The renewal questionnaire from your liability insurer arrives, as it does every year, and this year it has grown a section. Does the firm maintain an inventory of the artificial intelligence tools in use? Does it have a written AI use policy? Have staff been trained in the use of AI? An answer is due before the policy renews. A week later a customer writes to the person who handles the account and asks one question in one line, which is whether the firm uses AI on their data. And at the next board meeting a director who approved last year's AI spending asks the question the other two were circling, which is what it changed.
You have answers, and they are assurances. The tools were bought, the pilots ran, people were enthusiastic, the vendor's deck said hours were saved, and someone in operations swears the work goes out faster now. Every one of those statements may be true. None of them can be checked, and the person asking knows it, because you cannot show where the business stood before the AI arrived, and a measurement taken only afterward has nothing to be compared against. So the questionnaire gets a paragraph, the customer gets a reassuring call, and the board approves the next budget on the same assurances as the last one.
I sat in the other chair for the last years of my career, responsible for AI governance at a global enterprise of more than 120,000 people, and I want to be honest about how I answered the question when it came to me. I had a list of projects. I had a count of tools in use, which I trusted a little less each time I checked it, and a stack of business cases, each written before its project started and none of them revisited after. What I did not have, for any one of them, was a measurement of the business taken before the work began and the same measurement taken after it. I could tell the board what we had built. I could not tell them what it had done, and I was the person whose job that was. The sentence I have arrived at since, after enough of those meetings, is the one this page rests on: nobody measures the business before AI, so nobody can show what AI changed.
It is a before-and-after question, and the before is the half nobody keeps. A leader who put AI to work a year ago is now asked what it changed, by the board that approved the spend, by an insurer whose questionnaire has grown a section, or by a customer who wants to know whether AI touches their data, and answers with the projects that ran and the hours a vendor said were saved. Nothing in the answer can be checked, because nobody wrote down where the business stood before the AI arrived, so the next budget is approved on the same assurances. An organization that adopts AI without a dated baseline cannot show what AI changed, and so cannot tell an investment that worked from one that only felt busy. The capable firms worry me most, because a capable firm has already started. It has three pilots and a vendor contract and an agent that somebody in operations connected to the billing system over a weekend, and it has no baseline, which is why it has to back up before it can go forward. The essay beside this one is about the step that produces the baseline, and I will not repeat it here. This page is about what a baseline is, what it makes possible, and what a firm holds once it has one.
A pre-AI baseline is the firm's mission and goals mapped to its actual performance, frozen and dated before AI touches the business, and re-measured each quarter afterward. Every word in that sentence does work, and "frozen" does the most. A number that can be quietly revised when the result disappoints is not a baseline. It is a negotiation. The baseline is written down once, with the date it was taken and the source it was read from, and from that day forward every result the firm reports about its AI is written as change since that date.
The measures are deliberately ordinary: revenue, gross margin, operating expense, headcount, revenue per employee, customer count and churn, and the cycle time of the three core processes the firm names as the ones it lives by. I chose them because every firm already has them. The annual statements hold the financial measures going back years, the billing or practice system holds the customers, and the timestamps in whatever system the work runs through hold the cycle times, so most of a baseline is a matter of reading what the firm already keeps and writing down the day you read it. Nothing about it requires a consultant's model of your business. It requires the discipline of taking the reading before the AI arrives, which is the part everybody skips, because by the time anyone thinks to ask the question the tools have been running for a year and the before is gone.
There is one rule I would single out, because it is the one a skeptical director should ask about. The first improvement a firm makes after its baseline is frozen is kept outside the processes being measured, so that it moves no number the firm will later be judged against. A baseline is only worth keeping if the person who set it cannot game it, and a firm that follows this rule can say so to anyone who asks.
What the firm holds afterward is a picture of itself with the baseline inside it, and I will describe it the way you would meet it, as pages on a screen, before I tell you its name.
The first page says who the firm is and why it exists: its stated goals, the drivers that serve them, its departments with the person accountable for each, and, in a short list that always draws the eye, the things it spends money on that serve no stated goal. The second page is the register, which holds every agent and tool found running in the firm, each with a Tail Number, the identifier the framework paints on an agent the way a registration is painted on an aircraft, so that any system it touches can tell which agent acted and whose it is, and each with an owner's name or the honest word that there is none, beside the ring of vendors the firm depends on. The third page is the gaps, each tagged to the pillar of the framework it falls under. The fourth is the baseline, the seven measures with the date each was frozen and where it was read from. The fifth is the plan, every gap with the person who owns closing it or the admission that nobody yet does. The sixth is before and after, the questions the firm was asked by its insurer and its customers, with what it could say then and what it can say now, and each answer linked to the evidence that supports it or marked as not yet answered.
Two numbers sit at the top of every page. The first is trust posture, one reading drawn from the five pillars of the Trusted Living Agent Framework, the map the page before this one drew. The second is results since baseline, the seven measures as change against the day they were frozen. Beside them is a third figure I have come to think of as the most important thing on the page, the verified fraction, which is the share of everything the record holds that has been checked against a second source rather than taken on someone's word. Every entry on every page is drawn by how it is supported. A fact checked against a second source is drawn solid, a fact seen in a log or a ledger is hatched, a fact the firm wrote down is dashed, a fact somebody said is dotted, and a fact that is not yet real, a projection, is drawn as an outline with its assumptions written beside it. A reader can tell at a glance how much of the picture is known and how much is asserted, which is more than most firms can say about their own annual report.
The record has three states in time, and the pages can be turned to any of them. The first is as found, the day the baseline was frozen. The second is the re-measurement, taken each quarter, in which every result is written as change since baseline and the record makes no claim about what caused it, because it does not know what caused it and says so. The person who sponsored the work signs the attribution, or declines to, and the page records which. The third state is the projection, a year ahead, drawn in outline, and it carries the rule I would carve over the door of the whole thing: a projection is never measured.
The picture has a name, and the name is the Trusted Living Twin.
I did not lead with the name because the name is the last thing a leader needs from it. The twin is the whole framework instantiated for one firm. The baseline is its foundation, the business results at the top are what it re-measures, and the register, the gaps and the plan sit between the two, so that a director can run a finger from a goal the firm wrote down to the agents and vendors that serve it to the number that says whether anything moved. What makes it living is that it is kept current and re-measured, so that the question at the top of this page is answered on a schedule instead of under pressure. The twin lives with the firm it describes: the record is the client's and stays in the client's own custody by default, and a firm that would rather not keep it can choose to have it kept for it, in a space of its own, by the people who built it.
I built the first one for a firm that does not exist. Hollister Reed is a fictional regional accounting and advisory firm, a few hundred people across several offices, several dozen partners and no chief information officer, with the kind of AI use a firm of that size actually has: tools adopted by whoever found them useful, a vendor list nobody has read end to end, and an insurer's questionnaire on the managing partner's desk with the answer due. Its twin is public, and I would rather you looked at it than took my description of it. The questions on its before-and-after page are the ones at the top of this essay, and the record's answer as found is that the firm could answer only with assurances. Its baseline page shows the seven measures frozen on a date, each with its source named: the annual statements for the financial measures, the payroll register for headcount, the practice system's timestamps for how long an audit report or a tax return takes from start to issue. Its verified fraction on the day it was found is low, because nothing in a firm's own account of itself has yet been checked against a second source, and I left that figure where it fell rather than dress it, since a twin that claims verification it does not hold is worth less than no twin at all.
The twin earns its name after the engagement ends, and that is the part I would ask you to look at hardest, because every consultant leaves a report and most reports are stale by the following quarter. The twin is re-measured each quarter, against the same frozen date, by the same seven measures, and two one-page documents fall out of each re-measurement. The board summary is one printed page: the two numbers, the verified fraction, the five pillars and the five drivers of the business. The receipt is the other, and it is the one I built with the insurer in mind. It carries the questions the insurer asked, what the firm could say at the last renewal and what it can say at this one, with the evidence beside each answer. Neither page is written by hand. Every page of the twin is drawn from one record, and the footer of each says so, which means the board summary cannot say something the record does not hold, and neither can I.
Comparison lines will come later, showing where firms of your size and kind stand on the same seven measures, once enough firms sit behind the line that none can be picked out of it. The demonstration shows none, and I would distrust one that did this early.
What all of this comes to, in the words we use for everything we sell, is that the twin lets you trust that your AI is paying back, or shows you that it is not, and proves either answer with a dated record that a board, an insurer or a customer can read for themselves. That is the third of the seven things a leader can come to trust about the AI in the company, and it is the one that cannot be bought later, because the before cannot be taken after.
I have not worked out why the before is so easy to skip. The firms I talk to keep careful records of what they spent and almost none of what the spending changed, and the people in them are not careless. My best guess, and it is only a guess, is that a measurement taken before the work is a promise, and a promise written down with a date on it is a harder thing to sign than a budget.
Five questions, one for each pillar, will draw the first outline of your own firm's twin: the five-question screen. The finished picture, on a firm where nothing is at stake, is beside it: the Hollister Reed demonstration.
