Blackstone& field note · article three · for the people being reported to
You do not need to understand how AI works to govern it. You need eight well-defined, testable concepts: a floor of shared vocabulary small enough to learn in an hour and solid enough to carry a funding decision.
This note is for board and investment committee members, finance, PMO and assurance leads, and business owners: the governance community that receives the delivery manager’s translation. Article two gave the delivery manager a language. This piece makes sure the room understands the sentence.
“...so is it green or not?”the room, before the floor
the status line from article two’s report, arriving in front of the room · the tags that surface underneath it are the floor at work: three concepts, one legible status
This series speaks in two voices. › The dark strips are the instruments reading out. The white cards are the room speaking. Articles one and two used the white cards for the delivery manager’s translation; this piece hands them to the room itself, because the room is who it is for.
The conversation reverts to the questions the old instruments were built for: “so is it on track for March?”
“A measured band, converging” reads like an excuse where “green” would have read like an answer. The vocabulary built to carry bad news gets heard as not carrying news at all.
It moves quietly to whoever in the room sounds most fluent, often the sponsor or the vendor: the fluent voice with skin in the game.
The result is the fake-green machine rebuilt one level up. The delivery manager reports the position accurately; the room, unable to weigh what it hears, supplies its own certainty.
When Gartner predicts that over 40 percent of agentic AI projects will be cancelled by the end of 2027, through escalating cost, unclear business value or weak risk controls, this is one of the mechanisms behind the numbers: not bad models, but rooms that could not read the instruments that would have caught the problem while it was still cheap.
the prediction is Gartner’s (25 June 2025) · the mechanism is our inference
Nobody on a stage-gate forum was hired for fluency in evaluation methodology. Their craft is scrutiny: they know how to read a burndown, a variance report, a RAG status, and they know where those instruments hide bad news.
Those scrutiny instincts are exactly right; they are pointed at instruments that no longer measure the work. An eval band and a drift record are where the work now shows itself.
Eight concepts, not a field of study. Each with a plain definition, a named owner, a place on the report, and a question the reader can now ask.
Anyone who can read a discounted cash flow can read an eval band. The DCF took longer to learn.
The floor is deliberately small, and every concept on it is testable: a plain definition, a named owner, a place it appears in the delivery report, and a question the reader can now ask that they could not ask before. That last part is the point. Literacy that does not change what you can ask is trivia.
how to read the cards · the small report map beside each owner line is the delivery report in miniature; the highlighted line, named underneath it, is where the concept shows up · the white card at the bottom is the room speaking: the question the concept buys
An agentic system’s output varies run to run; that variation is inherent to the method. Quality is a range you measure and hold, not a state you reach once. “Done” is a measurement that keeps being made.
The eval is a repeatable exam: the system is scored against examples of “good”, and every run is logged with dataset version, model version, score and date. The bar is the pass mark governance sets. The gate is the decision moment where the evidence decides whether the next stage is released, held, or stopped: the score against the bar, the cost line and the value case, weighed together.
The business-owned collection of examples that defines “good”, curated and signed by domain experts. The delivery team can propose additions; it cannot change the exam it sits.
The unit of delivery. Each card carries its own pass mark instead of a binary “done”. Work flows card by card through a visible queue (engineers will call it Kanban), which is where the familiar agile rhythm lives on.
Money is released in capped blocks at gates, on evidence, rather than committed once up front against a plan. The next tranche is a decision, not an entitlement. Maximum exposure between any two gates is the tranche.
Live quality decays as the world changes underneath the system. Monitoring watches the band in production and fires when it slips. A control firing is the control working, the smoke detector going off, not the house burning down; the same posture covers guardrails, where a blocked out-of-policy action is direct evidence the control layer works.
A person inside the system’s flow, at one of two settings. At key decision points, every case passes a person before the action lands (the system drafts, the person decides): the norm wherever an action is hard to reverse. Elsewhere, a designed share of cases routes to a person, with a tolerance and a tracked trend. Autonomy widens only as the eval band and drift record earn it.
What one resolved case costs, tracked as a trend against an approved envelope. The unit is a resolved outcome the business wanted, counted after the calls and tokens are spent.
Every concept has a number, an owner and a record: the band has a run log, the dataset has signatures, the tranche has a cap, the envelope has a trend line. That is what separates the floor from the hand-waving that has made “AI literacy” a tired phrase.
The questions are the floor in use. A governance community that has internalised all eight is far harder to fake-green, because each question demands evidence that either exists or visibly does not.
Article two closed on a one-page status report. Here it is again, rebuilt line for line, because it is also the syllabus: every concept on the floor is a line on this page. The numbered markers show where each one lives; the white cards alongside are the room reading it, question by question. The figures are an illustrative worked example carried through the series, not a specific client engagement.
A note on provenance: the top four sections are article two’s report exactly as banked there. The dataset reference, the colour chips and the three sections below the divider are added by this piece: the same report, three sections fuller and coloured, as the build matures into live running.
88 of 92 and converging. Forecast: two iterations to the bar; if not cleared within the tranche, the decision returns to the forum at the gate.
replaces the reported ● green · the chip beside this line is computed from the band, not chosen · the gap to the bar is visible, and visible is the point
Prove gate passed 14 June on real data. Build gate next; beta assessment follows it.
progress = which gate has passed and which is next · not percent-of-plan · behind the gate, work moves card by card, each against its own measure
Tranche 3 of 5 released and capped. Spend reports against the released tranche.
the next tranche is a decision, not an entitlement
Resolves nine in ten cases inside cost; the tenth goes to a person, and the rate is tracked weekly.
One drift event this period. The monitor fired at 09:40 Tuesday; affected traffic routed to a person from detection; back in band 31 hours later after re-grounding.
a fired control is the control working · exposure bounded while it worked
One case in ten routes to a person, against a tolerance of 15 percent set in the business case. Trend steady over the last six weeks. Payment release stays review-all by design: every case sees a person before money moves.
a designed rate with a tolerance, planned in the business case
Cost per resolved case is at 82 percent of the approved envelope and trending down. Not cost per call, not cost per token: the price of a result.
finance owns the envelope; delivery reports the trend
the markers 1 to 8 sit on the report lines they interrogate · the questions land in the same order, top of the page to the bottom · every one of them is answerable from the line it sits beside
Thirty seconds in, the instruments have been ignored and the conversation has reverted to the one question the old instruments were built for.
Band, bar, trend and tranche, read together in one pass. The question that comes back is the floor in use.
That gap is the whole piece.
The floor is small, so the programme to build it should be small. The failure mode of enterprise AI literacy is the opposite instinct: a 40-slide “AI 101” deck that explains transformers and prompt engineering, teaches nobody how to govern anything, and burns the room’s patience for the subject. How models work is the engineers’ literacy. How to govern the work is the room’s. Teach the second.
One hour, on your own numbers. Not “an eval is...” but “here is our eval, here is our bar, here is who signed our dataset”. Concepts stick when the example is the thing you are already accountable for.
Every delivery report carries the vocabulary in use; the first three carry a one-line glossary under each instrument. The room learns the language the way anyone does: by needing it weekly, in context.
The forum rehearses one gate decision on real evidence before money is at stake. One hour, one decision, no consequences. Where “a forum that can actually stop” gets its muscle memory, and continuation bias, the review reflex that never stops anything, is named before it has anything to protect.
Everyone holds the full floor; each role then goes deeper where its decisions live. The table below is the whole of it.
Someone owns the vocabulary the way someone owns the risk register: keeps the glossary current, inducts new forum members, retires jargon before it accretes. A natural Centre for Enablement function; without one, a named individual, not a committee.
The annual e-learning module, completed and forgotten. The floor is used weekly or it is not a floor.
The delivery manager bridges the languages, but a bridge is not a substitute for the floor. If the room can only reach the work through one person’s fluency, that person is a single point of failure and, eventually, a single point of blame.
One fluent voice in the room goes unchallenged because nobody else can weigh the claim. The floor exists precisely so that no one in the room has to take anyone’s word for anything.
regulators are pointing the same way · EU AI Act Article 4, a binding AI literacy obligation, in application since February 2025 · ICO AI guidance, trained and competent staff wherever AI touches personal data · AI Playbook for the UK Government (February 2025), the same upskilling expectation across the public sector
a tailwind, not the case · worth one line in the business case for the hour it takes · the real case: a room without the floor is the residual risk that no other control can close
Articles one and two built a delivery method whose status is far harder to fake and a governance machinery whose gates release on evidence. This piece closes the loop at the top of the room. With the floor in place, the translation lands as meaning, not jargon. The measured amber reads as what it is: a control doing its job early. The stop recommendation, when it comes, is heard as the machinery working. And “quality gate at 88 of a required 92 and converging” reads as what it actually is: a defensible position with a forecast and a decision point attached.
The business wanted confidence that money follows evidence, and that someone can say, in language it understands, exactly where things stand. Article two built the someone. This piece builds the understanding. Between them, the sentence finally has a speaker and a listener. And the floor has one more property worth naming: it travels. The eight concepts are not project-specific. The community that learns them on the first agentic build governs the second from day one, which is the quiet beginning of the capability this whole series is really about: an organisation that can keep doing this without us.
Next in this series: how much of this actually changes. Less than it looks: the lifecycle survives, the gates survive, the craft survives. The case that the shift everyone is bracing for is smaller than the fear of it.