the main benefit.
Between B and C lies a sensible intermediate stage: the knowledge system with read access to the job data. It can give information – "Which plan version applies to job 4711?" – without changing anything. The risk is low, the benefit considerable, and the interface is the same one that will later be needed for Model C.
The following cases are typified scenarios, not customer references. They show possible forms of use and openly disclosed calculation methods. They are not a promise about achievable savings. Every figure in the example calculations is an assumption that you should replace with your own values.
The company handles around 1,200 jobs a year. Enquiries arrive by email. Data is transferred manually into a job list. Orders are held as PDF files. Dispatchers have to pick out the most important conditions themselves – each in their own way.
An AI assistant analyses new enquiries, extracts the site location and period, creates a draft case, recognizes missing details, summarizes incoming orders, flags execution-relevant conditions and produces a work summary for dispatching and the crew.
| Item | Value |
|---|---|
| Cases per year | 1,200 |
| Assumed saving in administrative handling per case | 15 minutes |
| Hours saved per year | 300 h |
| Notional process cost per hour | €55 |
| Calculated potential per year | €16,500 |
To this come effects that are harder to quantify: fewer queries, faster quotations, fewer handover errors. To be deducted are software costs, introduction, training and ongoing upkeep – see Chapter 21.
The extracted data is not taken over unchecked. A case handler confirms every case. Without this step there is no time gain, only a shifting of the error to a later point – where it becomes more expensive.
Site management assigns internal crews at short notice. Plan changes are frequently passed on by phone. Different PDF versions sit on smartphones and in email inboxes. No one can say reliably which version is on which device.
A central digital deployment folder provides the approved plan version, the order, the work instruction, contacts, the material list, photo requirements and setup and inspection checklists. Changes are versioned. The crew confirms that the new plan version has been noted – and this confirmation is the actual core of the solution.
The economic effect here shows less in saved minutes than in avoided errors. A single setup based on the wrong plan version costs a journey out, a dismantling, a rebuild – and, in the unfavourable case, the argument with the authority.
The works depot looks after short-notice work zones, events, grounds maintenance and repair deployments. Knowledge about standard plans, internal responsibilities and available material is concentrated in a few experienced staff. When they are on leave or off sick, the department comes to a standstill.
An internal knowledge system answers questions on the basis of approved documents. In addition, material, vehicles and recurring types of deployment are captured digitally. A member of staff can, for example, ask:
The system provides the internal checklist, the responsible office, the required application documents, the relevant templates – and the note that a professional review remains necessary.
Works depots rarely set up traffic safeguarding at the same frequency as a specialist business – which is exactly why the routine is weaker and the need to look things up is greater. A knowledge system that knows the depot's own requirements does not replace expertise here, but lowers the barrier to applying it.
AI projects are often justified with general productivity promises. For a dependable decision, though, what is needed is concrete process data from your own business – even if that is more uncomfortable.
Annual benefit = processing time saved + avoided error costs + additionally invoiceable services + capacity gain – ongoing costs
| Item | Concrete starting points |
|---|---|
| Time saving | enquiry review · data capture · document search · producing reports · preparing applications · plan comparison · assembling invoicing documents |
| Error avoidance | wrong plan version · missing condition · undocumented additional service · material shortfall · double reservation · missed inspection appointment · incomplete photo documentation |
| Capacity gain | more jobs per case handler · more ongoing jobs per dispatcher · lower query effort · faster quotation preparation · better use of vehicles and material |
| Revenue protection | complete proofs of performance · documented extra trips · traceable standing times · better bases for variations · fewer invoice deductions |
The fourth item is the most frequently underestimated. In many businesses, the amount of unbilled or reduced services is larger than any realistic time saving – it just appears in no statistic, because it is never booked as a loss.
The last three items in particular arise permanently and are rarely itemized in quotations. A knowledge system whose documents no one maintains loses its value within a year – and then becomes a risk, because it continues to be trusted.
A company handles 800 jobs a year. The following effects are assumptions. Replace them with your own, measured values – and measure beforehand, not afterwards.
| Assumed effect | Time per job |
|---|---|
| less effort on enquiry and case creation | 12 minutes |
| less effort on reviewing the order | 10 minutes |
| less effort on the final documentation | 8 minutes |
| Total saving per job | 30 minutes |
| Calculation step | Value |
|---|---|
| 800 jobs × 0.5 hours | 400 hours |
| 400 hours × €55 process cost | €22,000 |
| Calculated potential per year | €22,000 |
From this, the cost items named in Chapter 21 have to be deducted. Whether a viable business case remains at the end depends decisively on how high the integration effort turns out to be – and on whether the time saved is actually used for other work or merely flattens the peak load. Both are legitimate, but they have to be assessed differently in economic terms.
A decision should not rest on the calculated saving alone. In safety-critical processes, better traceability and a lower probability of error can matter more than pure working time. Conversely: anyone who works only with soft arguments gets no budget. Both belong in the same paper.
Completeness requires the question of what happens if the assumptions do not hold. If the time saving halves to 15 minutes per job, €11,000 of calculated potential remains. With annual costs of €12,000 for software, upkeep and governance, the case would be negative – measured by time.
The calculation would then have to run through the other three items: how many invoice deductions were there last year? How often was a setup made based on the wrong plan version? How many extra trips went unbilled? If these figures are not available, that is itself already an answer – and an argument for beginning to capture them.
An AI pilot should be limited, measurable and reversible. Reversible means: it must be possible to end it without the business coming to harm. This condition already rules out some use cases.
At the start there is no technology question, but a stocktaking. Suitable guiding questions:
At the end of this phase there is a defined use case – expressible in one sentence. For example: Incoming customer enquiries are analysed automatically. The system creates a draft job and a list of missing details. Approval is given by the back office.
The prototype should contain only the necessary functions – and really only those. For the use case named, that means: read in the email and attachments, extract eight defined data fields, recognize missing details, create a draft case, show the result for review.
No automatic job placement. No automatic quotation. No unchecked transfer into downstream systems. Each of these functions is technically easy to add – and makes the pilot irreversible.
The test is run with historical and current cases. Historical cases are the more valuable here: their outcome is known, and no one is under time pressure.
| What is measured | Why |
|---|---|
| Recognition rate | How many of the defined fields are found at all? |
| Number of incorrect details | How often is a found value wrong? More dangerous than a gap. |
| Time saving | Against a measured baseline, not against a feeling. |
| Number of corrections required | Determines whether any time is saved at all. |
| Staff acceptance | Is the system used voluntarily when no one is watching? |
| Technical stability | Outages, waiting times, error messages. |
| Critical errors | Cases in which an error could have had consequences. |
| Exclusion cases | Constellations in which the system must not be used. |
The last point is the most valuable yield of the test phase and the one most often forgotten. A pilot that, at the end, describes cleanly where the system does not work has delivered more than one that merely names a success rate.
Set the stopping criteria down in writing before the start. A pilot without a defined stopping condition is never ended – it is only funded further and further, because money has already been invested.
Four groups of metrics are relevant. Anyone who measures only the third will assess the application wrongly – anyone who measures only the first will never be allowed to roll it out.
The number of results taken over unchecked. It shows whether the approval architecture really exists or stands only on paper. If it rises, security falls – regardless of how good all the other figures look. It is at the same time an early indicator: it always rises when the system works so well that no one looks any more.
What matters is that all four groups are measured once before the pilot. Without a baseline, it cannot later be said whether an improvement has occurred – and the discussion shifts to impressions.
Rate each statement with 0 points (not present), 1 point (partly present) or 2 points (sufficiently present). The maximum is 40 points. Answer the questions as things actually are – not as they should be.
| No. | Statement | Points |
|---|---|---|
| Processes | ||
| 1 | Our job processes are described. | |
| 2 | Every job has a unique number. | |
| 3 | Responsibilities and approvals are defined. | |
| 4 | Plan changes are versioned traceably. | |
| 5 | Disruptions and deviations are captured in a structured way. | |
| Data | ||
| 6 | Job data is maintained centrally. | |
| 7 | Material stocks are available digitally. | |
| 8 | Staff qualifications are recorded and up to date. | |
| 9 | Site photos are assigned to projects. | |
| 10 | Inspections are documented digitally. | |
| Documents | ||
| 11 | Bodies of rules and work instructions are held centrally. | |
| 12 | Outdated documents are flagged or removed. | |
| 13 | Documents have owners with professional responsibility. | |
| 14 | Access rights are documented. | |
| 15 | Plan and document versions are traceable. | |
| Organization | ||
| 16 | Management supports the pilot. | |
| 17 | A professional process owner is named. | |
| 18 | IT, data protection and the works council are involved. | |
| 19 | Users are trained. | |
| 20 | Success criteria are defined. | |
| Result | Interpretation |
|---|---|
| 0 – 14 Foundations missing | An AI project would probably create additional disorder. First improve job structure, document storage and responsibilities. That is not a setback but the real work. |
| 15 – 28 Pilot possible | A limited use case with manageable risk is possible. Document analysis, knowledge search or report preparation are recommended. |
| 29 – 40 Good starting position | Alongside assistance applications, integrated processes, mobile documentation and dispatching support can also be examined. |
This checklist is meant to be filled in – once per use case. If, in any one of the five groups, you cannot answer more than two questions, the use case is not yet ready for a decision.
These questions belong in every provider enquiry. A provider who fails to answer several of them in writing is not suitable for a regulatorily observed trade – regardless of the quality of the demonstration.
The following sentences come up in almost every project. None of them is ill-intentioned – all are understandable. And all regularly lead to a project failing or remaining without effect.
A general chatbot knows neither the current jobs nor the approved plans and orders. It produces text but rarely solves the actual process bottleneck. After three months it turns out that a third of the workforce uses it for phrasing emails and the rest not at all.
Better: Choose a specific case and connect the necessary data sources.
For safety-relevant planning, that is not appropriate. Automatic checks can support, but cannot assume professional responsibility – and cannot take it off your hands when it later matters.
Better: Combine check rules, source locations and documented human approval.
An unchecked collection of documents leads to contradictory answers. And because the system resolves these contradictions in fluent German, at first no one notices.
Better: Classify documents, name responsible owners, maintain validity and version status.
Without rules, confidential data is entered into unsuitable systems or results are used unchecked. The wish to experiment is legitimate – it just needs a framework.
Better: Define approved applications, usage rules and training. Two pages are enough.
In traffic safety, completeness, traceability, fewer deviations and better responsiveness count too. Anyone who measures only minutes will switch off good applications and keep bad ones.
Better: Consider economic and safety-related metrics together.
Projects that are too large become hard to steer and deliver results only late. By then the requirements have changed and those involved have lost interest.
Better: Start with a limited process and expand step by step.
IT can operate and connect systems. It cannot judge whether an extracted condition was interpreted correctly or whether a standard-plan proposal fits the local situation.
Better: Name a professional owner for each application – from planning, dispatching or the back office.
Four maturity levels, as of July 2026. The assessment refers to a medium-sized business with the usual software equipment – not to what is technically feasible in a research project.
| Maturity level | Applications |
|---|---|
| Realistic immediately No integration needed, low risk |
summarizing tenders · extracting job data · preparing queries · searching internal documents · comparing document versions · producing draft reports · voice input for site notes · classifying photos · preparing inspection lists · automatically assigning incoming documents |
| With manageable integration effort Interfaces required |
job creation from emails · digital deployment folder · mobile setup and inspection processes · automatic summary of orders · material-requirement lists from job templates · deadline monitoring · route planning for inspection runs · customer status reports · assembling final files |
| Demanding but attainable Presupposes a good data foundation |
target-actual comparison between plan and setup · AI-supported dispatching · forecasting material and staffing needs · automatic detection of faults on photos · linking with traffic and sensor data · dynamic prioritization of inspections · semi-automated conformity checking |
| At present only with considerable caution Safety and liability questions unresolved |
autonomous approval of traffic control plans · independent change of a traffic arrangement · fully automatic assessment of safety-relevant clearances · autonomous decision that a defect is not dangerous · automatic closure of disruption cases · unchecked transfer of AI results into official applications · fully autonomous traffic control by a traffic-management company |
All the applications in the fourth group have one thing in common: with them, human approval falls away. That is no accident. The line between "demanding" and "only with considerable caution" runs not along technical difficulty, but along the question of who signs at the end.
Begin with a process that costs time today, is prone to error or limits growth. The question is not "What can AI do?", but "Where do we lose time, money or proofs today?".
Every application needs a professional owner. IT alone cannot judge the quality of a traffic-safety process – and should not have to be liable for it either.
Define in writing which decisions are always made by qualified staff. This document is short, but it is the foundation of everything else.
A central project and document structure is often the most important prerequisite – and the part of the project that looks least like AI and achieves the most.
Dispatchers, planners, site managers and fitters know the actual bottlenecks. An application without their input is seldom used for long. Where a works council exists, this applies doubly – and legally in any case.
Before an AI is used on ongoing jobs, it should be tested on known cases. That is the only test in which you already know the right answer.
For professionally relevant outputs, it must be traceable which data was used, which result was produced, who reviewed it, which corrections were made and who gave the approval. These five details are, in a dispute, more valuable than any success rate.
After three months at the latest, it should be apparent whether the application brings measurable benefit. If not: stop. A pilot ended cleanly is a result, not a failure.
Look for a bottleneck, not a technology. Name a responsible person, not a project. Define a stopping condition, not a vision. And record who signs.
Artificial intelligence will not change traffic safety through an autonomous planner or a fully self-steering work zone. Anyone waiting for that is waiting for the wrong thing.
The practical change begins in less spectacular, but economically relevant places:
None of these points is worth a headline. Together they change how calmly a business works – and how well it can prove what it has done.
It lies not between companies with and without a chatbot. It lies between companies that continue to scatter their job information across numerous standalone tools, and companies that join their processes, documents and data into a usable operating model.
AI can be a capable assistant in this. It can make work calmer: less searching, fewer queries, fewer loose ends on a Friday evening. But it cannot carry responsibility.
Responsibility stays with the company and with the expert staff. That is not a limitation but the very reason this technology can be used at all in a safety-critical trade.
The answer depends on your job chain, not on a product list. KrambergAI analyses together with you where, in your business, time, money and proofs are actually being lost – and which application changes that.
The result is not a general AI presentation, but a prioritized implementation overview for your business – with effort, expected benefit and a clear order.
| Offering | What it suits |
|---|---|
| AI potential analysis for traffic safety | You know something is possible, but not where to start. |
| Digital job and process check | The process has grown, the handovers are unclear, the data situation is unknown. |
| Company knowledge for traffic safety | Knowledge is tied to individuals; onboarding takes too long. |
| Pilot for enquiry and document analysis | Many enquiries, high effort on the initial screening. |
| Digital deployment and inspection documentation | Proofs are laborious, plan versions ambiguous, inspections on paper. |
| Industry-specific AI employee | Recurring case work is to be relieved on a lasting basis. |
We build digital products that make work calmer. Not faster at any price, but more relieving, more controllable and more provable. Data protection under the EU GDPR and development in Germany are, for us, not extras but a prerequisite – precisely where public clients are involved.
All details were last checked on 17 July 2026. Bodies of rules, deadlines and statistics may have changed since.
This ebook serves as professional and entrepreneurial orientation. It replaces neither legal advice, nor an examination of the specific individual case, nor the professional planning or approval of a traffic-safety measure.
What is authoritative is the applicable statutory provisions, official orders, technical bodies of rules, contract documents and project-specific requirements in each case. The classifications under the AI Act shown here are indicative and not legal advice; the classification of a specific system has to be examined case by case.
All example calculations rest on disclosed assumptions and are to be understood as calculation models. They constitute no promise of achievable savings or results. The practical examples are typified scenarios and not customer references.
The statistical details and legal positions used reflect the state as of 17 July 2026. In particular, the deadlines of the AI Act were subject to change at that time; the Digital Omnibus on AI had been adopted by Parliament and Council, with publication in the Official Journal of the European Union still pending. Check the current position before every decision.
Unchanged distribution within your company and to project participants is expressly welcome. For the use of excerpts in your own publications, we ask for prior agreement.
KrambergAI GmbH
krambergai.com
Title: AI in Traffic Safety at Road Works – A Practical Guide
Status: July 2026
Version: 1.0
Extent: 53 pages, A4
Audience: traffic-management companies, road-construction firms, works depots, infrastructure operators
Would you like to know which of the ten use cases suits your
business? We start with a stocktaking of your job chain – and will also
tell you when an AI project is not the right step at present.
krambergai.com