The bottleneck isn't drafting. It's the partner who has to put their name on it.
Picture the 75-person shop: four or five practice leads, a couple dozen seniors, a wide base of analysts and associates. The economics run on leverage — how much associate output one partner can review and stand behind per billable week. That ratio is the whole game. AI changes that ratio, but not in the direction the demo promised.
Here's the trap a firm this size walks into. An associate drafts an SOW scope, a findings memo, or a client deliverable with AI. It looks finished. It is not finished. The partner now has to read it more carefully than a human draft, because the failure mode is different: a human associate gets the structure right and fumbles a detail; an AI draft gets the prose confident and fumbles the judgment — the caveat that should have been there, the client-specific risk it smoothed over, the number it asserted without a source. If partner review takes longer than the associate's draft saved, your leverage ratio just got worse. You're paying senior rates to debug junior-rate output.
That's why readiness at this size is a workflow question, not a tooling question. The RSM middle-market AI survey, the San Francisco Fed analysis of AI and small businesses, and the OECD report on AI adoption by small and medium-sized enterprises all land on the same discipline: name the workflow, name who owns the data, and name the result you'll measure — before anyone touches a tool. For a services firm, the first workflow worth picking is one where a partner's review currently bottlenecks the whole team. Use the SMB AI readiness assessment to score where you actually stand on source quality, ownership, and measurable delivery impact rather than on vendor enthusiasm.
What 75 people exposes that 15 people hid
At 15 people everyone shares the same methodology because it lives in three people's heads. At 75, you've grown past that. One practice has a clean, documented engagement template; another runs on a senior associate's memory and a Slack thread from 2024. Feed AI the documented practice and it produces something defensible. Point it at the tribal-knowledge practice and it confidently invents the methodology — because the source it needed was never written down. Readiness fails along the seams between your practices, and a firm this size has seams.
So before client work touches AI, build the review model the way the NIST AI Risk Management Framework would have you build it: name the specific use case (say, drafting the recurring sections of a fixed-scope engagement report), enumerate how it fails for that use case, set the controls, and assign a named partner as accountable owner. Then handle client data the way the CISA AI Data Security Best Practices demand it: which client material is permitted as a source, who approved it, where outputs get logged, and what happens when the system hits an engagement it has no context for. For a professional services firm, that last control is not optional — confidentiality is the product, and a model that leaks one client's framing into another client's deliverable is a malpractice problem, not a quality problem.
Pick one workflow inside one practice that already has clean source assets. Sequence the source cleanup, the partner-review design, the prototype, and the adoption push using the 90-day AI implementation plan. Resist the instinct to roll it firm-wide at once — the seams between practices are exactly where an unsupervised rollout breaks.
Measure the leverage ratio, not the activity
The Deloitte State of AI in the Enterprise 2026 keeps making the same point firms keep ignoring: governed workflows produce value, scattered experiments produce dashboards. For a 75-person firm the proof you're hunting for is narrow and specific — did this workflow let a partner review more associate output per week without their rework climbing? That's the leverage ratio moving. Everything else is noise.
So track the pair that actually matters together: how much faster delivery moves AND how much partner rework that speed costs. Concretely — utilization lift, time-to-engagement-start, QA defects caught in review, deliverable rework hours, review cycle time, and adoption among the managers who own delivery quality (not the enthusiasts who'd adopt anything). If drafting got noticeably faster but partner rework ate the savings, you haven't proven anything except that the tool can type. The real decision is whether you can safely standardize this workflow as the firm's default — make it the way that practice works, not an option a few people use.
Run it through the AI ROI model without fake savings so the win is real hours and real margin, not estimated time multiplied by a fully-loaded rate. Only when one partner-owned workflow clears that bar should you take it to the next practice — and that next practice has different source quality, so you re-test, you don't copy-paste. When you're ready to map the full sequence across practices, build the AI roadmap.
