What an AI consultancy should deliver in the first 90 days

In the first 90 days, an AI consultancy should put one working system into production and measure it against success criteria you agreed before the build. That means a benchmark made from your own data, a first working version within weeks, and a system running inside your architecture by the end of the quarter. By day 90, your team should own the code and know how to run it. If the quarter only ends with a roadmap and a demo, you paid for the easy part.

The gap between that bar and today’s results is wide. In Deloitte’s 2026 State of AI in the Enterprise report, only 25% of respondents had moved 40% or more of their AI pilots into production. Yet 54% expected to reach that level within three to six months.

What should an AI consultancy deliver in the first 90 days?

It should deliver five things, in this order:

1. Success criteria: Agreed in writing before anyone builds, with a benchmark made from your own data. Every later demo gets judged against it.
2. A working first version: Ready within weeks, running on real data from your systems and used by the people who will own it.
3. Production foundations: Testing, security, monitoring and operations built in from the start, so the first version does not need a rebuild to go live.
4. A system in production: Deployed in your own cloud and connected to the tools your teams already use.
5. A full handover: Code, tests, agent instructions, runbooks and dashboards, plus the knowledge to run and change them.

The order matters. A consultancy that builds before the success criteria are signed will define success afterward, usually around whatever it managed to build.

How long does it take to get an AI pilot into production with an AI consultancy?

For one well-scoped use case, plan on about a quarter from kickoff to a system in production.

The first working version should arrive within weeks. The rest of the quarter goes to what a pilot skips: real data volumes, integration with your core systems, a security review and the people who will run the system every day. Programs that span several business units take longer. Each new use case should get faster, since it reuses the same foundations.

Unrealistic timelines do more damage than slow ones. In a Gartner survey of 782 infrastructure and operations leaders, run in late 2025 and published in April 2026, only 28% of AI use cases fully succeeded and met ROI expectations. Of the leaders who reported a failure, 57% said they had expected too much, too fast.

Here is what a realistic quarter looks like, and what should worry you at each stage.

New whitepaper

Pilots aren't AI strategy.

Read Hard-Wired ↗

How do you measure the ROI of an enterprise AI program before it scales?

You measure it on the first system, against a baseline set before the build, and you scale only when that number moves.

Pick one metric close to the P&L, such as cost per case, time to quote or revenue per customer. Then, capture its current value from your own data before anything gets built. Without that baseline, every result after launch is an opinion.

Expect the costs to come first. A short-term ROI test undervalues AI because the second and third use cases reuse the data pipelines, integrations and checks the first one paid for. Judge the first system on whether it moves its metric, and judge the program on how quickly each next system reaches production. Our analysis of why enterprise AI adoption stalls at scale covers the organizational side of that curve.

The largest gains often come from the system around the model. For a global industrial manufacturer, we rebuilt how a troubleshooting agent finds information, and its answers became 2x more accurate with the same AI model. For the same client, our systems now process about 45,000 support emails every night.

What should you ask an AI consultancy before the 90 days start?

Ask questions that force a commitment on each deliverable:

• What are the success criteria, and how will you build the benchmark from our data?
• When will our people use the first working version?
• What changes between that version and the production system?
• Where will it run, and who owns the code, the data and the outputs?
• Do model or cloud vendors pay you incentives that could steer the technology choice?
• What do we receive at handover, and could our team run the system without you?

Vague answers on day zero become missed deadlines on day 60. For the full set of criteria and red flags, see our guide on how to choose an applied AI company.

What happens after day 90?

The second system should reach production faster than the first, and that speed is your real return on the first quarter.

BOI is an AI deployment studio. We build production-grade AI systems inside your architecture, independent of any model or cloud provider, and we hand everything over. Pick the one process where a live system would change your numbers this quarter, then hold your partner to day 90.