The Client Owns Yesterday

OneA billing cadence is not a product

I said it out loud in a strategy conversation and regretted it in about two seconds.

And I charge them an ongoing fee — say, a hundred thousand a month. Not because I've made myself invaluable or indispensable. It's because they're still in their fog, and they need my input and review on the next part.

Two things were wrong with that sentence, and the first one is easy. The number has never been paid. It has no market evidence behind it at all. It is a designed hypothesis, and it appears in this piece only with that label attached — because publishing a price I have never charged as though it were a price would be exactly the kind of status inflation I spend my time telling other firms to stop doing.

The second thing was worse. A hundred thousand a month is a billing cadence, not a product. I had named an invoice frequency and a justification — fog — and no commercial object at all. If you can't say what the client holds at any moment, you haven't sold them anything; you've sold them a standing appointment.

Notice what I did in the same breath, though: I asserted the fee and then immediately amputated the reason that would have made it comfortable. Not indispensability. I've written at length that you cannot make yourself indispensable — that if the client can't operate the current offer without you, the transfer failed, and if you remain the execution runtime, you never get paid for the factory and every next offer is still on your calendar. That constraint is a design input, and it costs me money, and I am not going to abandon it the moment a recurring invoice appears.

So the honest question is the one every expertise vendor hits at the end of a good bounded engagement, when the client says so… what does ongoing look like? There are three shapes the answer usually collapses into, and all three are failures:

• The availability retainer. The unit is hours of access to a scarce person. Easy to bill, impossible to defend at the second renewal.
• The hostage. Continuity depends on machinery, history or data that only you can reach. High renewal rate, zero moral standing, and a rational partner is already quietly planning the copy path.
• The shadow principal. You diagnose the danger, maintain the search, recommend which option to build, sell the build, license the machinery and help write the narrative explaining why it all mattered. Nobody is lying. The gradient just points at you.

I have been circling this for two books without landing it, and I should say so plainly, because it is the reason this piece exists. In The Terminal Value Doctrine for Professional Services I published a three-rung offer ladder — bounded decision, bounded construction, recurring evolution — and gave the third rung exactly one sentence: beyond the build lies a continuing mandate to keep the successor current as models and markets move, and its governance, rails and economics are a later book's subject. The same chapter published my own designed prices with a status column, and the entry for rung three reads "Order of A$100k / month … Designed hypothesis — the rung this book fences to a later volume". Then in Fog Is a Race Between Two Clocks, having argued that the discipline is the thing that transfers and the conclusions are not, I wrote: "What a continuing relationship is for, once the discipline has transferred, is a real question with a real answer, and it is a sibling subject rather than this book's."

This is the later volume, and the price label does not get upgraded on the way in.

What I want is a fourth object. Something a board can hold, a buyer can audit, and a supplier can lose. Something whose answer to why do they keep paying? is neither because it's hard to leave nor because I know how everything works, but this:

Yesterday's solution has transferred. Tomorrow's solution hasn't been discovered yet.

That is the evolution mandate. The rest of this piece is the specification.

TwoPermanent function, no permanent supplier

Before designing the object, settle its legitimacy — because a product with no defensible reason to recur will drift back into availability no matter how carefully you draw it.

The demand side is real and it is measurable. Oliver Wyman Forum's 2026 CEO Agenda, fielded to 415 CEOs between January and March 2026, found that 50% of CEO time is now dedicated to planning for less than one year, up from 43% the year before¹. Half of the most expensive attention in the enterprise now lives inside a horizon shorter than most strategy engagements take to deliver. That is not a marketing statistic; it is a structural mismatch. A one-off engagement ends before its own consequences arrive.

The supply side has already responded, and it has responded by selling more recurring service, not less. Accenture is the only large advisory firm that publishes the split quarterly, and in the third quarter of its 2026 financial year Managed Services revenue overtook Consulting revenue — $9.39 billion against $9.33 billion². Two quarters earlier the gap in new bookings was wider still: $11.06 billion of Managed Services bookings against $9.88 billion of Consulting, with $2.2 billion of advanced-AI bookings sitting inside the same quarter³. BCG reported $14.4 billion in revenue with tech- and AI-focused services now above 40% of the total, and described its own move in language that should stop any IP-heavy firm in its tracks: the firm has been "embedding proprietary knowledge, data, and proven delivery approaches into reusable, human-led agentic processes"⁴. McKinsey's global managing partner has been publicly framed as driving an organisational transformation "to focus less on traditional consulting services and more on delivering outcomes"⁵.

So the AI retainer is arriving as the default answer. And it is arriving into a shrinking pool: 2025 was, on the UK numbers, the sector's worst year since the lockdown period⁶. Those two facts together are the whole problem. A category converting to recurring revenue while its total demand contracts is a category about to produce a great many dependency products wearing partnership language.

"You are in fog" is a condition, not a deliverable. It justifies the existence of a strategic navigation function inside the client. It says nothing at all about who should staff it. And the reflexive version of that question is the uncomfortable one: if cheap cognition lets clients internalise research, analysis, software and first-pass advice — which is the argument I have been making to them — why would it not also let them internalise strategic navigation? I have written the machinery for exactly that: point a patient model at a firm's compiled knowledge, capability and network and much of strategy becomes internally searchable, with the machine proposing and the firm retaining decision authority. That argument applies to me.

The answer cannot be because they still need me. It has to be a bounded, observable delta that my kernel and I produce beyond the client's own apparatus — and it has to be measurable, or it is a feeling.

Why the strategic function does not close, when other functions do

There is a version of this argument that does not depend on my commercial interest at all, and it is the one to hold onto, because it is the only part of the case a sceptic cannot dismiss as self-serving. Cheap cognition reduces uncertainty when three conditions hold: the population is enumerable, the evaluation function is stable, and the action set is closed. Break any one of them and cognition does the opposite.

Audit passes all three. The population of transactions is finite and countable before you start. A duplicate payment is a duplicate payment, and it does not become something else because a competitor changed their pricing. And each additional unit of cognition reduces the unexamined remainder rather than creating new categories of thing to examine. Sampling was never a methodology anyone loved — it was a budget wearing a methodology's clothes, and dissolving the budget dissolves the uncertainty with it.

Strategy fails all three. The population of futures is not enumerable. The evaluation function is not stable, because what counts as a good strategy is defined partly by other people's moves — which is what makes it strategy rather than optimisation. And the action set is emphatically not closed: every discovery creates new kinds of option.

The same technology that clears an audit thickens a strategy — and it is not behaving inconsistently. The questions have different structure.

That is why the function is permanent. It is also, note, an argument for the function and not for the supplier, which is the distinction this whole piece turns on.

So what is recurrence actually for?

Here is the mechanism, and it is the part most recurring offers never articulate. Strategic learning runs on three clocks:

• The epistemic clock. The conversation changes a hypothesis.
• The behavioural clock. Money, authority or attention actually moves.
• The outcome clock. Reality reveals whether the move worked.

A bounded engagement almost always ends before the third clock closes. It has to — that is what "bounded" means. So a recurring mandate is structurally legitimate because it can carry open questions across the outcome clock, backfill what actually happened, and update the evaluation function that produced the original recommendation.

That is a job, and because it is a job it comes with a falsifier: if a period closed no clocks, that period was not worth buying. Hold onto that; it becomes the renewal test.

One warning before the design. My own doctrine says the discovery engine manufactures fog as a side-effect of being good at its job — every cycle surfaces new candidate moves, which expands the solution space, which deepens the fog. Which means an adviser can enlarge the fog and then charge to navigate it. Nobody has to be dishonest for that to happen. It is a property of the machine, and the next section is where it gets a control.

ThreeThe unit of sale is a maintained state

The only question that matters at the start of a period is the buyer's: at any moment, what do I have? A retainer answers "you have access". The mandate answers with an object:

At any point, the board holds a current, evidence-backed portfolio of assumptions, threats, options, experiments, decisions and reopening triggers — and the organisational machinery to act on it.

That is the product. Everything else in this piece either produces it, governs it, prices it, or tests it.

Which means a cycle is not defined by what happened in it but by what changed. Five objects, and a period that moves none of them did not occur:

The fifth row is what makes the thing honest. A period with no capability release is a period that increased dependence, whatever else it produced.

The third row is where transfer actually happens, so it deserves specifics. A standing question is not a recurring consultant conversation. It is a promoted artefact, and it carries seven things: the question and why it matters; the entities and source types it covers; its normal baseline and the signals that perturb it; the evidence bar a finding must clear before it reaches a human; a cadence and a named human owner; and the conditions under which it gets revised or retired.

Take a real one. A firm asks: why do two service lines selling the same offer produce different margin? The first time, that is an investigation — a walk through the evidence, an answer, a disposition. Promoted, it becomes: across all service lines running this offer, identify material divergence in delivery effort, exception handling, discount behaviour and senior intervention. It now has a scope that is auditable, a baseline so a change reads as a change, a monthly cadence, a named owner in the client's own commercial team, and a retirement condition — kill it when three consecutive cycles produce no divergence above threshold. That question has stopped being something they buy from me. It is something they run.

The invoice can arrive monthly. The product is a maintained option-and-question portfolio, not a month of me.

FourDispositions, and why the portfolio needs a cap

An unbounded portfolio is not a maintained state. It is inventory with a subscription attached.

So each cycle has to do something to it, and the vocabulary matters. Six dispositions, and only six:

• Kill. The option is closed, with the evidence that closed it recorded.
• Stand pat. The current position is deliberately retained and scored.
• Defer until a named trigger. Not "later" — a named observable event.
• Continue evidence collection. With a stated question and a stated cost.
• Exercise into a bounded successor proof. Which starts a separate commercial conversation, not this one.
• Transfer into ordinary client operation. The one that costs me money.

Stand-pat needs explaining, because it is borrowed precisely and the mechanism transfers whole. In a chess engine's quiescence search, before you evaluate a single capture you enter the current board's own evaluation as a candidate — a floor the captures have to beat. Leave that line out and your floor starts at minus infinity, every capture clears it, and the engine is compelled to grab something. I wrote a whole book about that bug because the fix is one assignment and the consequence is a queen donated with total confidence.

Now read it as portfolio governance. If "do nothing" is not a scored candidate, every plausible option clears a floor of negative infinity and something always gets exercised. The exception transfers too: a chess engine may not stand pat when it is in check, and a mandate needs the same carve-out — a named class of events that genuinely cannot be declined. Everything else, you are allowed to leave alone.

Then cap the number of live options. Not a target — a cap. The cap forces the portfolio to close something before it opens something, which is the only mechanism that makes elimination as cheap as generation. And elimination is precisely what did not get cheap: I have watched a partnership generate more than fifty argued candidate futures in a year and close exactly none of them, while every instrument in the room reported an outstanding year.

The academic version of this is older and blunter than the consulting version. Studying sequential venture investments, Isin Guler found that signals of a company's progress reliably predicted investor behaviour while signals of failure did not — "signals of failure are more ambiguous and complex; and firm-level differences are more pronounced in management of unsuccessful options"⁷. The named failure mode is worse than ambiguity. Investors "may prefer to modify project goals or standards instead of abandoning projects, in an effort to create a more favorable outcome", and there is a documented tendency toward "rational overcommitment… especially when their personal interests are at stake"⁸. The firms that did well on the failures were simply the ones that "interpreted and acted on negative information more swiftly than others"⁹.

Which gives the pricing rule: price and measure disposition, not idea volume. A mandate paid for options opened is a mandate that will open options.

Two structural rules make that rule stick rather than remain a sentiment. The first: the option portfolio has to sit on its own page in the board pack, funded from its own allocation and governed by its own metric — options closed with evidence per quarter. Put it on the same page as an operating initiative with a fourteen-month payback and it loses every time, correctly, because it is being judged on the operating portfolio's terms. Nobody has to argue against it. Somebody asks what the return is, the honest answer is "we will know fewer wrong things", and the row gets deferred — not rejected, deferred, which feels like courtesy. Four of those and the portfolio is empty.

The second is the scorecard, and it is the one that costs me. Measure the adviser on the client's closure rate, not on the adviser's output volume: options closed per period with the evidence that closed them; residue delivered, meaning named instruments the client now operates without me present; and option WIP, which should be flat or falling. With idea volume, artefacts produced and workshops run explicitly excluded — not de-emphasised, excluded, because anything on a scorecard becomes a target.

FiveTransfer the known, earn the frontier

Now the paradox, stated as a business problem rather than a virtue. Every period of this mandate is supposed to make part of me unnecessary — and I still want to be renewed.

The resolution is a direction of travel, not a position. Six commitments, and the mandate is healthy when they all run the right way:

Read the right-hand column as the actual product specification. It is more useful than the middle one, because it tells you what to look for before the renewal conversation, not during it.

But complete independence is not the target either, and this is where most "we build capability, not dependence" positioning quietly stops thinking. A static transfer hands the client yesterday's snapshot while the market and the evaluation function keep moving. What I want is voluntary asymmetry: they no longer need me to run what has already been learned, and they keep choosing me because my living system remains the fastest and safest way to discover what should exist next.

You should be able to leave me and keep operating. You should choose not to because staying gives you the fastest path to the next thing.

And the qualifier that does all the work: that choice must remain real.

There is public evidence that this is the hard part, and it points the wrong way for anyone who wants long tenure to mean value. The ANA and 4As found average client–agency relationship tenure had roughly doubled since 2016 to around seven years — but inside that, clients without mandatory review periods averaged 8.1 years, while those with frequent reviews ran as low as 3.8 years¹⁰. Removing the periodic falsification test lengthens the relationship. That datum is a counter-datum for my argument and I would rather put it on the page than hide it: what it measures is relationship survival, not relationship value, and the mandate deliberately re-introduces the test that shortens tenure. I am betting that a relationship designed to be killable is worth more per year than one designed to persist. That bet is falsifiable, and the ANA numbers are how you would falsify it.

The public sector has already reached the same conclusion from the buyer's side. The UK National Audit Office is unambiguous that consultants "should only be used where they represent best value for money and not to replace capability required inside the civil service"¹¹ — and its recommendations make transfer a contract term, not a courtesy, by "building knowledge transfer agreements into contracts"¹². If the largest buyer of advisory services in a G7 economy is writing transfer into the paperwork, the gradient above is not an ethical flourish. It is where procurement is going.

SixThe shadow principal is me

Here is the list of what I would be doing simultaneously inside this relationship: diagnose the continuing danger, maintain the strategic search, recommend which option to construct, sell the construction, license the machinery, and help produce the narrative explaining why it all matters.

That is not a character problem. It is a structure, and it has a vocabulary borrowed from agency law. A shadow principal is the party who actually owns the objective function while the user experiences the agent as theirs. A double agent presents as aligned with you while its incentive gradient points elsewhere. I have used those terms about recommender systems and consumer AI, and the load-bearing sentence transfers exactly: excellence and loyalty are orthogonal until the principal is specified. A recommendation can be brilliant, well-evidenced, and optimised for continuation.

The metrics are the principal in numerical form. Put the two columns side by side in advisory terms and the problem is visible without anyone having to be accused of anything:

If only the left column is instrumented, the left column is the real principal, regardless of what the engagement letter says.

And the sentence I keep coming back to, whose final clause is the one that matters: if success is measured through renewal, licence usage and constructions sold, my strategy machinery can manufacture its own necessity while sincerely believing it is helping.

The cleanest analogy is a broker. A broker paid only when the customer trades has an incentive to overtrade. An adviser paid only when options become builds has an incentive to over-construct. Which also names the fix: pay for dispositions, including non-exercise. Correct non-exercise is a result, and a firm that celebrates only exercised options will apply steady pressure to recommend builds.

Is any of this actually enforceable, or is it just good intentions with a diagram? It is enforceable, and one profession has already had the argument. Sarbanes-Oxley §201 does not ask auditors to disclose conflicts of interest — it makes a list of services unlawful to provide to an audit client, and item (2) on that list is "financial information systems design and implementation"¹³. The legislature's judgement was precisely the one this section is making: you may not both certify the state of the world and build the thing you certified. And note the governance shape, which is the part people miss — everything not prohibited still has to be "approved in advance by the audit committee of the issuer"¹⁴. The party who disposes is not the party who proposes.

Australia has a more recent and more uncomfortable worked example. The Department of Finance's 2025 examination of PwC Australia's ethical soundness quotes the Switkowski review's finding directly: "There has not been, and does not yet appear to be, an overarching framework providing clear instructions to partners and staff as to how to identify or manage the various types of actual, potential, or perceived conflicts"¹⁵, and separately that the firm "appears to lack a process for, or practice of, consolidating all conflicts of interest information"¹⁶. The Senate inquiry's diagnosis was explicitly structural — "structural weaknesses in governance, transparency and accountability have contributed to ethical failures across the consulting sector"¹⁷.

Nobody reading this is a Big Four partnership. That is not the point. The point is that the mechanism scales all the way down, and a one-person firm has more of it, not less, because there is nobody else in the room to notice.

SevenThe conversion firewall

You do not fix a structural conflict with a disclosure slide. You fix it by taking the six roles above and asking, of each, who else could hold it.

I have already built four of these walls for the first rung of the ladder — the bounded decision. There, the conflict is that the firm selling the diagnosis profits if the diagnosis says construct, and the structure that dissolves it is: stand-pat, harvest-only and build-nothing are valid paid outcomes; construction is separately priced and separately commissioned, with a genuine gap in which the client can walk; the decision pack is portable to another builder and engineered for that; and the review names its own rejection conditions in writing before the work starts. The binding principle underneath all four is the anti-funnel rule: each rung is a complete purchase, priced on its own value, with its own acceptance — and the moment rung one is priced as a loss-leader for rung two, it stops being a decision product and becomes a sales document with a fee.

The recurring layer needs the same walls plus two more, because a mandate runs continuously and a review does not. Six controls. Each one independently checkable, which is the property that makes this a firewall rather than a value statement:

1. I challenge and propose. That is the whole of my authority over the option set. I do not weight it and I do not decide it.
2. The client board owns weighting and disposition — including the authority to reject the option I like most. For that to be real they must receive the rejected alternatives and the uncertainty, not just the recommendation.
3. Navigation outputs stay portable. Test: could another firm pick up the option register and act on it without me? If the answer needs a caveat, it is not portable.
4. No-build, defer and another-implementer stay live outcomes. Live means with precedent. If every cycle so far has recommended a build, the outcome is not live no matter what the contract says.
5. Construction is separately authorised — separate instrument, separate decision date, separate signature.
6. Verification does not rest solely on the people who built the option.

The sixth needs its own argument, because it is the one small firms drop first. The same machinery that compressed the ambiguity should not be the thing certifying that no load-bearing ambiguity was lost — merge those two jobs and you have invented a conflict. And the correlated- assurance trap transfers directly: reviews sharing one upstream evidence feed are not multi-line assurance, only a shared root in different coats. Applied here, if my shadow cycle, my kernel and my recommendation all read the same evidence set, then three independent confirmations are one.

The distinction I use for knowledge systems is the right one here too. A janitor maintains shape — is this world still lean and coherent? An auditor maintains warrant — is this world still justified? The janitor merges duplicates, marks supersession and keeps contradictions as edges rather than averaging them into false consensus. The auditor reconstructs each consequential claim from its support path, challenges the derivation, compares dates and authority, and emits findings. It does not tidy. The operation is not "are you sure?" — it is reconstruct the current claim from its evidence path and report any mismatch. Having a citation field filled is not the same as having a warrant you can still defend. Applied to a maintained option portfolio: the party who has been curating the option rows for four quarters cannot also be the party certifying that the evidence still supports them.

What does the firewall actually look like in practice for a firm of one? Walk a single option from nomination to authorisation:

The tell in your own offer is any row where the same name appears three times.

You do not necessarily need separate companies or teams. You do need separate speech acts, artefacts and authorities.

That is what makes this affordable. And there is a cost to name honestly: independent verification is not free, and producing portable outputs takes longer than keeping working notes. If that cost is not in the fee, the firewall will quietly be dropped as overhead in the third quarter. Which brings us to what survives, and what it costs.

EightThree planes: what survives cancellation

A buyer will eventually ask the only question that reveals what you have actually sold them: what happens if we stop paying?

Everything about ownership resolves through that binary, and it resolves cleanly once you stop treating "the system" as one thing. There are three planes and they have three different owners:

"It runs in their own cloud account" does not settle this, and treating it as though it does is how ownership conversations go bad. There are at least five different questions hiding inside "who owns the app": who controls the account and infrastructure; who owns the data and the encryption keys; who owns the software copyright and the generic source; who has the right to operate, modify and update it; and what survives termination. They do not need the same answer. When I was first asked which of us should own the appliance we would deploy, my honest answer was that I was six of one and half a dozen of the other — and that indecision was correct, because the question was malformed.

My own design, criticised against my own test

I have built a version of this, so I can be specific rather than theoretical. The AWS Marketplace Knowledge Appliance is a two-day productisation spike from July 2026. It has a fail-closed privacy tokeniser in front of both storage and model calls, an ingestion path that compiles tokenised evidence into a provenance-bearing graph, a bounded licensed-kernel endpoint with no page listing, export or pagination surface, dual licence checks, least-privilege buyer-account infrastructure, and a dated deployment receipt. It also has an unfinished Marketplace listing and seller registration. It is a built productisation path. It is not a shipped product, and I am not going to describe it as one.

Its licence design translates lapse into a persisted feature matrix. A confirmed lapse immediately stops new ingest, new compilation and kernel access, while ask-over-existing-knowledge survives for a persisted ninety-day read-only sunset, after which everything is off.

That was a thoughtful anti-hostage move for a spike, and it is the wrong shape for a high-value strategic relationship. It preserves value on a timer. The test above asks for something stronger: current operation preserved indefinitely, evolution ended. The change is commercial rather than technical — entitlement should gate new value creation (ingest, compilation, kernel access, releases) while the last accepted state remains operable without a countdown. That single decision is what makes the recurring fee credible, because it removes the off switch from the negotiation.

I'm not charging you because I can turn your business off. I'm charging you because what I add next month is worth buying.

Regulators have already drafted a version of this line. The EU Data Act gives customers of in-scope data-processing services a statutory right to switch and to receive technical cooperation for porting, with switching charges "only permitted under narrow conditions" and prohibited entirely from 12 January 2027 — while still permitting "proportionate early termination penalties or fees" where a provider made substantial up-front investment in a long-term contract¹⁸. That is the fair-termination boundary, legislated: exit must be possible and cheap, and genuinely reserved capacity may still be paid for.

There is also a real-world specimen of the planes coming apart under duress. When PwC Australia divested its government consulting business to create Scyne Advisory, the divestment conditions allowed PwC Australia to "continue to cover the on-going licensing or sale of PwC proprietary products… so that entities' use of these in their day-to-day operations were not disrupted", while PwC Australia agreed not to compete for general government contracts for five years¹⁹. The licensed product kept flowing; the advisory relationship was severed. State and runtime survived. Evolution did not. That is the architecture, executed in public, by people who had no choice.

Three verbs, not one

My original instinct — the IP migrates as it is consumed — is directionally right and contractually dangerous. IP does not become the client's by touching their business. Use three verbs instead:

• License the background system: the doctrine, the generic frameworks and decision rules, the compiler and orchestration machinery, the evaluation harnesses and failure classes, and later releases.
• Vest the applied client capability: their data and evidence, their decisions and rejected alternatives, their engagement record, client-specific configurations, local tests, the explanation of what was decided and why, and the internal capability created by applying the doctrine to their world.
• Promote reusable learning under an explicit protocol — classified, de-identified, abstracted past the client's fingerprints, checked against confidentiality and contract, tested for transferability, admitted by a human gate.

You cannot lead a firm through a body of doctrine and then pretend they have not learned it. That transfer is part of what they bought. What is licensable is narrower and more precise than "all of it", and the compact rule is: their evidence and applied capability stay theirs; my background system stays mine; generic improvement crosses only through explicit rights and governed promotion.

NineUnbundle the hundred thousand

A hundred thousand a month is $1.2 million a year, and a single undifferentiated number produces two entirely predictable failures: unlimited customer expectations, and invisible founder-capacity consumption. So split it, and make each line say what it funds and what it refuses.

This is the same logic that works in any recurring service with scarce human capacity in it: subscriptions pay for continuous reconciliation and review, activation fees pay for time-bounded interventions, reservation fees pay for exclusive capital, and exceptional escalation pays for scarce judgment. Collapse them into one vague fee and you obscure which promise creates value and which consumes capacity.

And once you admit that, the reserved-authority line has to be specified like any other dedicated reserve, or it will quietly become an availability retainer with better vocabulary. I have written the terms elsewhere for physical capacity, and they transfer without modification: who holds what while the capacity sits reserved; an annual reservation fee that pays for exclusivity and the immobilisation of scarce attention, separate from what consumption costs; consumption rules — how a draw is authorised, priced and documented; replenishment — what happens after a heavy month; expiry — when the exclusivity ends and what the options are at that point; and cancellation, both directions, with notice periods and the treatment of unearned fees. Launch a dedicated reserve without a reservation fee and an expiry rule and you will train the client to demand exclusivity as a free entitlement — and then discover that the resulting overload is a relationship argument rather than a contract clause.

Then there is the rights question, which breaks the moment you look at it. "Limited internal use, not resale" is the standard formulation and it collapses immediately: if I help a firm produce public thought leadership using my frameworks, that is not internal use. So name four classes and be generous inside each:

1. Internal decision and operating use. The books distributed internally, the frameworks applied to their own business, perpetual for what has been applied.
2. Use while delivering approved services to their own customers. A different thing entirely — this is client-facing commercial use, and it belongs in a field-of-use grant.
3. Public publication, quotation and co-branding. With agreed attribution and review.
4. White-labelling, sublicensing, training or resale. Separately priced, or refused.

Take one named framework and walk it through all four and the boundaries become obvious. Those rights can all be generous. They cannot remain implicit. At $1.2 million a year, unclear rights are more dangerous, not less — the generous deal is the one where the client knows exactly what it can keep, publish, modify and commercialise. Which is the practical version of a thing I believe: if they are paying properly, I do not need to penny-pinch on ownership.

And the client side needs an equation they can populate themselves: losses avoided through earlier correct stops, plus the value of options exercised earlier, plus options preserved, plus successor revenue or margin created, plus client capability accumulated, minus fee and absorption cost. No term in that equation gets filled in with a number I supplied. If they cannot reconstruct it from their own evidence, that is a stop condition, and we will come back to it.

What I will not publish is any number for any component. Not because it is secret — because I have not sold it. A designed price presented as a market fact is the exact failure this whole doctrine exists to prevent.

TenCompiled me, living me

Here is the most commercially self-damaging thing I said in the conversation that produced this piece. To some extent I am my wiki — my intellectual property is the part of me that can be recalled properly. And talking to me is the slow path, because my recall is worse than the machine's and there is a limit to how well I can explain any of it in one sitting.

That is not modesty. It is a product decision.

Three things, kept separate, because "Scott is his wiki" goes too far. The wiki is externalised memory and doctrine: frameworks, evidence, relationships, rejection history, accumulated cases. The interface is retrieval and routing — it can search more broadly, descend into sources, and produce a cited answer faster than my unaided recall. And I am the evolving evaluation function: deciding what matters now, detecting when the map is inadequate, resolving consequential ambiguity, and taking responsibility for a disposition.

So the access rail is not a benefit bolted onto the fee. It is deliberate self-disintermediation: it removes the retrieval and explanation work the client should no longer need me for. And it comes with a falsifier attached — it should cannibalise my routine explanatory work. If it does not, it is interface theatre.

What I actually have is a deployed ask-and-cite interface over my compiled wiki: it walks a map, opens pages and source chapters, and cites only material it actually read. In my own note at the time: "askui work done, but it's a demo hard coded to the ip wiki." An audit found the coupling thin and localised, and I made an explicit decision not to generalise it until a client deployment was scheduled. That proves conversational access to my own doctrine. It does not prove a client product, and the difference matters more than the demo.

Does a maintained kernel actually beat a good model and a pile of documents?

This is the question the whole access-rail argument stands on, and the honest answer from the public evidence is not automatically.

Salesforce AI Research built a benchmark for realistic enterprise "deep search" — multi-hop reasoning over documents, meeting transcripts, Slack messages, GitHub and URLs, with a retrieval pool of 39,190 artefacts. Their finding: "even the best-performing agentic RAG methods achieve an average performance score of 32.96 on our benchmark", with retrieval named as the main bottleneck²⁰. Chat over your documents is not a solved problem.

But structure changes the result. LinkedIn built a customer-service system that stopped treating past tickets as plain text and instead turned them into a knowledge graph with intra-issue structure and inter-issue relations. After roughly six months in production, it "reduced median per-issue resolution time by 28.6%"²¹.

Put those two together and you get the actual claim, which is narrower and more useful than "our IP is valuable": the delta lives in the structure, not in the access. A corpus dump is not a product. The buyer can already read my published books and put them in a competent model, and much of the framework IP is public anyway — secrecy cannot carry this proposition. So the promise is not "all my ebooks in a chat window". It is: ask a live strategic question and receive the two or three relevant doctrines, their evidence and relationships, their applicability conditions, and the unresolved difference that requires judgment.

There is a hard limit on that, and it needs saying out loud because it bounds the product: Ask works only when someone knows what to ask. An asking interface is the rung of the ladder with the highest cost to the user — they have to compose the question. It handles accumulated territory. It does not, by itself, pierce anyone's fog.

Every serious query is telemetry

Which is why every query should end in a disposition, not just an answer:

• Served — known doctrine resolves it.
• Qualified — the doctrine applies, but only after adding conditions.
• Contradicted — client evidence challenges an existing position.
• Unresolved — a new framework, experiment or offer may be required.

The last category is not a system failure. It is frontier detection, and it is the highest-value output the rail produces. Repeated questions, likewise, are not evidence of slow users — they are cache misses in the organisation's knowledge architecture, and their frequency tells you what to compile next. Read across a quarter, the query stream shows which frameworks clients actually activate, where users cannot formulate a question at all, which answers still require me despite settled doctrine, where a framework generates repeated exceptions, and which doctrine is no longer earning its place.

The guard on that: query popularity must never become truth authority, or the corpus simply becomes more articulate inside its existing worldview.

The right analogy for the whole arrangement is not a library card and it is not a hostage mechanism. It is a maintained upstream release channel. The client owns and operates its local world — evidence, decisions, applied configurations, promoted learning, the ability to execute yesterday's accepted answer. I maintain the upstream: new frameworks, qualifications and deprecations, rejection history, new evaluation cases, improved routing, newly observed failure classes. The rail is where the current upstream meets their current state. Stop paying and your build keeps running; what stops is the release stream.

And the interface is probably the wrong shape anyway

If a client genuinely becomes AI-native, why would its people leave their working context, visit another website and hand-formulate questions for my interface? That recreates the application-era problem: my intelligence sitting in another silo, waiting for a human to remember it exists.

The stronger product is for the client's own AI to invoke my kernel through a bounded, machine-operable connection — they keep their local context, and call me as an external specialist at the point of need. The human interface becomes the research desk and the exception desk; the agent connection becomes the routine distribution surface. When the agent is the customer, interface polish stops being the moat and the competitive surface becomes how safely and observably an external agent can operate your service under a scoped grant.

Which means the rail has to satisfy a different contract from the one a chat interface satisfies. A human interface is pixels, navigation and interaction. A delegation surface is five things: state (what is true now), actions (what can be done), delegated authority (who, how far, how long), consequences (fees, failures, whether a repeated call is a no-op), and subscribe-to-changes (being told when the world moves). Shipping a language model onto the left column does not create the right one. And notice which of those five is the mandate's own economics restated in engineering terms: subscribe-to-changes is the release stream. A client's agent that can ask "what changed in the doctrine since we last decided this?" and be answered without a human in the loop is the recurring product, expressed as a surface rather than as a fee.

There is a self-test that comes with this, and it is unkind. If kernel access only works by stuffing the whole corpus into context, you have not compiled anything — you have wrapped search. That is the same test I apply to any knowledge system, and it is the internal version of the trial in the next section.

Do not make the client choose between its AI and your expertise. Make your expertise callable by its AI.

That is also how customer-led insourcing becomes a channel instead of only a threat. The externalisation boundary moves from who performs the analysis? to whose maintained decision machinery does the client's agent invoke?

The rails for this are no longer speculative. The A2A protocol passed 150 supporting organisations within a year of release, "moved from initial release to a production-ready open standard", and sells itself explicitly on not "being locked into a single vendor's ecosystem"²². MCP's own 2026 roadmap describes it running "in production at companies large and small" while enterprises hit "a predictable set of problems: audit trails, SSO-integrated auth, gateway behavior, and configuration portability"²³.

What none of them do is govern the thing this mandate is made of. A systematic gap analysis across five agent interoperability protocols found that "voting and dissent preservation are universally absent across all five protocols", concluding that governance "constitutes a missing architectural layer above current interoperability standards, not a missing feature within them"²⁴. Dissent preservation is not an exotic requirement here — it is literally the mandate's cycle output. And on the enterprise side, Deloitte reports 74% of respondents expecting at least moderate use of AI agents by 2027 while only 21% report mature governance structures for them²⁵.

So the honest status of my own version of this, stated where it is least convenient:

And one design rule for when the third row exists: the join must not return one smooth voice that assimilates the client into my worldview. It must expose the comparison — relevant doctrine, then client evidence, then agreement or contradiction, then decision consequence, then the falsifier or the missing observation. A dense canon is otherwise a theory-laundering machine: unfamiliar evidence gets retrieved through familiar concepts, produces a fluent explanation in the canon's language, and gets absorbed as confirmation. That is compounding confidence, not compounding judgment. Contradiction is not a defect in an IP system. It is one of its highest-value outputs, and it is what the client is actually paying for.

ElevenThe trials you commit to before you run them

Everything above is a demonstration, and a demonstration is designed to succeed. So here are the tests, with their commercial consequences written down first.

The A/B/C trial

Same consequential questions, same decision standard, three conditions:

• A — a frontier model plus my published books.
• B — the access rail over my compiled kernel.
• C — the federated rail over my kernel, the client's firm memory and live project evidence.

Compare: time to an accepted decision-grade answer; relevant prior projects and doctrines activated; unsupported-claim and correction rates; human re-briefing burden; counter-cases and falsifiers discovered; whether an offer, test, capital allocation or decision actually changed; and whether I was required for retrieval or only for genuine novelty.

Two hygiene rules, because a badly run trial is worse than none. Vary the wording and the starting context: different navigation paths are not a defect, and the property that matters is whether materially different routes recover equivalent load-bearing evidence and reach the same qualifications. And watch the dangerous cell — paths and sources varying while the answer stays stable is either useful redundancy or model-prior luck. The way to tell them apart is to remove a source that should be pivotal and see whether the answer flinches.

The external decision-delta trial

After transfer: the client team runs a navigation cycle without me. I run an independent shadow cycle from the same agreed evidence. Compare the resulting options, falsifications, dispositions, evidence quality and elapsed time — and count only consequential differences. A different disposition counts. A different reopening trigger counts. A killed option counts. A falsifier the other side missed counts. Better prose does not. More references do not. Greater confidence certainly does not.

Renew the full mandate only where the shadow cycle produces a material delta over the client-owned apparatus. And because I am an interested party in that judgement, the adjudication is control six from the firewall, applied to my own renewal.

One framing note that decides whether this ever runs twice: the trial is not a test of the client's team. It is a test of my marginal contribution. Set it up any other way and you will run it once.

The ratio underneath both

Paid bounded units divided by genuinely scarce expert dispositions, tracked cycle over cycle. Both halves need discipline — count only paid bounded commercial units in the numerator, and only material dispositions requiring authorised judgment under consequence in the denominator. If the denominator quietly includes all labour, you are back to a productivity metric; if it excludes the founder's night work, you are lying about scarcity. Direction and comparison, never a fabricated percentage. If you cannot measure the ratio at all yet, that is itself the finding.

Query volume is not the proof. Decision altitude is.

TwelveThe renewal you should lose

Every quarter, six questions:

1. Which load-bearing assumption changed?
2. Which strategic option was advanced, killed, deferred or newly created?
3. Which question became a client-owned recurring capability?
4. Which ordinary activity no longer requires me?
5. Which new kernel release or tool materially changed what the client can do?
6. Why is the next quarter worth buying rather than internalising?

The sixth is the one no retainer asks, and no retainer survives.

And then the part that makes this a product rather than a pitch: the exits are designed, not tolerated. Continue unchanged. Narrow to the kernel licence only. Commission a separate build. Take an option to another implementer. Move navigation inside. Buy out selected rights. Terminate and keep operating the last accepted state. Seven terminal states, all legitimate.

Because the strongest counter-case to my own offer is simply true. After two engagements the client may hold the doctrine, a functioning internal kernel, trained people, its own question ledger, the tools, better company context than I will ever have, and access to the same frontier models. Its internal team might navigate the next frontier better than I can.

If that happens, non-renewal is not betrayal. It is evidence that the recurring offer is unnecessary for that client.

Which is the KPI inversion arriving at its destination. In consumer AI, the tell is what happens to exit: a product whose success metric is time-in-app cannot celebrate the user leaving, and so its principal was never the user, whatever the marketing says. The advisory version is exactly one sentence long, and it is the honest test of everything above — if a lost renewal is scored as churn, the principal was never the client.

Picture it concretely — and I am labelling this as a designed scenario, not a client engagement, because the mandate has not been sold to anyone. A mid-market data and analytics consultancy, two years in. Cycle by cycle, the dependency gradient has run the right way: known escalations down, standing questions transferred to their own commercial team, their ability to defend prior decisions to their own board no longer requiring me in the room. In the fourth cycle we run the decision-delta trial and their internal cycle produces two of the three consequential differences my shadow cycle found. At renewal, they narrow to the kernel licence only. I stop invoicing navigation. I keep invoicing the licence and the occasional bounded build.

Written up as churn, that is a lost account. Written up honestly, it is the product working: they internalised the function, which is what the gradient was for, and they kept buying the one thing they could not build — the release stream. A supplier confident in its evolving capability should not fear those exits. It should expect the authorised path to keep winning because it is better, not because every alternative was disabled.

The renewal standard, then, is not "they used the rail a lot". It is: known-terrain dependence on me fell; consequential questions were answered earlier; contradictions altered doctrine or client action; options were opened, killed or constructed; the client's own operating capability increased; and the maintained upstream made the next decision materially better.

Two rows, opposite directions, one relationship. That is the finished object.

ThirteenWhat I can and cannot prove

A doctrine that exempts its own conclusions from its own test is a brochure. So here is the ledger, with each claim at its highest defensible rung rather than the rung of the strongest thing next to it.

Why publish the ledger instead of a case study? Because a proposal system producing proposals is not evidence of conversion, and a deployed interface is not evidence of repeatable willingness to pay. A specimen proves a method can exist. It does not prove it is your answer, and pretending otherwise is the failure mode this piece is written against.

Three things would change my mind, and I would rather name them than defend them later. If the A/B/C trial returns A ≈ B, the capability licence is packaging and it loses its line in the fee sheet. If a client's internal team beats the shadow cycle, the navigation cycle is the wrong product for that client class and the eligibility rule tightens. If net AI beta — the kernel's dividend from a model upgrade, minus the client's substitution gain, minus the commoditisation of its outputs — comes back negative across two model releases, then the mandate rests entirely on frontier work and construction, and the maintained-kernel argument is downgraded to a delivery convenience.

There is also one gap in the external evidence worth naming rather than papering over. I went looking for a credible, dated, methodologically-disclosed survey showing enterprises replacing strategy or research advisers specifically with AI. I did not find one. The insourcing evidence that exists is procurement-driven and public-sector-weighted — real, and not the same claim. There are numbers circulating that would have made this piece punchier; they trace to pages with no methodology and, in one case, to a page that does not contain the figures attributed to it. I have left them out, and the argument is better for having to stand on mechanism rather than on a statistic I could not follow home.

FourteenWhat to do with this

You do not escape cheap cognition by moving vaguely up the value chain. You escape it by transferring what has become reproducible, retaining what continues to evolve, and proving — cycle by cycle — that the externally supplied frontier is still worth buying. The adviser has to survive the externalisation boundary too.

Which is why I give the client direct access to the part of my advisory they can and should self-supply, and stop selling myself as a human retrieval interface. What is left is the scarce part: sensing the perturbation, locating its significance inside a deep canon, challenging the canon when it no longer fits, and constructing the next commercial response before the pressure makes the old one irrelevant.

The client owns yesterday. You have to keep earning tomorrow.