What a 90-day AI roadmap looks like for a professional services firm

Every AI Executive Assessment I run ends the same way: a written report and a 90-day roadmap. Prospects always ask what the roadmap actually contains. Fair question. So here is one, in full.

The firm below is a composite — a pattern I’ve seen repeatedly across professional services work, both in my advisory practice and in delivery projects with Last Rev, the platform engineering firm I co-founded. It is not a named client and not a disguised one. Picture a firm of roughly 100 people. Document-heavy: proposals, engagement letters, deliverables, client files. A few partners already use ChatGPT on their personal accounts. There was one AI pilot last year. Nobody can tell you what happened to it.

That last detail is the norm, not the exception. MIT’s State of AI in Business research found that about 5% of enterprise AI pilots achieve rapid revenue acceleration — the rest stall with little measurable P&L impact. S&P Global’s 2025 survey data is blunter: the share of companies abandoning most of their AI initiatives jumped to 42%, up from 17% the year before, and the average organization scrapped 46% of AI proofs-of-concept before production. Meanwhile Thomson Reuters’ Future of Professionals report found only 22% of firms have a visible, defined AI strategy — and the ones that do are twice as likely to see AI-driven revenue growth.

The lesson I take from three years of shipping production AI: the failures aren’t caused by weak models. They’re caused by scattered pilots with no owner, no baseline, and no sequence. A 90-day roadmap is the sequence.

Days 0–14: the assessment

Two weeks, fixed end date. Five things happen:

Stakeholder interviews. Three to five conversations — managing partner, operations lead, a senior practitioner, someone from IT if IT exists. I’m listening for where hours actually go, not where people think AI belongs.

A workflow and data-source map. Which workflows consume the most billable and non-billable hours, and which systems hold the documents those workflows depend on. Usually it’s four or five systems and one shared drive nobody wants to talk about.

An AI usage and risk review. What employees are already doing with AI, governed or not. There is always shadow usage. At an ad-tech client, we found 18% of AI interactions contained sensitive data before governance was in place. After: zero. You can’t fix what you haven’t measured.

Top-5 use cases, scored. Every candidate workflow gets scored on ROI and feasibility. Not ten. Five, ranked, with the math shown.

An architecture recommendation. Private, commercial (Claude, ChatGPT, Copilot), or hybrid — driven by the data sensitivity found in the risk review, not by ideology. I don’t resell any of these, so I can argue whichever one the evidence supports.

The output is a board-ready report and a working session with the leadership team. Then the 90 days start counting.

Days 15–45: foundations before features

Nothing user-facing ships in this window. That’s deliberate.

Governance goes in first. Three artifacts: data tiers (what’s public, internal, client-confidential), an approved-tools list, and a one-page acceptable use policy. One page. If your AI policy is 30 pages, nobody has read it and everybody is improvising.

Knowledge hygiene on the one corpus that matters. Not “clean all the data” — that’s a two-year project disguised as a prerequisite. For this composite firm, the corpus is past proposals and final deliverables: deduplicated, current versions flagged, access permissions verified. One corpus, done properly, beats five corpora done vaguely.

The first workflow goes into build. For most professional services firms, the top-scored use case is proposal or RFP response drafting, because it’s high-hours, document-grounded, and has a clear before/after. In one delivery engagement — a construction services firm — RFP responses went from 40–60 hours to 8–12. That’s the shape of result a well-chosen first workflow can produce. It’s also why the choice gets two weeks of scoring instead of a hallway vote.

Days 46–90: one workflow live, measured

The workflow ships — with a baseline. Before launch, the current process gets timed. Hours per proposal, drafts per week, whatever the unit is. If you don’t capture the before, the after is an anecdote, and anecdotes don’t survive budget season.

The adoption program runs. Team training on the actual workflow, not generic prompt tips. Two or three champions — practitioners, not IT — who field questions and surface friction. Usage tracked weekly. A tool nobody opens has an ROI of zero regardless of what the demo looked like.

The second workflow is queued, not started. It’s scoped and scheduled. It does not begin until the first workflow’s numbers are in.

The first monthly executive report goes out. Adoption metrics, measured results, risks, blocked decisions, the next-30-day plan. This is the operating rhythm the whole program runs on afterward.

By day 90: one workflow live with measured before/after numbers, governance functioning, a trained team, a queued second workflow, and a reporting cadence. That’s it. That’s the whole roadmap.

What’s deliberately not in the first 90 days

• No fleet of agents. One workflow, instrumented, before any autonomy anywhere.
• No company-wide rollout. Broad access without governance is how you get the 18% problem.
• No replatforming. The roadmap works with the systems the firm already has. MIT’s research found internal builds succeed at a third the rate of buying and partnering — and the flashiest projects aren’t where the ROI is anyway.
• No custom model training. A firm of 100 people does not need its own model. It needs its documents organized and one workflow that works.

The heuristic underneath all of this: ship one measured workflow before scaling anything. Every failed AI program I’ve examined violated it. Every successful one I’ve been part of followed it, whether or not anyone said it out loud.

Sometimes the roadmap says wait

Not every assessment ends in a green light. If the firm is mid-migration on its document system, if the one corpus that matters is genuinely unusable, or if leadership can’t name an owner — the honest recommendation is to fix that first and revisit in a quarter. I’ve made that call. It costs me a project and it’s still the right answer, because a roadmap built on a foundation that isn’t there just becomes another entry in the 42%.

The 90 days above aren’t ambitious. That’s the point. Ambition is what the second 90 days are for — once the first workflow has numbers behind it.