The board asks for an AI strategy. The big consultancies answer with a transformation roadmap. Here is what the work actually looks like, month by month, inside a real mid-market plant — and why none of it starts with buying software.
Picture a composite of the manufacturers we know best: roughly 200 employees, $40M in revenue, two plants, an ERP that was implemented a decade ago and customized into something nobody fully understands, a quoting process that lives in one estimator's head, and a CEO who has now been asked about AI at three consecutive board meetings. There is budget. There is urgency. What there is not — and what no one on the org chart can supply — is judgment about where AI actually pays in this business.
That judgment gap is the job. Not evangelism, not a lab, not a 90-slide operating model. A fractional Chief AI Officer exists to do, part-time and at senior level, what the company cannot yet justify a $400,000 hire to do full-time: find where AI returns money in the operation, get the first systems into production, and build the discipline that keeps them there. What follows is the honest version of that engagement.
Weeks one to three: the diagnostic nobody wants and everybody needs
The engagement does not begin with tools. It begins with an uncomfortable inventory. The CAIO walks the quoting process with the estimator, sits in the Monday production meeting, pulls three years of job-cost history, and asks the questions that map where knowledge actually lives. In a 200-person manufacturer, the findings are remarkably consistent: the highest-value knowledge is tribal, the ERP data is 80% usable and 20% fiction, and the most expensive daily activity is skilled people re-deriving answers the company already paid to learn once.
The output is not a vision document. It is a ranked map — every viable AI use case in the operation, scored on two axes: estimated payback and implementation effort. And the most valuable rows are usually the ones crossed out. A board-ready diagnostic says no in writing: no, computer vision on the line is not your first project; no, a custom model is not justified; no, the chatbot the software vendor pitched will not survive contact with your data. The discipline to sequence — and to decline — is most of what the title is for.
of AI pilots never reach production — and the failures are operating problems, not model problems: unclear success criteria, untrusted data, no named owner. The minority that ship return roughly 171% on average. The diagnostic exists to put a company in the second group before a dollar is spent on the first.
Forrester · Anaconda · 2026 Adoption ResearchDays 1–30 of deployment: the first system is always unglamorous
In nearly every mid-market manufacturer, the highest-payback first systems come from the same short list, and none of them photograph well for a press release.
Quoting and estimating is usually first. The company has years of historical job costs — actual hours, actual material, actual margin — and an estimating process that uses almost none of it. A system that grounds new quotes in that history does two things at once: it compresses quote turnaround from days to hours, and it stops the silent margin leak of underpriced complexity. When one estimator holds the pricing logic, this is also succession insurance, because that estimator will eventually retire, and the institutional knowledge currently walks out with them.
Document and tribal-knowledge retrieval is the quiet second. Maintenance manuals, engineering change orders, customer spec sheets, the folder of PDFs named FINAL_v3_USE_THIS_ONE — a retrieval system over the company's own documents gives every engineer and maintenance tech the answer in seconds that used to require finding the one person who remembered. It is the least exciting deployment imaginable, and on the shop floor it is frequently the one people refuse to give back.
Revenue operations is third, and it is where the front office meets the same discipline. Scoring the existing pipeline on real win patterns — which deals are actually likely to close this period, which are structured optimism — turns the forecast from a feeling into a fact. For a manufacturer with three to ten salespeople and no RevOps function, this is routinely the fastest payback in the entire map.
Notice what these three share. The data already exists. A named owner already lives inside each workflow. And the payback is measurable inside a quarter. That triad — existing data, existing owner, measurable quarter — is the entire selection filter, and it is why the first systems are chosen by an operator and not by a software vendor's demo calendar.
Days 30–60: governance, or the part everyone skips until an auditor asks
Somewhere in the second month, the work turns to paper — and this is the portion of the engagement that looks least like "AI" and protects the company most. A usable AI policy for a 200-person manufacturer fits in a few pages: what data may enter which tools, what never leaves the building, which outputs require human review before they touch a customer or a price, and who owns each deployed system by name. Customers' procurement teams are beginning to ask for exactly this document in supplier audits. Insurers are not far behind. Having it written before it is demanded is the difference between a checkbox and a fire drill.
The same month, the CAIO runs the upskilling that makes the systems survivable: not "AI training" in the abstract, but teaching the estimator, the maintenance lead, and the sales manager how their specific system works, where it fails, and how to correct it. The goal is unromantic and essential — every system must run without the CAIO in the building, because the CAIO is, by design, not always in the building.
companies now have a Chief AI Officer, and 66% of executives expect most companies to hire one within two years. The mid-market question is not whether the function arrives — it is whether it arrives at $400,000 full-time or 8–20 senior hours a week.
IBM CAIO Survey · 2025Days 60–90: production, measurement, and the only review that matters
By the third month, the engagement is judged on one standard: working systems in production use, with numbers attached. Quote turnaround before and after. Forecast accuracy before and after. Hours of search-and-re-derive eliminated per week. A 90-day review that presents activity instead of these numbers is a strategy deck wearing a hard hat — and the board should treat it as such.
This is also where the honest version of the timeline matters. Production research puts median time-to-value for AI deployments near five months; a disciplined 90-day result is earned by scoping small, owned, and measurable — not by heroics. Two or three systems genuinely in production beats eight pilots in perpetual evaluation, every time, in every plant we have seen.
What the role is not
It is worth saying plainly what this job is not, because the market is busy selling the confusions. It is not a research function — a 200-person manufacturer should never train a model. It is not a vendor-selection concierge — tools are chosen after the process is defined, not before. It is not delegation of judgment — the CAIO ranks and recommends, but the named owners and the executive team decide. And it is not permanent by default. The role exists to make itself replaceable: after several quarters of production systems, the right structure is usually a permanent hire — often a director of AI or data rather than a C-suite seat — stepping into systems that already run, with the fractional executive handing off rather than holding on.
The board's question — what is our AI strategy? — turns out to have a short answer. It is a ranked map of where AI pays in your operation, two or three systems in production with named owners, a governance document your customers' auditors will accept, and a leader accountable for all of it at hours the P&L can defend. That is the whole job. Everything else is a press release.
This is the engagement our Fractional Chief AI Officer practice runs — a founder-led AI Opportunity Diagnostic, a vetted operator deployed against it, and working systems in 90 days. The diagnostic is fixed-fee and credited toward the retainer.