Nine AI proposals in the board pack. Roughly $2.4 million of asks against a $900,000 envelope. The finance committee has already done the sensible thing: sorted the list by projected annual benefit, drawn a line under the third row, and recommended those three. Everything below the line goes into next year's cycle. The meeting is scheduled for ninety minutes and the first sixty will be spent defending the ROI estimates in column four.
I have sat in versions of this meeting often enough — most of them through Webspot, the AI strategy firm I co-lead in Beirut — to know that the ROI column is almost never the thing that goes wrong. The estimates are usually defensible. What goes wrong is that the list is treated as nine independent bets, and it is not. Below is a composite board pack — plausible proposals, illustrative figures, no real client — worked through twice. Once the way most committees do it, and once the way I would.
The board pack — nine proposals, budget for three
Here is the list as it arrives, already sorted by projected first-year benefit, with the asks rounded:
- 1. Customer-service assistant — $420k ask, $1.1m projected benefit.
- 2. Contract review copilot — $310k ask, $740k projected benefit.
- 3. Demand forecasting — $260k ask, $600k projected benefit.
- 4. Sales-call summarisation — $140k ask, $310k projected benefit.
- 5. Invoice matching — $180k ask, $290k projected benefit.
- 6. Internal knowledge search — $160k ask, $240k projected benefit.
- 7. Quality-inspection vision — $300k ask, $230k projected benefit.
- 8. HR CV screening — $120k ask, $180k projected benefit.
- 9. Document labelling clean-up — $210k ask, $60k projected benefit.
The line gets drawn after row three. Proposals one, two and three are funded; the assistant, the contract copilot, the forecasting model. Nine sits at the bottom, and it deserves to on those numbers — $210,000 to clean and label a decade of documents, returning almost nothing on its own. In most cycles I have watched, that row is not deferred. It is quietly killed.
What the ROI column does not show
Every one of those nine initiatives touches four shared assets, and the board pack has a column for none of them.
The first is clean labelled data — not raw archives, but documents someone has actually classified, with agreed field definitions. The second is a working human review step: a queue, a named reviewer, a service level, and an escalation path for when the model is wrong. The third is integration access — the authenticated, permissioned, tested ability to read from and write to the systems where the work actually lives. The fourth is staff fluency, meaning a population of people who have supervised AI output long enough to know what a plausible-but-wrong answer looks like.
Each initiative either produces one of those assets or consumes one. Read the list again with that lens and the top three stop looking like a portfolio and start looking like a bottleneck. The customer-service assistant consumes all four and produces none. The contract review copilot consumes a labelled clause corpus that does not exist and a review step nobody has built. Demand forecasting consumes integration access to the ERP and clean product master data. The committee has just funded three of the heaviest consumers on the list simultaneously, in an organisation that owns none of what they consume.
What happens next is predictable and almost invisible in the reporting. Each project builds its own version of the missing asset, inside its own budget, to its own standard. The contract copilot spends five months and a large share of its budget having paralegals label clauses — work that never appears as a labelling line item, because it is booked as contract-review delivery. The assistant team builds an escalation queue for support tickets that no other function can reuse. Forecasting negotiates its own ERP read access on terms specific to forecasting. Twelve months later the board has three systems in partial production, three incompatible private versions of the same plumbing, and a genuine puzzle about why the returns arrived at roughly half the promised rate. The shortlist was right. The order was wrong.
The same nine, re-sequenced
Now map producers and consumers before the line gets drawn, and the same nine proposals rearrange themselves into three waves.
Wave one funds document labelling clean-up (ROI rank 9), invoice matching (rank 5), and internal knowledge search (rank 6). Labelling produces the clean data asset that three other initiatives consume. Invoice matching is the cheapest honest place to build a real review step and the first authenticated write into the ERP, because finance already has controls, an audit habit, and people who check things — it produces the review pattern and the integration foothold while returning $290,000 on its own. Knowledge search consumes the labelled corpus and produces staff fluency at almost no blast radius; when it is wrong, someone reads a slightly unhelpful document and moves on. Wave one's combined ask is $550,000, comfortably inside the envelope, and its combined direct return of roughly $590,000 is deliberately modest. It is not supposed to be the return. It is supposed to be the runway.
Wave two funds contract review (rank 2), sales-call summarisation (rank 4), and demand forecasting (rank 3). All three now start on assets that exist. Contract review begins against a labelled corpus instead of spending its first five months creating one. Forecasting inherits the ERP integration pattern rather than negotiating a new one. Summarisation extends the review habit into the commercial team.
Wave three funds the customer-service assistant (rank 1) and quality-inspection vision (rank 7) — the two highest-consumption, highest-exposure initiatives, now landing on a data foundation, an escalation path, integration access, and a workforce that has spent a year learning to supervise. That is also the point at which an initiative can stop being a project and become a standing role, which is the whole premise of BrianServes, my platform for deploying autonomous AI e-mployees: a defined scope, a named owner, and an escalation path, resting on shared assets that already exist. HR CV screening (rank 8) I would not schedule at all until someone has answered the governance question its ROI number silently assumes, which is a separate conversation from this one.
Note what did not change. Not one proposal was removed for being a bad idea. The rank-1 initiative is still the rank-1 initiative and still gets built. It simply gets built third, on foundations, instead of first, on nothing.
The four questions to add to your next AI approval paper
You do not need a new governance framework for this. You need four lines added to the template, answered by the sponsor before the paper reaches the committee. This pairs directly with the dependency line in the four-line board-ready AI roadmap, which is where these answers should eventually live.
- Which shared assets does this initiative consume, and do they exist today? Name them specifically — this labelled dataset, this review queue, this system integration, this trained population. "We will handle it in the project" is not an answer; it is an unpriced second project hiding inside the first.
- Which shared assets does this initiative produce, and which other proposals are waiting for them? A low-ROI initiative that unblocks three others is not a low-value initiative. It is the one you should be arguing about first.
- If we fund this alone, what will it build privately that a later initiative will have to build again? This is the question that surfaces duplicated cost before it is spent, and it is the one sponsors find hardest to answer honestly, because the duplicate work is usually what makes their own timeline look achievable.
- Does the stated ROI assume a start date the dependencies actually permit? A benefit projection built on a January start, in an initiative that cannot truly begin until its inputs exist in July, is not a forecast. It is a rounding error with a spreadsheet around it.
What changes at the approval table
Three things change, and they are governance changes more than analytical ones.
The board approves a sequence, not a shortlist. The paper that comes to the table proposes waves with named dependencies between them, and the decision recorded in the minutes is the order and the gates, not just the three names. Second, someone owns the shared assets across the portfolio — a single accountable person for the labelled data, the review pattern, the integration layer, and the fluency programme, so that no individual project gets to define those to suit itself. Third, foundational initiatives stop being scored against direct return. They are scored on how much they unblock, and if you cannot state that in a sentence naming the specific initiatives they unblock, they are probably not foundational.
None of this requires a bigger budget, and none of it requires better ROI estimates. It requires accepting that in an AI portfolio the initiatives are coupled, that the coupling is invisible in a ranked list, and that a board's most consequential act is usually choosing what comes first. That instinct is closely related to the budget split that quietly breaks most AI strategies: in both cases the number on the page is fine, and the structure underneath it is what determines whether the money turns into anything.
Nine proposals. Three slots. The shortlist is the easy part of that decision, and it is the only part most boards actually discuss.
Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.
