The chief operating officer of a general contractor spent a Tuesday in a hotel conference room deciding where AI belongs in his business. Twelve people, one facilitator, sticky notes on the window. By four o'clock they had a list. Estimating. Safety reporting. Something about RFP responses. Someone wrote "chatbot for the field" and nobody pushed back.
He walked out with a deck. The question he walked in with went home with him: why his estimators rebuild the same cost assumptions from scratch on every bid, and what that habit costs him at the margin line.
His company is about to spend real money teaching machines to do work faster. It decided where using a method that has not changed since the iPad launched.
The room is the method
Here is how AI strategy still gets made at most companies. You book two days. You invite the leadership team plus a few people who are good in meetings. You run fifteen interviews beforehand and pull the themes into a deck.
Then everyone brainstorms use cases, clusters them, dot votes, and plots the survivors on an impact-versus-effort grid. The upper-left quadrant becomes the roadmap.
Congratulations, you have an AI strategy.
The format has a real pedigree. It works well when the executives in the room hold the relevant knowledge. Pricing. Market entry. Org design.
AI is a different kind of question. The value sits in the mechanics of how work actually gets done. Which spreadsheet gets rebuilt every Monday. Which 200-page document gets read by four people who each need a different paragraph out of it.
Those details live with the people doing the work, and almost none of them are in the room.
What the room cannot see
Three failures follow, and they compound.
Coverage. Twelve executives and fifteen interviews represent a sliver of a company with thousands of people in it. Entire functions get summarized by whoever showed up, and the parts nobody spoke for never appear at all. Their absence looks exactly like having no opportunities there.
Voice. Meetings reward confidence and seniority. The superintendent who has run field operations for twenty years and hates talking in front of the CEO holds some of the most valuable process knowledge in the company. He contributes a sticky note.
Imagination. Ask people to name AI use cases and they name the AI they have already seen. Copilot. A chatbot. Meeting summaries. The list converges on the same six ideas at every company, which makes it a survey of the market's marketing rather than a map of your opportunities.
None of this is a failure of effort. Traditional discovery runs on two or three strategists, and every single conversation costs them four things: finding an hour on an executive's calendar, prepping for it, facilitating it, and doing the analysis afterward.
Fifteen interviews is not a sampling decision. It is the number that fits.
Breadth is what gets cut, and breadth is the whole point of a first pass.
Executives are also working from a stale picture of their own company. McKinsey asked C-suite leaders what share of their employees use generative AI for at least 30 percent of daily work. The leaders said 4 percent. The employees said 13 percent.
Adoption inside these companies is already three times deeper than the people planning for it believe.
Discovery that scales
Every constraint above comes down to how many conversations a strategy engagement can afford. AI removes it.
AI-guided interviews do not raise the ceiling on how many conversations an engagement can hold. They remove it. The three weeks that bought fifteen interviews will buy five hundred, and the cost of the five hundred and first is close to zero.
Every function, every region, every shift.
The interview asks role-tailored questions, so an estimator gets asked about bid assembly and a project manager about change orders. It follows the interesting answer instead of moving to the next item on the guide. It runs on the respondent's schedule, at 6 a.m. before a job walk or at 9 p.m. after one.
And because every conversation comes back as structured text, patterns can be found across all of them at once. That last part matters more than the volume.
A controller describing a reconciliation headache and a service lead describing a billing dispute are often describing the same broken handoff, three departments apart. Neither can see it, and a workshop cannot either, because those two sat in different sessions if they sat in sessions at all.
Scale also changes the question worth asking. "What AI use cases can you think of" measures how much AI marketing someone has absorbed. "Where does work break down" measures how well they know the business, which is the thing you were after.
It also gets sharp answers from people who have never thought about AI. A scheduler will tell you which three calls she makes every morning to confirm information that already exists in two systems.
That is an opportunity with its data sources attached, from someone who would never have been in the room.
The machine gathers, people decide
Scale on discovery is not automation of judgment.
An AI can collect every one of those perspectives and cluster them cleanly. It cannot tell you that the ERP migration makes three of those ideas impossible until next year, or that the loudest complaint in the transcripts is a symptom of a management problem no model will fix.
Strategists do that work. They validate what came back, challenge the claims that fall apart against how the business runs, and turn a pile of signal into a sequence someone can fund. Then they put their name on it, which is the difference between a recommendation and an output.
Sponsorship works the same way. The executive who commissioned the assessment is a source, not a filter, and an assessment the sponsor controls returns the sponsor's assumptions with better formatting.
Assess the opportunities, then concentrate
Broader discovery creates a second problem. Most companies are short of a reliable way to compare opportunities against each other and against what the business is already trying to do, and scoring exercises often make it worse.
A company builds a model. Value, feasibility, strategic fit, risk. Sensible weights. Each business unit scores its own proposals and sends the sheet back.
Nearly everything clears the threshold, the ranking comes back close, and the decision falls to whoever argues hardest in the room. The exercise produces a tie with arithmetic attached.
Assessment is a different act from ranking. It puts every opportunity through the same four questions, applied by the same people:
- What is it worth?
- Can it be built with the data and systems that exist today?
- Does it fit where the business is going?
- Is the organization ready to absorb it?
Dependencies, risks and tradeoffs then set the order.
What comes out should be short: confidence about the three or four opportunities worth funding first, not a well-formatted inventory of forty.
Fifty ideas, four decisions
A leading general contractor came to us with the problem in its usual shape. No shortage of AI ideas, no way to rank them, and a data landscape nobody had assessed.
Discovery ran across senior leadership and subject-matter experts in estimating, project management, field operations, procurement, safety, and finance. It produced more than 50 candidate use cases.
Each was scored on business impact, feasibility, data readiness, and implementation complexity, and the strongest received business cases with projected ROI.
Four categories rose to the top: estimating accuracy to protect project margin, predictive project risk to cut cost and schedule overruns, automated proposal and RFQ response to lift win rate, and specification analysis to speed preconstruction and reduce rework.
Discovery also surfaced the finding nobody wanted. Critical data sat in disconnected systems with limited structure and governance, so most of the priority use cases would stall at implementation however good the business case looked.
That is readiness assessed per opportunity rather than once for the company, and it is the finding that set the order.
The roadmap sequences the data platform work first, across three phases: foundation, which pairs infrastructure with early wins, core development, which pushes proven use cases into more departments, and scale, which operationalizes governance and adoption.
Discover, assess, prioritize, roadmap
We run this as GenROI. AI-guided discovery across the whole organization, strategists validating and challenging what comes back, assessment against value, feasibility, strategic fit and readiness, then a roadmap that ties every priority to the data foundation it depends on.
Every call carries the reasoning behind it. A plan built that way is shared, because much of the business helped build it, and defensible, because the reasoning travels with it.
None of it comes out of a two-day workshop, not because the people in the workshop are wrong, but because it reaches the twelve people who were invited and no further.
Which gives you a test for the AI strategy on your desk right now. How much of the business did it actually see?
If the answer is twelve people and a Tuesday, what you have is a list, and you are about to fund it.
The other several hundred people in your business have an answer to that question too. GenROI is how we go get it: discovery across the organization, opportunities assessed against the same four criteria, and a roadmap that names what has to be true before any of it ships.
