Introducing AI in a small office: the first 90 days

In a small design and engineering practice or construction company, the start works if you choose one process, record the minutes per case for a week and then try it out for four to six weeks with two to three people. Then you re-measure with the same log and decide. You do not need success figures from other offices for this, and none are given here.

1. Week 1 Measure One process, one week, note the minutes per case
2. Week 2 Rules and tool Contact person, data rules, collect concerns, then choose a tool
3. Week 3 Training and start Short training per role, pilot with two to three people
4. Weeks 7 to 9 Re-measure After four to six pilot weeks repeat the same measurement
5. Week 10 Decide Continue, adjust or stop, hear the team
6. Week 13 Write it down Record the rules, pick a second process or pause

Text version of the graphic

• Week 1: Measure. One process, one week, note the minutes per case
• Week 2: Rules and tool. Contact person, data rules, collect concerns, then choose a tool
• Week 3: Training and start. Short training per role, pilot with two to three people
• Weeks 7 to 9: Re-measure. After four to six pilot weeks repeat the same measurement
• Week 10: Decide. Continue, adjust or stop, hear the team
• Week 13: Write it down. Record the rules, pick a second process or pause

Which process first?

The first process should come up often, consist mostly of reading and assembling, and be checked by a person at the end. For the first round, also choose a process without particularly sensitive data. Depending on the company, these candidates suggest themselves:

Company	First process	More detail
Design and engineering practice, project manager	Meeting minutes after the meeting	Meeting minutes with AI: from recording to sign-off
Contractor	Incoming invoice through to approval, or checking a variation	Checking interim payment applications: invoice, BoQ, measurement, Checking variations: variation, BoQ and contract side by side
Property owner, housing company, management	Searching the old records, or the monthly report	Capturing knowledge: old projects as a reference library, Prepare a monthly client report with AI in project management

Decide on exactly one process. Anyone who starts three at once cannot judge any of them at the end.

Week 1: the measurement log

A case is one set of meeting minutes, one invoice or one variation. How to keep the log is described in Measure workflows before introducing AI: log one week. The people who handle a case should do the measuring, because only they know when they start and stop. The later contact person collects the slips and adds them up on Friday. One line per case with these details:

• date, project and case, for example “Variation 07, drywall lot”,
• size by your own definition: small, medium or large,
• start and end,
• of which minutes for searching and assembling,
• of which minutes for rework after queries,
• who handled it.

An example with invented data: variation 07, medium, from 14.05 to 16.05, of which 40 minutes assembling the variation, the bill of quantities (BoQ) and the correspondence, handled by the project lead. The case therefore takes 120 minutes, and a third of that is assembling.

Agree the way of counting beforehand. Does finding the documents count, or does the case only begin with the check? If fewer than about five cases arise in a week, extend the measurement to two weeks. That is a suggestion, not a standard.

Separating assembling from judging has a purpose. The AI is more likely to take over the first part. If that part is small, the process is not worth using as a starting point.

Week 2: contact person, data rules, tool

Name a fixed contact person. They collect questions, try out tasks and pass on findings. Without them every question ends up with the managing director, and the pilot becomes a sideshow. The managing director decides on the process, the period and the data rules, and holds the date of the re-measurement.

Settle this week which data may go into which tool. In 2024 the Data Protection Conference (Datenschutzkonferenz, DSK) published guidance that serves as a checklist for selecting, implementing and using AI applications. Its focus is on language models. Further papers have followed, for example on technical and organisational measures for developing and operating AI systems (June 2025) and on generative AI with the RAG method (October 2025). The BSI (the German Federal Office for Information Security) sees its paper on generative AI models as a possible basis for a systematic risk analysis. None of these papers replaces your data protection officer’s look at your case. What a rule for employees should contain is described in AI policy for the office: what it should contain. Why bans alone do not hold is explained in Shadow AI: when staff use private chat services. Consumer chat services are out of the question for client and project data.

Only when the process and the data rules are settled do you choose the tool. It has to fit the data rules, not the other way round.

Also collect the concerns in the team, openly and if possible anonymously, and answer them before the start. The evaluation of the pilot assesses the process, not individuals. If there is a works council (Betriebsrat), involve it from the beginning. What rights it has is for a legal adviser to clarify.

Week 3: training and start

Train briefly and per role. A project lead who checks variations needs different exercises from someone in accounts who prepares invoices. As a suggestion for the pilot: one hour of shared kick-off on the rules and one hour per role, which is the short form of the role module, with tasks from the measured process, but with invented or redacted documents. The contact person collects questions from practice in the weekly meetings. The training per role is described in AI training by role: content for the office and the site: a kick-off of one hour, a role module of one to two hours and, after two to four weeks, a follow-up session. It follows if you continue the pilot.

You also practise what the AI cannot do. Every output is a draft that a person reads and approves. Figures are recalculated by a spreadsheet or a program, not by the language model.

The training is also a measure within the meaning of Article 4 of the AI Regulation. Under it, providers and deployers of AI systems must take measures to support the development of the AI literacy of their staff, without guaranteeing a particular level for each person. Whether your office is a deployer within the meaning of the regulation is for a legal adviser to clarify in the individual case. Which stages of the regulation matter to an office that only uses AI is summarised in EU AI Act: which dates matter for offices that only use AI, and the position on Article 4 is described in AI policy for the office: what it should contain.

The pilot: four to six weeks

The two to three people work with the tool on that one process, and everything else runs as before. The contact person keeps a list: what worked, where was rework needed, what question came up? Meet once a week for 15 minutes, otherwise the pilot fizzles out. Note minutes and rework during the pilot as in week 1, so that you can compare in the re-measurement.

Do not intervene during this time to improve the result. A poor draft is a finding and belongs in the log. Choose four weeks for a process that comes up often, and six for one that arises only a few times a week.

Re-measuring and deciding

Measure exactly as in week 1: the same process, the same way of counting and, if possible, the same people and the same size classes. Include the rework, because a draft that saves ten minutes and costs twenty minutes of correction is no gain. Compare separately how many minutes the assembling and how many minutes the rework take per case and size class.

A measurement from a few weeks and two to three people is small. It is enough for a decision within your own company, not for extrapolating to the whole office. How to turn the measured minutes into a calculation is shown in Is AI worth it in a small office? A calculation with assumptions. Hear the team and then decide: continue, adjust or stop.

Writing it down by week 13

Write on one page what applies: which process, which tool, which data, who the contact person is and how things are looked up. Then decide whether a second process follows or a pause makes sense.

What you can do now

1. Choose the one process and fix the measurement log for next week.
2. Name the contact person and the two to three pilot participants.
3. Collect the concerns in the team before a tool is chosen.
4. Clarify the data rules for the pilot with your data protection officer.