Everyone in the organization writes with it now — in the tools you sanctioned and the ones you never saw. Fluent, confident text, generated across every surface at almost no cost, faster than anyone can read. What it does not do is keep that output consistent, or accountable to what the institution has already said.
Plainly: Council is the coherence layer beneath everything an institution publishes. It keeps one approved record of what the organization has claimed, committed to, and decided — and checks every new draft against it, surfacing contradictions for a person to resolve before they ship.
Read the argumentEveryone writes with it now — in the tools you sanctioned and the ones you never saw. Text is the one interface every kind of work shares, so the machine reaches wherever text goes. Which is everywhere.
This is recent. Eighteen months ago this was a rounding error; today it is the majority of what most institutions publish. The cost of producing it fell to nearly zero. The cost of confirming it is true, consistent, and approved did not move at all. That gap — between what an organization can now say and what it can still vouch for — is what just went vertical.
No one approved most of it. No one has read all of it. No one could. The institution has never said this much, and never known so little of what it is saying.
For most of institutional history, one person could hold an organization's commitments in their head: what had been promised, to whom, and whether it was still true. They could catch a contradiction before it went out. As organizations grew and their output scaled, that single point of coherence became impossible to sustain, and each new technology moved the institution's own words further beyond any individual's reach.
A single person could track the institution's commitments, figures, and decisions, and keep them consistent.
Records outgrew individual memory, but were still written and read by people, at human pace.
The institution began publishing continuously, to audiences that compare what it says across channels.
Output is generated faster than anyone can review it, across every channel, with nothing ensuring the parts agree.
Each of those documents is internally fluent. What no longer exists is anything underneath that holds them to one another, or to what the institution has already committed to.
The result is not error in the ordinary sense, which can be checked and corrected. It is drift: the same claim stated differently in three places, a figure that changes between reports, a commitment made last quarter that no one is tracking. The cost does not appear in any single document. It accumulates, and it falls on whoever has signed them.
The condition 01 → the shape 03
One true figure, propagated across every surface, drifting over time. Click any surface — alone, each reads fine. The fracture is only visible from underneath. Turn on the coherence layer to see what no single document can.
Drift looks like a content problem. It is the surface of a deeper one: a shift in where an organization's value is actually defensible, and it runs against the intuition most leaders are working from.
For most of computing's history, machines operated only on what a firm could write down: its files, its figures, its documented procedures. The valuable things that resisted codification stayed with people — the judgment calls, the read of a counterparty, the practiced sense of what good work looks like. Economists call this tacit knowledge, and because no earlier machine could reach it, it became the most defensible asset a firm held.
ExampleWhen an acquirer buys a company and the founders soon leave, it often discovers it purchased everything legible — the code, the patents, the customer list — and lost the thing it was actually paying for. The capability walked out the door, because it was never in the files.
The conventional view held that tacit judgment was safe because it could not be automated. But it was never protected so much as inaccessible: it could not be specified, so no rule-based machine could reproduce it. Modern AI removes that barrier. It learns from examples rather than stated rules, which makes it well suited to exactly the pattern-recognition we were never able to put into words. The capability that felt most secure turns out to be the most exposed.
ExampleNo one can write the decision rules for spotting a mispriced security or a weak legal argument. Yet a model trained on enough cases now performs both, without anyone ever having to state the rule it learned.
Once the work itself can be automated, what sustains a profession is no longer competence but permission: the legal requirement that a licensed human own the decision and bear the liability when it fails. The protection becomes institutional rather than technical. The professions that endure will do so because regulation, not capability, keeps a credentialed person in the loop.
ExampleA model can read a radiology scan about as accurately as a physician. A licensed radiologist still has to sign the read and carry the malpractice exposure. The signature — the assignment of liability — is the moat, not the diagnosis.
When fluent, competent output becomes essentially free, competence stops being the scarce input. What stays scarce is answerability: a party that will stand behind the work and has something to lose if it is wrong. A model can produce the work, but it cannot be held accountable for it — it cannot be sanctioned, sued, or made to care. Accountability is the one function that does not transfer to the machine.
ExampleTwo memos reach the same conclusion; one carries the signature of someone who is dismissed if it proves wrong. That assumption of consequence, not the analysis, is the part that cannot be automated.
The argument 02
Drag the reach of automation. The asset everyone treats as safe is the next thing taken — and the one thing that holds is held by a wall, not by height.
An institution is only what it can stand behind. Council is the layer that holds it to its word — built around the single thing the machine made scarce. The rest of this page is what that looks like in practice.
Council maintains the institution as a single coherent record beneath the places its people work — drafting, sourcing, reviewing, approving, and archiving. It performs three functions, and their order reflects their priority.
Council monitors for drift. When a current draft contradicts a prior report, when retired language reappears, or when a commitment comes due with nothing behind it, the system flags the conflict and routes it to the person authorized to resolve it. Detecting the institution's inconsistencies with itself is the core function; the other two exist to make it possible.
What the institution is bound to: not everything it knows, which stays with its people, but what it has committed to — approved figures, approved language, decisions, and promises, each with its evidence and the date it was last verified.
Every document is drafted from approved internal sources, routed through the institution's existing approval process, and logged as it is produced — preserving the speed of AI-assisted work without sacrificing consistency.
Drafting and retrieval are becoming essentially free; coherence and accountability are not. Council is built around the functions that remain difficult.
The category is crowded with adjacent products that resemble Council but address a different problem. Most fail on one of two dimensions: they are controlled by someone other than the institution, or they cannot detect when their own contents have stopped being true.
Built to speak for any institution, it is grounded in none: it draws on the open web, answers to no specific organization, and is fluent about the world in general rather than about yours.
Knowledge bases accumulate documents but are not maintained: they fill, go stale, and are rarely consulted. A record no one trusts enough to act on is not institutional memory.
A copilot answers the request in front of it and retains nothing once the session closes. It produces output, but keeps no durable record, enforces no consistency, and tracks no commitments.
A repository that cannot tell when its contents have gone out of date is only an archive presented with confidence — more hazardous than an acknowledged gap.
The institution remains accountable for what it says and does; that does not transfer to the system. Council's narrower role is to ensure the institution knows what it is saying across every surface, in time to act.
The internal model is proprietary, but its scale is worth stating plainly. Each figure below corresponds to a deliberately designed structure within the system.
The content of these categories, and the reasoning for this structure rather than another, is not disclosed here — but it is specified in full, and open to serious counterparties under NDA.
Prompted against the open web, the model fabricates freely.
Restricted to its actual figures, approved language, and prior decisions, the same model becomes exact.
That constraint is the technical premise of the system, and the less interesting half of the design. The more important half is what it makes possible: not better individual answers, but coherence across them, and a record the institution can be held accountable for. Constraint is how Council achieves reliability; it is not why Council matters.
The wedge is chosen for evidence, not revenue. Nonprofits are not the prize; they are the clearest place to prove the value is real, because the cost of incoherence surfaces fastest there. The same buyer profile — reputationally exposed, multi-channel, answerable to outside scrutiny — defines the institutions that can pay the most.
They produce a high volume of written material for audiences that actively compare it — boards, funders, regulators, the public — so the same claim must stay consistent across all of them. A visible contradiction costs not embarrassment but trust, which is effectively their only capital. They feel the pain acutely, they are demanding and candid partners, and it is the sector the founder knows best — the fastest path to undeniable proof.
Health systems, universities, public agencies, and firms under a compliance regime share the exact buyer profile and a far larger budget for it — they already spend heavily to keep their record defensible. Council is priced as infrastructure, per institution; the proof earned in the first market is the reference that opens these. The expansion is the business.
What they buyNot "answerability" in the abstract — a concrete job: the ability to answer "is this still true, and did we already commit to it?" before anything ships; faster approvals because provenance is automatic; and an auditable record of what was said and sanctioned, for the moment someone outside asks.
The fair question for anything in this space: what stops a frontier-model vendor or an incumbent copilot from adding "memory and consistency" and swallowing it? Three structural answers, in order of weight.
Council's asset is the maintained record, and it deepens every day it runs: each approved figure, resolved contradiction, retired phrase, and tracked commitment makes the next check sharper and the cost of leaving higher. A competitor starts blind; the incumbent holds years of the institution's own adjudicated truth. That is proprietary state, accruing per customer — a moat of one.
A copilot's job is to generate more, faster; Council's is to constrain and reconcile. A frontier vendor sells one model trained on everyone; Council's premise is a model bounded to one institution and answerable only to it. The posture that makes Council trustworthy — adversarial to its own output, surfacing the model's own contradictions — cuts against the generation vendors' business and indicts their hallucinations. They can bolt on memory; they will not take that stance against themselves.
Council becomes what an institution points to when asked what it committed to, and when it was approved. Systems of record are audited, budgeted as infrastructure, and load-bearing — not swapped like a writing assistant. A feature rides on top of the workflow; Council sits beneath it, and outlives any single tool above.
The logic the argument applied to professions applies to Council itself: the durable position is not the cleverest output but the one thing that accumulates and cannot be cheaply reproduced. Council is built to be that, for the institution it serves.
I am Ravi Narain. I developed the argument on this page and am building the system it implies, currently as a solo founder in Seattle. The narrow scope is deliberate: a bet this specific is better built deep than broad, with a small number of design partners rather than a broad launch.
Two kinds of conversation are useful now. The first is with organizations that recognize the condition in section 01 and would be willing first design partners. The second is with investors who think in terms of where AI creates durable value inside institutions — not where it lowers the cost of output. If either describes you, write directly.
I respond personally, if not always quickly.
Council is in early development, and access is limited while the first institutions are onboarded. A conversation can begin before the software is ready.