Audiences AIThe AI that sits around the work
I keep seeing the same AI budget take shape inside smaller companies. A founder brings in a consultant, or makes a first AI hire, and within a month there is a working list: draft the proposals, summarise the sales calls, route the support tickets, chase the invoices. Every item on it is real work. Every item shows well in a meeting. And nearly all of it sits around the work rather than inside it.
The layer that keeps getting skipped is older and less photogenic: the statistics and machine learning that sit inside the commercial choices a business repeats every day. Which customer gets this week's offer, and at what depth. What price holds the margin on the fastest-moving line. Which account has gone quiet in the way that precedes a cancellation. How much stock to commit before the season shows its hand. A company makes these calls thousands of times a year, mostly by habit and spreadsheet, and each one has a modelled version that is slightly better. Slightly better, at that volume, is where the money is.
The distinction worth holding onto is where the ceiling sits. Automation takes effort out of work that already exists, so its value is capped by the cost of the task. If proposal drafting consumes half of one person's time, then perfect automation of proposal drafting is worth half a salary, forever, however good the models become. A decision carries a different bound. Its value scales with the pool it touches: the revenue that flows through a price, or the customers sitting in a renewal file.
Put illustrative numbers on it. A business doing twenty million dollars a year hands its proposal drafting to a model and saves, generously, forty thousand dollars of writing time. That figure is the ceiling, and it never moves. The same business lifts realised price by one point with a competent pricing model and books two hundred thousand dollars, this year and every year the discipline holds. One point is a modest ambition for pricing work. McKinsey has published the immodest version: a petrochemical company that rebuilt its pricing with machine learning added three percentage points of return on sales within a year. A margin result, produced by the least fashionable models in the building.
The demo problem
So why does the budget keep landing on the other side? Because the two kinds of work sell themselves in opposite ways. An automation demos in the first meeting. The founder types, a draft appears, and everyone present can price the time saved, because everyone present has done the task. A pricing model has nothing to show for weeks. It wants the company's own transaction history, cleaned. Its first output is a backtest. Its eventual win is a counterfactual, the mis-set price that never happened, and nobody applauds a counterfactual. Smaller companies can reach more AI talent now than at any point I can remember, and most of it is pointed at whatever performs best in the room.
A quieter force runs underneath. Automation is safe to attempt: if the draft generator writes a clumsy paragraph, someone fixes it and nothing is lost. Decision work touches the levers that hurt. A bad pricing rule surfaces in margin within a quarter. A mis-aimed offer hands discounts to customers who were already coming. Work like that needs an owner willing to be wrong somewhere visible, and when no one volunteers, the safe project takes the budget by default.
The correction is a sizing habit, and it is boring on purpose. Before funding any AI line item, write two numbers beside it: the cost of the task it removes, and the size of the pool it touches. Rank the list by the second number. In the companies where I have watched this done honestly, somewhere near the top sits a repeated decision, usually a price or an offer, that touches more money in a year than the entire automation list combined and has never once had a modelled input.
The automations can stay. They pay for themselves in hours saved, and they are pleasant to live with. But when a founder walks me through an AI plan now, I hold one question for the end, because it decides whether the plan was pointed at the right layer at all: which repeated decision in this business is quietly leaking the most money?
Sizing that decision is the first week of a custom engagement.