At a recent industry roundtable, mid-market IT leaders from across different industries kept arriving at three questions about AI. The questions were not technical. They were the questions you ask when the board has read the same article every other board has read, the vendors are pitching the same demo every other vendor is pitching, and someone in the room has to actually decide what gets built.
The three questions:
How should we use AI?
Where should we use AI?
How big an AI effort should we pursue?
The honest answers do not match the vendor playbook. They are also more useful.
How: the deployments earning money have three shared traits
The AI deployments that survive past the demo phase share a pattern. They have deterministic boundaries — the AI does one thing, the system around it constrains the input, and the output is checkable. They keep a human in the loop on anything that touches a customer, a dealer, or money. And they have a clear kill switch — if the AI degrades, the workflow falls back to the prior pattern without an outage.
The deployments that do not survive share an inverse pattern. They were built to impress the steering committee. They have no boundary; "the AI handles unstructured input" was the selling point. They removed the human reviewer in the interest of efficiency. And there is no defined fallback because nobody planned for the AI to be wrong.
The practical anchor: if you cannot name the problem the AI solves in dollars and hours, you do not have a deployment. You have a science project. Science projects are fine — they belong in R&D budgets, with R&D timelines, and R&D risk tolerance. They do not belong in operations.
Where: the surface test
The right surface for AI is the one where the cost of being wrong matches the value of being right.
Three surface categories most mid-market manufacturers can actually distinguish:
High-volume routine — claims analysis, pricing data normalization, document classification, anomaly detection. The AI processes work that would have taken a person hours per day; the person now reviews exceptions. Wrong by five percent on a Tuesday is recoverable. ROI is clean, and the audit trail is the work itself.
Customer- or dealer-facing with reversibility — chat assistance, content drafting, search relevance. The AI generates the first pass; the customer corrects it implicitly by the interaction that follows. Wrong is uncomfortable but reversible inside the same session.
High-stakes irreversible — financial commitments, regulatory submissions, safety-critical processes. These are the wrong surface for current AI capability. Not because the AI cannot draft them; because the cost of a single wrong answer outweighs the value of a thousand right ones, and the verification overhead consumes the time savings.
The surface test, asked plainly: if this AI is wrong on five percent of cases, what does it cost us? If the honest answer is "a customer waits a day longer than they would have," ship it. If the honest answer is "we file an incorrect regulatory submission" or "we underbid a multi-year contract," hold.
How big: three buckets, not one budget question
The AI conversation in most boardrooms collapses into "how much should we spend on AI?" That is the wrong question. The right question is which of three buckets the next dollar should go into.
Embedded — AI features in the tools you already license. Copilot patterns, intelligent assistance baked into CRM and ERP and productivity software. Low incremental investment, broad coverage, immediate availability. Most mid-market organizations have more of this than they realize, and use less of it than they paid for.
Targeted — one named use case, scoped build, measurable ROI. Moderate investment. The hardest part is choosing the right use case; the second hardest is resisting scope creep into a platform.
Platform — organization-wide AI capability with shared infrastructure, standards, and governance. High investment, multi-year payoff, requires executive sponsorship and a board commitment beyond IT. Mid-market organizations should generally not invest at platform level until Embedded is fully exploited and at least one Targeted deployment has proven the ROI pattern.
The pattern observed across the roundtable, and across most mid-market shops I have worked with: over-pitched on Platform, under-invested in Embedded, confused about Targeted. The reframe is the discipline. The next dollar belongs in whichever bucket has the highest marginal return — and for almost every mid-market manufacturer at the start of the AI journey, that bucket is Embedded.
What the three questions surface together
Asked one at a time, these questions get vendor answers. Asked together, they form a posture: AI is a portfolio decision, not a project decision.
The portfolio has three buckets.
Each bucket has a different surface profile.
Each surface has a different test for being wrong.
The mid-market manufacturers building durable AI capability are answering the three questions in that order: how, where, and how big.
They are arriving at investment decisions that are smaller, more focused, and more measurable than the AI conversations they walked into.
Plant Floor to Cloud goes out every Tuesday.
Where is your organization on the three-bucket question? Reply — I read every one.