The AI proposals arriving in board packets have started to look alike. There is a vendor, a pilot scope, a per-seat price, and a savings estimate stated with more confidence than anything else in the memo. What is almost always missing is an account of what the work costs today. Not the salary line — the cycle time, the rework, the error rate people have quietly learned to live with, the number of times a human touches the same record before it is finished.
That gap is the whole problem. Without a baseline, every outcome can be declared a success after the fact, and the company ends up with a subscription instead of a margin.
The number the slide almost never contains
Before I support spend on AI in portfolio companies, I want one thing from management that is harder to produce than a vendor quote: the current cost of the specific workflow, measured the way they intend to measure it afterwards. How many invoices per week. How long a quote takes from request to delivery. What share of support tickets are resolved on first contact. How many hours a month go into pulling the same report.
This request does more work than it looks like it does. If a team cannot measure the process now, they will not be able to measure the improvement later, which means they cannot manage it either. Sometimes the honest answer is that nobody knows, and that answer is useful. It tells you the first investment is instrumentation, not intelligence — and instrumentation is cheaper, faster and almost always undersold.
It also reframes the conversation from technology to operations. A workflow you can describe in numbers is a workflow you can improve without AI at all. I have seen proposals disappear at this stage for the best possible reason: once the steps were written down, three of them turned out to be unnecessary.
Three conditions, and you need all three
The functions where AI earns its keep share a shape. First, the data already exists as a byproduct of doing the work. Tickets, invoices, contracts, call transcripts, code, order histories, specification documents. If the proposal begins with a data collection project, you are not funding an AI initiative, you are funding a data initiative with an AI story attached. That may still be worth doing, but it should be priced and sequenced as what it is.
Second, the work is repetitive and high volume. AI pays back on frequency. A task performed four hundred times a week compounds small gains into real capacity; a task performed twice a quarter does not, no matter how painful each instance feels. Judgment-heavy, low-frequency work is where senior people should be spending their time anyway.
Third — and this is where most enthusiasm dies — the function must tolerate error, and correction must be cheap. The right question is not whether the system will be wrong. It will be. The question is what happens when it is. If a mistake is visible immediately and fixed in seconds by the person already looking at the screen, the risk is manageable. If a mistake is silent, travels into a customer contract, a regulatory filing or a financial statement, and surfaces months later, the expected cost of being wrong swallows the savings whole.
Any two of these three conditions produce a pilot that demos beautifully and never scales. All three, and you have something worth capital.
The functions that pass, and the ones that flatter
Applied honestly, the filter is unglamorous. It tends to bless back-office and mid-office work: extracting fields from invoices, claims, purchase orders and specification sheets; drafting first-pass responses to inbound support that a human reviews before sending; searching across a decade of contracts for a clause nobody can remember agreeing to; assembling quotes and configurations in a technical sales team from a catalogue too large to hold in a head; sequencing receivables follow-up; assisting engineers inside a software product organisation. Dull, repetitive, data-rich, correction-cheap. Exactly where the money is.
The proposals that flatter the boardroom are usually the opposite. Demand forecasting built on a thin history of unusual years. Pricing intelligence in a business that has never run a price test and has no elasticity data to learn from. Strategy and market analysis generated in seconds, which impresses everyone and changes no decision. Personalised outbound at industrial volume, which borrows against a brand the company spent years building. Customer-facing chat placed in front of people who are calling precisely because something has already gone wrong for them, where the tolerance for a confident wrong answer is close to zero.
From a board seat you rarely have the operating detail to judge the technology, so judge the request instead. A few signals reliably separate substance from theatre:
- The tool was selected before the workflow was named.
- No single named person owns the outcome, only a committee that will "explore" it.
- The pilot has a start date and no kill criteria.
- The ask is for a platform rather than for a problem to be solved.
- The champion is animated about the capability and vague about the constraint it removes.
None of these is disqualifying on its own. Three of them together and you are being asked to fund a demo.
Allocating across several companies at once
Looking at more than one company changes the arithmetic in ways a single management team cannot see. The most common waste in a portfolio is not a bad bet; it is paying three times to learn the same lesson. Three companies, three vendors, three pilots in document extraction, three sets of implementation scar tissue, no shared conclusion. That is not diversification. It is a tuition bill paid in triplicate.
The overcorrection is worse. A mandated common platform imposed on businesses with different operating models, different data hygiene and different customers tends to be adopted in name and abandoned in practice. What actually travels between companies is method, not tooling: the baselining discipline, the three conditions, negotiated commercial terms, a scarce technical person whose time can be shared, and a candid account of what failed and why.
So I would rather sequence than spread. Fund it first where the data density is highest and the owner is clearest, insist that the result be written down in terms the other management teams can use, and let the others watch. This is ordinary post-closing value creation work, not a separate technology agenda, and it should be judged on the same terms as any other operating investment.
One caveat on the revenue side, which is a genuinely different bet from the cost side. AI is changing how buyers discover and evaluate companies before anyone in sales hears from them — I operate BrandAmplifi, which works in online reputation and search visibility, so I watch that shift closely. It is real and it matters. But it is a market-structure question about distribution, not an internal efficiency question, and folding the two into one line item on one slide is how boards end up approving neither properly.
Savings nobody harvests are not savings
Here is the question that makes management teams shift in their seats, and it should be asked before approval rather than a year later. If this works, what happens to the freed capacity?
There are only two honest answers. Either the company absorbs growth without adding people, or it reduces cost. Both are legitimate. What is not legitimate is a business case built on hours saved where the hours quietly return to the same team doing the same volume slightly more comfortably. In that case the cost base has not moved, the subscription is new, and margin is worse than before.
Deciding this in advance also forces honesty with the people whose work is affected. Teams can tell when a project is described as a productivity aid and structured as a headcount plan. That gap destroys the cooperation the project depends on, because the people who know where the exceptions hide are the ones being asked to document them. Most of this reckoning happens well away from the quarterly meeting, in the follow-through that defines what a board actually does between the meetings.
The advantage is the workflow, not the model
The capability itself is not scarce and will not stay expensive. Everyone in every industry will have access to the same models on roughly the same terms, which means the model cannot be the advantage. The advantage is the proprietary record of how a specific business does specific work — captured, cleaned, measured and improved — and the organisational discipline to keep doing that after the novelty wears off. That is closer to the ordinary craft of investing than to a technology thesis.
Which is why I have become comfortable saying no slowly and yes narrowly. Approve one workflow with a measured baseline, a named owner, a correction path and a decision about the freed capacity. Then do it again. We will look back on this period and find that the companies that compounded were not the ones that adopted first, but the ones that could tell you exactly what the work used to cost.