The delivery platforms told us a new customer at one of our restaurants cost fourteen dollars. The real number was forty-one. Same two digits, reversed. That turned out to be a fair summary of their math. Getting to the true number, then acting on what it revealed, is the most useful thing I’ve learned about AI this year.
The fourteen came with friends: seven-times ad returns, conversions credited to ads merely seen, results double-counted, every platform grading its own homework with its own math. That isn’t fog. Math that flattering has a job: it keeps you feeding the machine. The real data exists. It just never arrives in the monthly report. You pry it out of them directly, and you have to know what to ask for.
We pried, then built the Growth Intelligence Engine to force one true answer out of what came back: machine intelligence working over our own sales and costs and the outside feeds at once, each validating the other. The reading is the machine’s. The deciding isn’t. And we decided, precisely: turned the always-on ads off to run one clean experiment, kept the offer that pays and pointed it only at new customers at about a third lower cost, flagged a boost that looked fine until food cost entered the math, and started weighting the conservative platform’s numbers over the flattering ones. The subtlest find: our biggest spending weeks were buying our most expensive customers. So the fix isn’t spending less. It’s finding the band where a dollar earns, and grading every dollar against the register instead of the platforms’ homework. The point was never finding the flaw. It’s building the system that keeps the math flipped our way. I started by pulling every report myself and feeding the Engine by hand.
As of this weekend, the reports come to it directly, and the whole group fits on one screen, every number carrying its rule for what to do when it turns. The hour that went to collecting goes to deciding. Next quarter come the connectors, MCPs wiring it straight into OpenTable, Uber Eats, DoorDash, the catering platforms, our accounting, and Toast, our point of sale, so the outside data and ours arrive already fused, and the Engine surfaces the move on its own. Every rung buys the same thing: time at the top, calls on pricing, spend, and seasonality made while they still matter. It’s compelling enough that we’re considering taking it to market for the industry. The engine can be sold. The calls it informs can’t be.
That’s one of three rebuilds running across my businesses, and these notes come from inside them. Most companies are doing something else with AI: pilot it, license it, bolt it to the front of an unchanged business. The engine gets bigger. The business stays the same. Almost nobody is rebuilding the work itself.
At Makeable, our Creative × Capital studio, we’re building a direct-to-consumer program for an enterprise client in a multibillion-dollar category where single digits of share are worth tens of millions. The category’s problem was never the benefit, which is real and rarely disputed. It’s that the biggest competitor is hesitation, and the growth isn’t in the people already considering. It’s in whole audiences the category never spoke to, younger ones who don’t yet see themselves in it. So the program we’re building is designed to grow the category, not harvest it: to read the signals of curiosity as they first appear and use them to advance people from awareness to consideration to decision on a timetable the category has never seen, answering the specific hesitation that stalls each step. Readiness here won’t be waited for. It’s being built. What stays human is the hard part: knowing what the wanting is really about, the cultural read, the creative that earns trust. A campaign ends. This is being built to compound.
Through Vulcan, we’re inside a marketing-services firm remaking how it works. Its AI was scattered experiments; we’re codifying the firm’s own methods, training systems on them, and installing that intelligence inside the tools the team already uses. We’re mid-build, not done, and the objectives are set and measurable: reporting, monitoring, and first drafts running as systems the team supervises; subcontracted work coming back in-house; account people with time to counsel and strategists with time to think. The capacity we expect to free runs from hundreds of thousands in a single practice to millions across the firm, none of it from cutting people. Every job tilts toward growth, and headcount stops being how the firm scales. That’s optional today. In their business it won’t stay that way.
An operating business, a client program, a firm in transformation. One move across all three: intelligence inside the work, decisions kept human. The machine carries the volume. The person keeps the seat that decides.
Judgment and taste are this year’s words for that seat, and they’re being used with abandon, as if they were innate gifts you claim in a bio. They aren’t. They’re skills, honed over years, and both are drawn from the source the posts never name: wisdom.
Judgment is the call. Taste is the standard the call answers to. Wisdom is where both come from, and it doesn’t come from a dataset.
It comes from decades in the world: boardrooms and dining rooms, launches that worked and a venture of mine that didn’t, the pattern sense you only earn by answering for outcomes. The models are trained on everything ever posted. Wisdom isn’t online.
I’m a futurist by temperament; I’ve had a bromance with Ray Kurzweil for decades, so I’ll make his argument for him: every sensor and transcript is pulling the world into the training set, and on a long enough curve, wisdom may go online too. But a business doesn’t run on the curve. It runs in the meantime, and in the meantime the machine runs on yesterday’s patterns; even our systems that learn as they run learn inside a frame a person drew. The fourteen-dollar customer was yesterday’s pattern, reported by the machine that profited from it. The forty-one took a system we designed and a reason to doubt. And neither number is the point. The point is the string of decisions made because of them, ads off, offers re-aimed, dollars regraded, each one a call a machine wouldn’t have made and a person answers for. The next call is the one advantage no one else can own, and the opportunity isn’t out on the curve. It’s in the meantime, and we’re standing in it.
The last note said intelligence is inside the work now. Soon every business will have the same engine, trained on the same public everything. What’s left to own is yours alone: your data, and the wisdom to direct it. Almost no one has redesigned the work to let those lead. The engine was the easy part; the work is the whole job. Wisdom still matters. For now.
More Signal. Less Noise.