On this page
Every leadership team we speak with has run at least one AI pilot. A chatbot for support, a tool that drafts sales emails, a summary of meeting notes. The demo went well, people were impressed, and six months later very little has changed in how the business runs.
They’re not alone. In 2025, MIT’s NANDA initiative reported that about 95% of the enterprise generative AI pilots it studied had no measurable impact on profit and loss. The study was based on interviews, surveys and a sample of public deployments, so the exact figure deserves some caution. The pattern behind it matches what we see in the market: the models work, but the projects around them aren’t built to reach production.
This article sets out the 90-day roadmap we use to move one AI use case from idea to a live system with a measured result.
Why pilots stall
When we review stalled AI initiatives, the same causes come up again and again:
- No owner. The pilot belonged to an innovation team or an enthusiastic individual, not to the leader of the process it was meant to improve.
- No metric. Success was defined as “people like it”, not as a change in response time, cost per ticket or pipeline created.
- No integration. The AI lived in a separate tool. Using it meant copying data in and out, so people stopped.
- No safety net. Nobody could say how often it was wrong, so nobody trusted it with real customers.
None of these are AI problems. They are delivery problems, and they can be solved with ordinary project discipline.
Before day 1: choose the right use case
The single biggest factor in success is the choice of use case. A strong first candidate has four things:
- A business owner who wants the result and will change the process to get it.
- A metric that already exists, such as hours spent, response time, conversion rate or cost per transaction.
- Data you already have, in systems you control, such as your CRM, help desk or document store.
- A tolerable cost of error. For the first project, avoid anything where one mistake could cause legal, financial or reputational damage.
Good first projects for B2B companies often include account research before sales calls, routing and summarizing inbound requests, drafting first replies to support tickets, and extracting data from documents such as contracts, invoices and orders. Our article on AI use cases for sales and service teams covers more.
Days 1 to 30: design and baseline
The first month is about agreeing what “good” means and preparing the ground.
- Map the current process. Who does what today, in which systems, and how long it takes. This becomes your baseline.
- Define success in numbers. For example: “Reduce research time per account from 20 minutes to under 5, with reps rating the brief useful in at least 8 of 10 cases.”
- Collect real examples. Gather 50 to 200 real cases with the correct outcome, including awkward ones. These become your test set. We explain why in how to test an AI automation.
- Check data access and security. Confirm which systems the AI needs to read and write, who approves that access, and where data is processed.
- Decide the architecture. Most use cases are best served by a structured workflow that uses AI at specific steps, not by a fully autonomous agent. See workflows vs agents.
Output of month one: a one-page design everyone has signed off, a baseline and a test set.
Days 31 to 60: build and test
The second month is where the system is built, in small increments that the process owner sees every week.
- Build inside existing tools. Results should appear where people already work, such as the CRM, help desk, Slack or email, not in a new interface they have to remember to open.
- Test against the test set after every change, and track accuracy, cost per task and speed.
- Add review steps where mistakes would be costly. People approve before anything leaves the building. Our guide to human-in-the-loop design explains how to keep reviews fast.
- Run in shadow mode. For one to two weeks, the AI runs alongside the existing process without affecting it, so you can compare its output with what people actually did.
Output of month two: a working system that meets the agreed quality bar on real data.
Days 61 to 90: launch and measure
The final month turns a working system into a working process.
- Launch to a first team, not the whole company. Train them, explain what the system does and doesn’t do, and make it easy to give feedback.
- Monitor quality and cost every week: sampled reviews, user feedback, error alerts and running costs.
- Measure against the baseline. Compare the agreed metric before and after, honestly.
- Document and hand over. The process owner should know how the system works, who maintains it and what to do when it misbehaves.
Output of month three: a live system, a measured result and a clear decision: scale it, improve it or stop it.
What happens after day 90
A successful first project does more than deliver its own result. It gives the organization a repeatable way of working: a template for choosing use cases, a testing approach, a security review, and a team that has done it once. The second project typically moves faster than the first, and the third faster again.
That is the real return on a well-run pilot: not one clever tool, but the ability to put AI into production repeatedly and safely.
Questions to ask before you start
- Who owns the result, and what number will they report to leadership?
- What happens today when this process makes a mistake, and what should happen when the AI does?
- Which systems does the AI need to read from and write to?
- How will we know in week 6 whether we are on track?
If you can answer these, you are ready to start. If you can’t, that is where the first two weeks should go.
We run AI implementation projects end to end, from use case selection to a live, monitored system. See our AI and automation services or book a call to discuss your first use case.


