From AI pilot to day-to-day operations: four questions

Updated on 8 October 2026: we have added the transition period for Article 50(2).

The pilot is over, operations have not yet begun

A lot is written about choosing AI tools. About rolling them out, too. About everyday life afterwards, almost nothing: the weeks and months in which the AI is simply supposed to run without anyone paying it special attention.

Perhaps you know the moment. The pilot was convincing, the budget for routine operation has been approved, the project team disbands. And in the next meeting someone asks: who is actually looking after this now?

The same pattern often shows up in this phase. The pilot had a committed person with time for it, and it ran under observation and with selected cases. Exactly these conditions disappear as soon as the project becomes routine. Why things often stall between testing and operations is covered in the article on moving from testing to operations.

A successful pilot proves that something can work. It does not prove that it works in operations. In operations there is no project plan and no special attention any more. There is Monday morning, full inboxes and employees with other tasks.

After an AI rollout, four questions often remain open. Who looks after the AI? Who notices errors? What does it cost to run? Who is liable if something goes wrong?

An open worked example: the AI that pre-sorts quote requests

First, a note: this company does not exist. The example is constructed and is based neither on a client nor on our own operating data. It shows the points at which operations can break down.

The starting point. A manufacturer of industrial components receives enquiries every day by email and through the contact form. They include catalogue parts, custom parts made to drawing, spare part orders and messages that are not enquiries at all. Until now, the internal sales team has sorted them by hand, followed up on missing details and passed custom orders on to engineering.

What the AI does. An AI workflow takes over the first step. By this we mean a sequence of work steps that a language model carries out according to fixed instructions. It assigns each enquiry to a category, checks whether quantity, material and requested date are given, and drafts a reply.

How the pilot went. The head of internal sales looks after the pilot; we call this person the pilot champion from here on. They get a fixed time budget for it. For a few weeks, only the central enquiries inbox runs through the AI. They write the work instructions, collect good and bad example replies and read every draft before it goes out. They correct wrong assignments immediately. The team is happy, and management approves routine operation.

What changes in operations. To make the relief worthwhile, the team decides that the AI will answer pure information requests on its own from now on. Everything else still goes to a person as a draft. The contact form is added.

So far, this is a good story. But it leaves four points open:

Looking after it. The time budget disappears. The work instructions sit in a folder that only one person really understands. Who adjusts them when a new product line is added?
Errors. Instead of one experienced person, several colleagues now approve, under time pressure. And some of the replies nobody reads any more. Who notices a wrongly sorted custom order before the customer chases it?
Costs. The pilot budget covered the licence and setup. The pilot champion's hours appeared nowhere. What does the workflow cost if someone else has to put in those hours?
Liability. A draft gives a delivery date that production cannot meet. It is approved without checking, and the customer plans around it. Who is accountable for that?

None of these points is a technical defect. What has changed is the environment: less attention, more people involved, no fixed responsibility.

Question 1: Who looks after the AI when the pilot champion goes back to their actual job?

In the pilot, one person made corrections whenever the AI got something wrong. This work was in no plan. In operations, the person returns to their actual tasks. The work on the AI remains, and nobody is explicitly responsible for it any more.

This is no side issue. The study "The GenAI Divide: State of AI in Business 2025" by MIT (Project NANDA, July 2025) asks why many AI pilots show no measurable effect. Its finding: the cause lies in the approach rather than in the quality of the models. Tools that fit poorly into workflows and do not learn get stuck. Who looks after the AI is part of exactly this integration.

Three roles that usually sit with one person in the pilot

"Looking after the AI" means three different tasks. Name them separately, even if two of them end up with the same person.

Business ownership. Who owns the process? In the example: who decides which enquiry belongs in which category, and who is accountable when an order lands in the wrong pile? This role sits with the business unit rather than with IT.

Technical support. Who takes care of access, the connection to the inbox and the CRM, and the vendor's updates? If the vendor changes its model, someone has to notice.

Content maintenance. Who keeps work instructions, examples and business context up to date? New products, changed prices, a new owner for key accounts: the AI only learns about these if someone enters them. This role is easy to overlook because in the pilot it ran along unnoticed.

Whatever gets maintained needs a name

We make the thing being maintained tangible with the three-folder principle. Everything an AI needs for a task sits in three places: the business context (who you are, what you offer, and for whom), the data (such as price lists and product sheets) and the work instructions, with examples of good and unusable results.

These folders are a working tool like your quote templates. They are not a project deliverable for the archive. They need a fixed place and a responsible person. If a folder goes out of date, the AI works with outdated information.

The test question for your company

Is looking after your AI pilot written into a job description, a target agreement or a role plan? Or does it exist only in one person's head?

If only in someone's head, a holiday, an illness or a change of job is enough, and nobody maintains the AI any more. Each of the three roles needs a name, a deputy and a fixed time budget. That is not easy: in the Bitkom survey on artificial intelligence 2025, 51% of companies name a lack of staff as an obstacle to using AI.

Question 2: Who notices errors before the customer does?

In the pilot, errors stand out because someone looks at every case. In operations, this attention falls away. Then the only people who notice errors are those who look for them on purpose. Or the customer.

Back to the worked example

An enquiry contains a concrete order, but it is briefly worded and has an unusual subject line. The AI classifies it as an information request. Since routine operation began, the AI answers those on its own: a friendly standard reply with the product brochure attached.

In the pilot, the pilot champion would have seen this the same day. Now nobody sees it. The customer waits for a quote, gets a brochure and orders elsewhere. The system says: enquiry answered. Errors like this produce no error message.

Four levers

Set approval levels. For each task, decide what the AI may do on its own and what only goes out after human approval. In the example: the AI may sort enquiries and answer pure information requests on its own. Anything that looks like an order, a price or a complaint goes to a person as a draft. Binding statements on prices, delivery dates or availability do not belong in automatic replies. The line belongs where an error costs money or trust.

Spot checks instead of full control. It is enough to regularly review a fixed number of randomly drawn cases, above all from the part the AI handles on its own. That is where the wrongly sorted orders end up. The check is not about whether the reply is well worded. The question is: would an experienced colleague have classified the case the same way?

Create a feedback route. The people who work with the results every day see errors first. They need a fixed place for reports, a responsible person and a short response. Without this route, people quietly correct errors by hand. The problem stays; nobody sees it any more.

Expect gradual deterioration. AI results can change without anyone in the company having changed anything. A study from 2023 gave two versions of GPT-4 the same task: the March 2023 version identified prime numbers with 84% accuracy, the June 2023 version with 51%. The authors therefore recommend continuous monitoring. The same applies to your own data: if your range or prices change, the instructions may no longer fit.

Gradual deterioration

84% → 51%

Accuracy of GPT-4 at identifying prime numbers, March 2023 version compared with June 2023 version. Results can change without anyone in the company having changed anything.

Source: arXiv, 'How is ChatGPT’s behavior changing over time?' (2023)

A review rhythm instead of a monitoring project

You need neither a dashboard nor an IT project for this. For a low-risk task like this one, a fixed rhythm with a named owner is enough to start with. The more serious the possible consequences of an error, the tighter the control should be:

Weekly: a random sample from the cases the AI handles on its own. Briefly note the findings.
Monthly: go through reported errors and decide whether instructions, examples or approval levels need adjusting.
With every change: if the vendor changes the model, or your range, prices or processes change, known test cases are run again. In the example, that would be around ten typical enquiries with a fixed classification, including at least one briefly worded order. If they come from real customer emails, remove names and contact details first.

It takes little time, but someone has to plan for it.

Question 3: What does the AI cost once it is running?

In the pilot, the bill was manageable: licence or usage, perhaps some help with setup. The pilot champion's time appeared on no invoice. That is why pilot costs are no yardstick for operating costs. How running costs fit into the business case is shown in the case for augmentation rather than job cuts.

Large organisations get this wrong too. Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The analyst firm named poor data quality, inadequate risk controls, escalating costs and unclear business value as the reasons. In October 2024, Gartner added that anyone who does not understand how AI costs scale with use can get their cost calculation wrong by 500 to 1,000%.

So calculate item by item.

The items you should put a figure on yourself

1. Usage-based model or licence costs. Many AI services charge by volume: number of requests, length of texts, number of users. In the example, the pilot ran one inbox for a few weeks. In operations, the contact form and peak season are added. Your question: how does our vendor charge, and what happens to the bill when volume rises?

2. Internal support time. The tasks from questions 1 and 2: business ownership, technical support, spot checks, feedback. No other item stays as reliably invisible in the pilot. Your question: how many hours a month, by whom, at what internal hourly rate?

3. Maintaining instructions and data. The three folders go out of date. If the company introduces a new product line, the AI only sorts enquiries about it correctly once someone has adjusted the instructions and examples. Your question: who maintains the folders, on what occasion, and how often?

4. Adjustments when the model, prices or vendor change. If the vendor swaps the model or changes its prices, or you switch vendors yourself, instructions have to be reviewed, test cases repeated and connections possibly rebuilt. Your question: who checks after a change whether the results are still right, and how much time do we plan for it?

5. Training new employees. New colleagues have to learn to check AI drafts instead of simply adopting them. The AI Act addresses AI literacy too (more on this in question 4). Your question: how long does onboarding take, and who does it?

Put a rough estimate and a name next to each item. An item without an owner only gets noticed when the invoice arrives or quality drops.

Question 4: Who is liable if something goes wrong?

In the example, a draft reply gives the wrong delivery time, and the customer plans around it. Who bears the consequences? In practice, the customer calls you, not the AI vendor. The AI supplies the draft; your company sent it. Who is legally accountable for what in the end, and whether you have recourse against the vendor, is a legal question. We do not settle it ourselves in projects; we have it reviewed.

A case from Canada shows how a court can see it. The Civil Resolution Tribunal, a dispute resolution tribunal in British Columbia, ruled in 2024 on an airline's chatbot that had given wrong information about special fares (2024 BCCRT 149). The airline argued that the chatbot was a separate entity. The tribunal did not accept this and ruled in favour of the customer.

We do not assess what the case means for Germany or Austria. For operations, we draw a simpler lesson: telling customers "the AI wrote that" is not an answer that makes you look good.

What the AI Act says and what we make of it for operations

AI literacy (Article 4). According to the text of the regulation, Article 4 on the AI literacy of staff has applied since 2 February 2025. The AI Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, revised the article. You will find the wording there. In projects, we ask two things at this point: does every person who approves draft replies know where the AI typically gets things wrong? And is there a record of who was trained, when and how? In our experience, a short training session on the real workflow achieves more than a general AI course.

Transparency (Article 50). According to the regulation, the transparency rules of the AI Act have applied since 2 August 2026. According to the European Commission, providers of AI systems placed on the market before that date only have to meet the obligation to mark AI-generated content (Article 50(2)) from 2 December 2026. In the example, the AI answers information requests on its own. We test such points through the customer's eyes: does the customer notice that they are writing to an AI? We have legal counsel review who in the company is responsible for this and which notices are needed, and we do the same for a chatbot on the website.

Sensitive use cases. For recruitment or credit checks, we always have legal counsel review which rules apply. Regardless of that, in every AI operation we define who oversees the results, how operations are monitored and which logs are kept.

Liability and data protection: what we hand over for legal review

The European Commission withdrew its draft AI Liability Directive in 2025, the IAPP reports. The new Product Liability Directive (EU) 2024/2853 explicitly names software, and member states have until 9 December 2026 to transpose it. We do not answer the question of what this means for a specific AI workflow ourselves. We make sure the question reaches your legal counsel early.

Then there is data protection. Enquiries by email and contact form contain personal data, such as the names and contact details of contact persons. Before real enquiries run through the AI, we therefore go through the usual points with the data protection officer: legal basis, data processing agreement with the vendor, retention and deletion periods, possible data transfers to third countries. The handover checklist below does not replace this review.

What this means for your operations

Three questions help in operations:

Is it defined which AI results go out without human approval?
Can you show which training measures took place?
Can you trace afterwards which draft came from the AI and who approved it?

Our deployer check shows how we go through a company's use of AI as a whole.

From pilot to operations: a handover checklist

Go through this list once before you officially hand over the pilot, and do it for a specific workflow instead of for "the AI" in general. Every question can be answered with yes or no. A no does not mean you have to stop. It shows where someone still has to make a decision.

Responsibility

Is one named person responsible on the business side, meaning the person who decides how enquiries are sorted and answered?
Is it settled who provides technical support for access, connections and updates?
Is it defined who maintains the business context, data and work instructions?
Are these tasks written into a role description, with a time budget and a deputy?

Approval and control

Is it set out in writing what the AI may do on its own and what only leaves the company after approval?
Is there a fixed review rhythm with random spot checks and a responsible person?
Do employees know where to report an AI error, and does the report reach someone?
Are known test cases run again after every change to the model, instructions or data sources?

Costs

Does every running cost item have an owner, from usage through support time to training?
Is the support time that, in the pilot, came from one person's commitment now recorded as effort?
Is there a threshold above which rising costs trigger a review of the workflow?

Documentation and competence

Are the work instructions documented well enough that another person can understand and change them without the pilot champion?
Is there a simple record of who learned what and when?
Has data protection been reviewed for the workflow, from the legal basis to the data processing agreement?

Exit and shutdown

Is there a plan for how the workflow is shut down, and is it settled who decides that?
Does the process keep running if the AI is down for a week? Who sorts the enquiries then?
Is it defined what happens to data, instructions and examples in the event of an exit?

How many noes you accept is your decision. Responsibility and approval, however, need to be settled before routine operation. The rest can go on a list, but each item with a name and a date.

What runs differently on Monday morning

An AI workflow in routine operation hardly gets noticed, because the open questions have been answered. In the example, Monday morning then looks like this:

Responsibility: the head of internal sales knows that the workflow belongs to their area. The technology and the three folders each have a named person.
Errors: replies on orders, prices and complaints never go out without human approval. On Mondays, someone draws a random sample from the enquiries the AI answered on its own the week before.
Costs: usage, support and maintenance are in the budget with an owner. If the vendor changes its prices, that is a budget question and comes as no surprise.
Accountability: everyone knows that the AI supplies drafts and the company decides.

None of this is a lot of work. None of it runs by itself.

If the checklist gives you more noes than yeses, you know where the work after the pilot lies. If you would like to go through it with someone who looks at operations rather than at the next tool: in the AI Diagnostic we carry out exactly this stocktake. Who will look after it in your company from Monday?

Frequently asked questions

When is an AI pilot ready for routine operation? When you can answer the four operating questions with names and dates: looking after it, error control, running costs, accountability.

Do we need a dedicated position to run AI, or is an additional role enough? For a single workflow, an additional role can be enough if it is explicitly assigned and has a fixed time budget.

How do we check the quality of AI results without checking every case? With approval levels, regular random spot checks, a reporting route for employees and repeated test cases after every change.

Which running costs arise after the rollout that were not visible in the pilot? Above all the internal support time, plus maintaining instructions and data, adjustments when the model or vendor changes, and training. Put a figure on each item yourself.

Who is responsible when an AI makes a mistake in the company? Organisationally, the company that sends the draft: a named person approves, and there is a record of who approved what. Your legal counsel clarifies who is legally accountable for what and whether you have recourse against the vendor.

Where can we find reliable information on the AI Act? In the text of the regulation on EUR-Lex and in the amended version under the AI Omnibus. According to the regulation, Article 4 on AI literacy has applied since February 2025, and the transparency rules in Article 50 since August 2026. Your legal counsel clarifies what applies to your use.

This article describes our practical perspective. It does not contain a legal assessment; for that, we refer you to the linked primary sources and your legal counsel.

The AI pilot was a success. Now what? What has to happen in operations after the rollout