How to bring generative AI into your company: a 90-day plan

Most companies we talk to sit somewhere between “we should do something with AI” and “we bought licences and not much changed”. A few people use ChatGPT for emails. Someone tried a chatbot on the website. Nobody can say whether any of it saves time, and the work itself runs the way it did two years ago.

That’s normal. Buying access to a model is easy. Changing how a piece of work gets done is the hard part, and that’s where the value is. This guide is a plan for getting from good intentions to one process that runs better with AI, in about 90 days, and then doing it again.

It’s also the hub of our series on bringing AI into a company. Each section points to a longer article if you want the detail.

What generative AI is good for, and what it isn’t

Generative AI means models that read and write text, and increasingly images and speech. They learn from huge amounts of text to predict what comes next. That makes them good at language work and unreliable with facts nobody gave them. What is generative AI explains how they work.

In an ordinary company it earns its keep in five kinds of work:

• Reading messy input: emails, PDFs, scanned forms, call transcripts. Pulling out the order number, the amount, the actual request.
• Sorting: is this email a complaint, an invoice or a question? Which team gets it?
• Drafting: replies, quotes, job ads, a summary of a long email thread.
• Answering questions from your own data: “What does our contract with this supplier say about late delivery?”
• Turning data into words: a monthly report a client can actually read.

It’s weak at exact arithmetic without tools, at facts it wasn’t given, and at giving the same answer every time. And if a simple rule does the job (“invoices over €5,000 go to the finance director”), use the rule. It’s cheaper, faster and never makes things up.

The 90-day plan

From idea to a working process in 90 days

1. 1

   Weeks 1–2

   Pick one process and measure it

   How long it takes today, how often it goes wrong, who owns it.

2. 2

   Weeks 3–4

   Design the pilot

   What the AI does, what a person checks, which data it needs.

3. 3

   Weeks 5–8

   Run it on real work

   One team, real cases, a person reviews every output.

4. 4

   Weeks 9–10

   Measure

   Time per case, error rate, how much people actually use it.

5. 5

   Weeks 11–13

   Roll out, fix or stop

   Decide with numbers, then pick the next process.

The plan is deliberately small: one process, one team, one quarter. Companies that try to “roll out AI” everywhere at once usually end up with many experiments and nothing in daily use.

Weeks 1–2: pick one process and measure it

Start from the work, not from a tool. Ask each team what they do twenty times a week that involves reading or writing something. Good first candidates have:

• volume: it happens daily, not twice a year
• messy input or text output, which rules alone can’t handle
• results a person can check in a minute
• data you can reach (a shared mailbox, a folder, a database)
• an owner who wants it fixed

Pick the first use case

Value →

Quick wins

Big bets

Nice to have

Skip for now

• 1 Order emails into the ERP
• 2 Support email triage
• 3 Quote drafts from past quotes
• 4 Meeting notes
• 5 Fully autonomous sales agent
• 6 AI-generated social posts

Effort →

1. 1 Order emails into the ERP
2. 2 Support email triage
3. 3 Quote drafts from past quotes
4. 4 Meeting notes
5. 5 Fully autonomous sales agent
6. 6 AI-generated social posts

Then measure the baseline before you change anything: how many cases a month, how many minutes each, how often something goes wrong and what that costs. A week of tally sheets or the timestamps from your helpdesk is usually enough. Without a baseline you can’t tell whether the pilot worked, and the next project won’t get a budget.

How to find the first AI use cases worth building has a scoring sheet for comparing candidates. How to measure the ROI of an AI project shows how to measure, with a worked example.

Weeks 3–4: design the pilot

Write it down on one page:

• What the AI does. “Reads incoming order emails and attachments and fills in a draft order in the ERP.”
• What a person does. “Checks every draft before it’s confirmed. Handles anything the AI marks as unclear.”
• What data it uses, where that data lives, and whether you may use it this way (GDPR, client contracts).
• What success means, in numbers from the baseline: “Under 2 minutes per order including the check, no more errors than today.”
• What happens when it fails: the case goes back to the old route, and nobody is stuck.

Collect 30–50 real past cases with the correct result. They become your test set. Every version of the pilot runs against them before real users see it.

During a pilot a person reviews every output. That isn’t distrust; it’s how you learn where the model is reliable. Human in the loop covers how to make the review step fast and how to loosen it as trust grows.

Weeks 5–8: run it on real work

Give the pilot to one team, on real cases, inside the tools they already use where possible. A pilot that lives in a separate browser tab, where people copy text in and paste answers out, gets dropped in the first busy week.

Meet once a week. Look at the cases the AI got wrong, fix the instructions or the data, and run the test set again. Ask users what slows them down. Most improvements here are unglamorous: a better layout on the review screen, a rule that catches a common mistake, a clearer instruction.

Weeks 9–10: measure

Compare against the baseline: minutes per case including the review, error rate, share of cases that went through the new process, and the running cost. Then ask the team whether they’d keep it. If they wouldn’t, find out why before you go further.

Weeks 11–13: roll out, fix or stop

Decide with the numbers on the table:

• Roll out if it saves time without adding errors. Daily use needs more than a pilot: integrations, permissions, monitoring, support. From AI pilot to production lists what changes.
• Fix and rerun if the result is close and you know what to change.
• Stop if it doesn’t pay off. A stopped pilot that cost a few weeks is a cheap lesson, not a failure.

Then pick the next process. The second one goes faster, because data access, review habits and measuring are already in place.

Who does what

AI projects stall more often for lack of an owner than for lack of technology. You need four roles. In a small company one person can hold two of them.

• Sponsor. A manager who decides, frees up budget and data access, and asks for the numbers in week 10. Usually the owner, the operations director or a head of department.
• Process owner. The person responsible for the work today, such as the head of customer service or the office manager at an accounting firm. Owns the baseline, the success criteria and the go or no-go call.
• Users and a champion. The team that does the work. One of them tries early versions, collects feedback and shows colleagues how to use it.
• Builder. Whoever sets up the tool or builds the integration: an internal developer, an IT admin for an off-the-shelf product, or an outside team.

Add the person who answers data protection questions today. They should see the plan in week 3, not the week before launch.

Choosing the tool: assistant, workflow or custom system

There are roughly four levels, and most companies end up using more than one.

How much the AI does on its own

1. Chat assistant People paste text in and copy answers out 1
2. AI inside your tools Features in your helpdesk, CRM, office suite 2
3. AI workflow Fixed steps, the AI reads or drafts 3
4. AI agent Chooses its own steps and tools 4

← A person drives The AI acts alone →

1. 1 Chat assistant People paste text in and copy answers out
2. 2 AI inside your tools Features in your helpdesk, CRM, office suite
3. 3 AI workflow Fixed steps, the AI reads or drafts
4. 4 AI agent Chooses its own steps and tools

A chat assistant (ChatGPT, Claude, Microsoft Copilot or Gemini on a business plan) is the right start for individual work: drafting, summarising, thinking a problem through. It’s cheap and quick to roll out. On its own it doesn’t change a process, because the work still depends on each person remembering to use it. Pair it with a short AI usage policy, some training and a few good prompt templates.

AI features in tools you already pay for are worth checking before you build anything. Many helpdesks, CRMs and accounting packages now draft replies, summarise tickets or read invoices. If the feature fits your process, use it.

An AI workflow is a fixed sequence of steps where one or two of them use a model. An email arrives, the model extracts the order, rules check stock and credit limits, a person approves, the ERP is updated. For a typical company this is where most of the measurable value sits, because it changes the process itself. Our post on workflows, AI workflows and agents explains the difference in detail.

An AI agent decides its own steps and calls tools in a loop until it reaches a goal. It helps when the path can’t be written down in advance. When it can, an agent is slower, costs more and is harder to test. Read what an AI agent is and when to use one before you build one.

Five questions decide between buying and building:

1. Does an off-the-shelf tool fit the way you actually work, or would you bend the process to fit the tool?
2. Does it need data from your own systems, and can the tool reach it?
3. What will it cost at your volume in two years, not in the trial month?
4. May the data go to a public AI service, or must it stay in the EU or on your own servers? See self-hosted LLMs or API models.
5. Is this process part of what makes you different from competitors?

Build or buy works through these questions. What an AI project really costs covers build, running and maintenance costs. For reference, our own projects reach a first version in 6 weeks to 3 months, and one process connected to one or two systems typically costs €5–10k.

Is your company ready?

You don’t need a data warehouse or a fifty-page AI strategy. You need a few practical things in place.

Readiness check

• One process with an owner Someone who wants it fixed and can decide
• A baseline Volume, minutes per case, error rate
• Data you can reach Mailbox, folder, database or API, not paper
• Permission to use it GDPR basis, client contracts, internal rules
• Real cases to test with 30–50 past examples with the right answer
• A review rule Who checks the output, and when
• Time from the team A few hours a week during the pilot
• A decision date When you roll out, fix or stop

The most common gap is data. If the information lives in people’s heads or in paper folders, start by getting it into a system. That is often a worthwhile project in its own right, with or without AI.

Risks and how to keep them small

Wrong answers. Models write fluent text that can be wrong. Give the model the source material instead of relying on what it “knows”; RAG explained shows how. Check outputs against rules, and keep a person on anything that goes to a customer or moves money. Why AI makes things up has the design techniques.

Data going where it shouldn’t. Some consumer AI accounts may use what you type to train their models. Use business plans or APIs with proper contract terms, decide which data may go where, and write it down. For personal data, GDPR applies as usual. AI, GDPR and the EU AI Act covers what a European company needs to do. One duty already applies to almost every company that uses AI: since February 2025, the AI Act expects it to take steps on its staff’s AI literacy, meaning people understand what the tools do and where they fail.

Agents doing too much. An agent that reads email and can also send payments can be tricked by a hidden instruction inside an email. Give it the narrowest permissions that work. Keeping AI agents safe explains the defences.

Costs creeping up. Usage-based pricing looks tiny in a pilot and can surprise you at full volume, especially with agents. Work out the cost per case early and check it monthly.

People stop reading. When the output is usually right, reviewers start approving without looking. Spot-check approved items and track how often reviewers change the draft.

Why rollouts stall

Where European companies are

• 20% of EU enterprises used at least one AI technology in 2025
• 55% of large enterprises did
• 71% of those that considered AI but didn’t use it named a lack of expertise

Use is growing, but smaller firms are well behind, and the main brake isn’t the technology. In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after the proof-of-concept stage by the end of 2025. It named poor data quality, weak risk controls, rising costs and unclear business value.

In our experience the reasons look more ordinary up close:

1. The tool came first. Someone bought licences or built a chatbot, then went looking for a problem.
2. No owner. The pilot belongs to IT or an “innovation team”, not to the person whose numbers would improve.
3. No baseline. Nobody can show it saves time, so the next step never gets a budget.
4. Toy data. The demo used ten clean examples. Real emails have attachments, typos and three languages.
5. No integration. People copy and paste between the AI and their real systems, then quietly stop.
6. The team wasn’t involved. People see it as extra clicks or as a threat to their jobs, and go back to the old way. Talk about the job question honestly and early.
7. Waiting for perfect. The model handles 90% of cases well, and the project waits for 100% instead of letting a person handle the rest.

Each of these is cheaper to prevent than to fix. The 90-day plan above is built around them.

What to do next

• This week: list five tasks your team does daily that involve reading or writing. Pick the one with the most volume and an owner who cares.
• Next week: measure it. Cases per month, minutes per case, errors.
• Within a month: decide between an off-the-shelf tool and a workflow, and start a four-week pilot with a person checking every output.

If you’d like help with the first process, our internal tools and workflow projects start exactly there. A 30-minute call is usually enough to know what to build first and roughly what it will cost.