Token Capital: the asset your balance sheet cannot see

On a Tuesday afternoon in Rotterdam, a trade finance analyst at a consumer goods distributor resolves a retailer deduction. A grocery chain has short-paid an invoice by €14,000, citing a promotion that ran in week 37.

The analyst asks an agent to investigate. The agent pulls the promotion agreement, reconstructs the entitlement calculation, finds that the retailer double-counted a display fee already settled in week 33, drafts a rejection with the evidence attached, and routes it for approval. The analyst reads the reasoning, corrects one assumption about a regional listing fee, and approves. Eleven minutes, start to finish. The old process took two days.

Here is the question almost nobody asks: what does the company own on Wednesday morning?

If the answer is “a resolved claim and a productivity statistic”, the company rented intelligence for eleven minutes and kept nothing. If the answer is “a corrected reasoning trace that makes the next thousand deductions cheaper, faster, and more autonomous”, the company just formed capital.

Same model. Same task. Same eleven minutes. Entirely different economic events.

That difference is the subject of this essay, and I will argue it is the most consequential distinction in enterprise strategy this decade. Every company is now spending on AI. Almost none of them can answer the only question that will separate the winners from the spenders: what remains inside the firm after the intelligence has been consumed?

I call what remains token capital. This essay defines it precisely, explains why your accounting system is structurally blind to it, shows how it compounds, shows how it leaks, and gives you the instruments to measure and govern it. It is a long argument, but it ends in a single sentence you can take to your board, and I will get you there.

Start with an uncomfortable fact about the models themselves: they are converging into a commodity input.

Frontier models are astonishing, and they are also increasingly interchangeable at the point of enterprise use. Your competitor can call the same API you call. The capability gap between the best model and the second-best model — measured on the tasks that actually matter to a deduction analyst or a demand planner — is shrinking quarter by quarter, and the price per unit of intelligence is collapsing faster than any industrial input in history.

When something becomes abundant, cheap, and available to everyone at the same time, it stops being a source of advantage and becomes a utility. Electricity. Bandwidth. Cloud compute. Now thinking.

History is unambiguous about what happens next, and the precedent is worth thirty seconds, because it predicts the next ten years of enterprise AI with unsettling accuracy.

When factories first electrified at the end of the nineteenth century, they replaced their central steam engine with a central electric motor and changed nothing else. Productivity barely moved, and it stayed flat for nearly thirty years. The economist Paul David showed why: the gains never came from the dynamo. They came decades later, when factory owners finally redesigned the factory around the new input — replacing the single drive shaft with small motors on every machine, and reorganizing the floor around workflow instead of around power distribution.

The value was never in the electricity. The value was in everything the firm rebuilt downstream of the electricity.

The same was true of the last technology wave. Nobody sustained an advantage by “having cloud”. Advantage came from what companies built on top of it, and from the operational knowledge encoded in how they built it.

The firms that win will not be the ones with access to intelligence, because everyone will have access. They will be the ones that build the largest, best-governed stock of firm-specific assets that convert generic intelligence into trusted, machine-executed work.

That stock is token capital.

Definitions of new concepts tend to be either poetic or useless. This one is neither.

Every word in that sentence is load-bearing.

Accumulated stock

Token capital is not this month’s API bill or this quarter’s pilot. It is what survives the interaction. Spending is a flow. Capital is a stock. Confusing the two is the defining accounting error of the AI era, and I will come back to it.

Firm-specific

An asset your competitor can buy tomorrow is not capital — it is procurement. A model subscription is not token capital. A prompt copied from LinkedIn is not token capital. Your promotion agreement ontology, your ten thousand expert-resolved deduction cases, your validated entitlement calculator, and the recorded judgment of your best key account manager encoded into an evaluation rubric: those are token capital, because nobody else can have them.

Governable

An asset you cannot control, audit, version, restrict, or delete is not an asset — it is an exposure. This is where token capital differs from the older, softer idea of “organizational knowledge”. Knowledge in someone’s head is human capital: valuable, perishable, and it resigns when they do. Token capital is that same judgment made durable, inspectable, and executable by machines, under the firm’s control.

Trust, without repurchasing the judgment

This is the phrase that does the real work, and it leads to the idea at the centre of this essay. The output of token capital is not answers. Answers are cheap now; any model produces answers. The scarce output is trusted execution — work the firm is willing to let happen with progressively less human review.

Token capital is, in the most literal sense, trust converted into an asset.

The five-question test

There is a simple test for whether something you built is actually token capital. Five questions. Five yes answers means capital. A single no means you bought a tool or ran an experiment — both fine, neither compounds.

Now the second claim: the most important asset being created inside companies today appears on no balance sheet, in no management report, and in no board pack, anywhere.

This is not because CFOs are negligent. It is because accounting was engineered for a different economy. Financial statements are superb at representing assets you can drop on your foot: property, inventory, machinery. They handle purchased intangibles adequately. But internally generated capability has always been accounting’s blind spot — under both IFRS and US GAAP, nearly everything a firm spends to build its own knowledge assets is expensed immediately and vanishes into the P&L. Baruch Lev spent a career documenting the consequence: as intangibles became the dominant source of corporate value, financial statements explained less and less of what companies were actually worth.

AI spend inherits this blindness completely, and then makes it worse. Every euro of model subscriptions, agent platforms, integration work, and evaluation engineering hits the P&L as cost. The capability those euros create — if they create any — is recorded nowhere.

In the absence of a real measure, companies measure what is easy — and what is easy is activity. Licences purchased. Prompts submitted. Tokens consumed. Pilots launched. “Hours saved.” I want to be precise about what these numbers are.

Stop asking how much we spent on AI. Start asking what we own because of it.

The blindness is also the opportunity. Markets eventually reprice what accounting cannot see; they did it for brands, for software, for data. The firms that build instruments to see token capital now — years before their competitors and decades before the accounting standards — get to compound in the dark. The rest will discover the category when an acquirer’s due diligence team prices it for them.

What is token capital physically made of? Resist the urge to list forty artifact types; taxonomies are where strategy goes to die. There are five layers that matter, and they form a stack — each one multiplies the value of the layers below it.

1 · Context — what the firm knows, made machine-legible

Every company sits on decades of operational truth: agreement hierarchies, customer structures, promotion mechanics, pricing waterfalls, accrual states, the difference between what the contract says and what the relationship tolerates. Almost all of it is trapped in formats built for human eyes, or buried in the heads of people who will retire.

Context capital is that truth — structured, current, permissioned, and connected — so an agent reasons from your reality instead of from the internet’s averages. This layer is unglamorous, and it is where most token capital programmes actually succeed or fail, because a brilliant agent with poor context is a confident liar with system access.

2 · Capability — skills and tools that do the work

A skill is a validated procedure: classify this claim, reconstruct this entitlement, detect this duplicate, draft this settlement. A tool is governed access to a system of record.

The critical word is validated. An unvalidated prompt is an opinion. A skill becomes capital only when it has been tested against expert judgment, versioned, and certified for a defined scope. The difference between a prompt library and skill capital is the difference between a folder of recipes and a certified production line.

3 · Memory — the trace corpus

Every time an agent works and a human corrects it, something economically precious is created: a trajectory. The situation, the reasoning, the mistake, the correction, the outcome.

A single trace is trivia. Fifty thousand curated, outcome-labelled traces from your own operations are a proprietary dataset that no competitor and no model vendor can reproduce, because the events only ever happened inside your firm. This corpus is the fuel for everything else: better evaluations, targeted fine-tuning of small models, and the evidence base for extending autonomy.

It is also, not coincidentally, the asset firms most frequently give away for free. I will come back to that with some anger.

4 · Evaluation — the firm’s judgment, industrialized

Evaluation capital is the machinery that decides whether machine work is good: golden datasets of expert-resolved cases, calculation tests, prohibited-action tests, drift monitors.

This is the most underrated layer in the entire stack, for one strategic reason: evaluations are what make you model-agnostic. A firm with a rigorous evaluation harness can adopt any new model in days, with proof of safety, and can keep its providers in permanent competition with each other. A firm without one is married to whatever it first deployed — and does not even know how well it is performing.

5 · Codified judgment — the tacit made explicit

At the top of the stack sits the rarest material: the decision patterns of your best people. When does a great key account manager settle a disputed claim rather than fight it? Which retailer’s forecast do you quietly discount by eight percent? What does “acceptable risk” mean here, in numbers?

For a century this knowledge lived only in apprenticeship and left in resignation letters. Agentic systems force it into the open, because an agent cannot inherit judgment by sitting near someone for five years — the judgment has to be written down, encoded in policies, thresholds, and rubrics. This is the layer where token capital and human capital fuse, and it deserves its own statement.

Token capital does not replace your experts. It is the first technology in history that lets their judgment outlive their tenure.

Here is where this essay departs from every other treatment of AI value, so I will slow down.

Everyone measuring AI today measures productivity: hours saved, drafts accelerated, tickets deflected. Productivity is real, but it is the wrong unit of account, because it measures the human getting faster, not the firm getting more capable. The right unit of account for token capital is autonomy earned: how much consequential work the firm can safely delegate to machines, at what threshold, with what evidence.

At TiMiNa we operationalize this with the Graduated Autonomy Ladder, the framework we use in every agentic deployment. Every agent capability sits at exactly one of five stages.

The ladder was designed as a safety instrument. It turns out to be something more important: it is the valuation instrument for token capital.

A capability’s position on the ladder is a precise, auditable statement of how much organizational trust it has earned — and trust is exactly what the asset is made of. Promotion up the ladder is capital formation, and it can never be granted by enthusiasm or vendor promises. It is purchased with evidence, which means it is purchased with the lower layers of the stack. You cannot move a deduction agent from Propose to Execute Under Threshold without a golden dataset proving its accuracy, a trace corpus showing its failure modes, and controls governing its blast radius.

The ladder turns the entire abstract discussion of “AI maturity” into a hard question with a hard answer: what fraction of this workflow’s decisions execute at what autonomy stage, and what evidence file supports that placement?

Writing the value down

This gives us something the token capital conversation has been missing entirely: a way to write down value.

Every term in that expression is measurable. None of them appears in any AI dashboard I have ever seen inside an enterprise, all of which are busy counting prompts.

Governance is not the tax. Governance is the factory.

One implication deserves to be pulled out, because it inverts how most executives think about AI risk.

An ungoverned agent is permanently stuck at low autonomy, because no rational firm will extend trust without evidence and controls — which means an ungoverned agent is an asset that cannot appreciate. The compliance function, of all things, turns out to hold the keys to the capital account. Firms that treat AI governance as friction will remain in pilot purgatory forever, and they will deserve it.

Capital is only interesting if it compounds. Token capital compounds through one loop, and the loop is worth stating plainly: work produces traces; traces plus expert correction produce validated assets; validated assets earn autonomy; autonomy produces more work at lower cost; and the cycle turns again.

Each rotation makes the next rotation cheaper. That is the flywheel, and everything in a token capital strategy exists to spin it.

The number to start tracking on Monday

The compounding shows up in a single figure I ask every client to start tracking, because it is the best summary statistic of an agentic transformation: cost per trusted decision.

Take one workflow — deduction resolution. In the manual world, a resolved claim costs, conservatively, €35 of analyst time. Deploy an agent in Shadow and the cost initially rises, because you are now paying for the machine and the human. This is the phase where impatient companies quit. It is also exactly the phase where capital is forming fastest, because every human correction is minting trace capital.

Move to Propose and the cost per decision falls to perhaps €20. Earn Execute Under Threshold for the routine seventy percent of claims and it falls to €8. Add a specialized model trained on your own corpus and it falls again.

That is not a productivity curve. It is a capital accumulation curve wearing a productivity costume.

What happens to the second workflow

The second, more strategic effect is what happens to the next workflow — and this is the part vendors never explain, because it is not in their interest to.

The first agentic workflow is expensive. You are building identity infrastructure, tool gateways, evaluation harnesses, governance processes, and organizational muscle from zero. Perhaps eighty percent of that investment is not workflow-specific at all. So the second workflow — promotion settlement, say — inherits the platform, the customer ontology, the agreement context, the audit patterns, and a leadership team that now knows what good looks like.

In our engagements the second comparable workflow typically costs thirty to forty percent of the first, and reaches equivalent autonomy in half the time. By the fifth, marginal capabilities are being assembled substantially from existing token capital rather than built from scratch.

This is increasing returns — the signature economics of intangible capital — and it produces the strategic conclusion of this section.

Two years of head start in trace accumulation and evaluation infrastructure is not two years of advantage. It is an advantage that widens every quarter it is left unanswered, in exactly the way a competitor’s factory never did.

I will not soften this section.

Token capital has a sign. It can be negative. Right now, inside most companies, it is.

Consider what actually happens when your employees do serious work inside consumer AI tools on default settings, or when your agent platform’s architecture routes your reasoning traces into someone else’s improvement pipeline.

• Your demand planner explains, in a beautifully specific prompt, how your company adjusts forecasts for Ramadan in your Middle East markets.
• Your key account manager pastes the negotiation history with your largest retailer to draft a counterproposal.
• Your pricing analyst walks a model through your margin waterfall, exception by exception.

Each of those interactions produces exactly the asset this essay is about: context, judgment, corrected reasoning. The only question is who keeps it.

When the trace lands outside your governance boundary, you have not merely failed to form capital. You have transferred it. The tacit knowledge that took your organization decades to develop — the judgment that constitutes your actual competitive advantage — flows out one helpful prompt at a time, into systems that serve you, your competitors, and eventually your substitutes, identically.

And here is the detail that should make every chief executive sit up: you are paying subscription fees for the privilege. The polite term for this is capability leakage. The accurate term is that companies are currently funding the largest uncompensated transfer of institutional knowledge in commercial history, and booking it as a productivity win.

What the answer is not

I want to be careful and honest here, because this argument is often made stupidly. The answer is not to ban the tools — that is how you get shadow AI, which leaks worse and governs never. The answer is not paranoia about model vendors, most of whom now offer serious enterprise controls.

The answer is architectural, and it fits in one principle.

Own the trace.

Route agentic work through infrastructure where the interaction record, the corrections, and the learning belong to you — contractually and technically. Keep your context layer, your evaluation harness, and your trace corpus in assets you control, and treat the model behind them as a replaceable component. Sovereignty over token capital does not mean building your own models. It means making sure that when learning happens, it accrues to your balance sheet rather than your vendor’s.

Every capital asset decays, and pretending otherwise is how balance sheets lie. Token capital decays in ways executives need to internalize, because some of its half-lives are startlingly short.

That fourth row is the dangerous one. Capabilities without evaluation harnesses do not just decay — they decay invisibly, which is far worse. And trust, the substance of the asset, depreciates fastest of all: a single badly handled autonomous failure can knock a capability down two rungs of the ladder overnight, destroying in a day what took a year of evidence to build.

False capital

Distinguish decay from something more dangerous. False capital is the stuff that looks like token capital in a steering committee deck and fails the five-question test.

In the coming wave of AI-washing, boards will be shown a great deal of false capital. The test is always the same: show me the evidence file, show me the autonomy stage, show me what survives the departure of the people and the vendor.

Capital has receipts.

You cannot manage what you refuse to write down, so write it down. Not with twenty-seven-part frameworks and maturity spider charts; with one ledger and five numbers. This is the entire measurement system, and a capable team can stand it up in a quarter.

The ledger

The ledger is an inventory: every validated capability in production, each with the fields below. If a capability is not in the ledger, it does not exist for governance purposes.

This single artifact does for the agentic firm what the fixed asset register did for the industrial one. It will feel just as boring, and prove just as indispensable.

The five numbers

On top of the ledger, five numbers — reported quarterly, to the executive committee, next to the financials.

One change to how money is presented

This may be the highest-leverage recommendation in the essay: split the AI budget into two lines.

Consumption is intelligence used up in operations: subscriptions, inference, copilots. Formation is investment that creates ledger entries: context engineering, evaluation development, skill validation, trace infrastructure. Accounting standards will not let you capitalize the second line yet. Management accounting answers to no one — run it internally anyway.

The moment a leadership team sees that ninety-five percent of its AI spend is consumption, the strategy conversation changes permanently, in one meeting. I have watched it happen.

Theory earns its keep in the specifics, so here is the arc as we build it in FMCG, with realistic numbers.

A European distributor processes 40,000 retailer deductions a year against roughly €300 million of promotional spend. Fully loaded cost per resolved claim: about €35, or €1.4 million a year — before counting the two to three percent of invalid deductions that go unchallenged because analysts have no time to fight them, which is several million more in silent margin leakage. The capability lives in eight experienced analysts, two of whom are near retirement.

human capital        high
token capital        zero
key-person risk      significant

Months 1–4 · Shadow

An agent investigates every claim in parallel with the analysts, executing nothing. Measured productivity: negative — and that is fine, because the real production is happening in the trace corpus. Twelve thousand claims where the agent’s reasoning is corrected by experts. A claim taxonomy hardening with every dispute. An entitlement calculator being tested against reality. A golden dataset of adjudicated cases curated from history.

The retiring analysts’ judgment is being written down for the first time in the company’s existence. Formation spend to date: about €400,000. Ledger entries: four capabilities, all in Shadow, evidence files growing weekly.

Months 5–10 · Propose, then the first promotion

The agent now drafts every resolution; humans decide. Accuracy on routine claim types crosses 95% against the golden dataset, and the evidence file supports moving duplicate detection and standard promotion claims — about 60% of volume — to Execute Under Threshold at a €5,000 limit, with sampled batch review.

Cost per trusted decision on covered volume falls to roughly €8. Recovered invalid deductions, now that every claim gets fought: €1.1 million in the first year — a number the CFO can see, produced by an asset the CFO cannot.

Year 2 · Where the argument cashes out

The company starts its second workflow, promotion settlement verification, and finds that the customer ontology, the agreement context layer, the tool gateway into the ERP, the evaluation harness, and the governance process all already exist. Build cost: 35% of the first workflow. Time to Execute Under Threshold: five months instead of ten.

Meanwhile the deduction trace corpus — now 45,000 labelled trajectories — quietly enables one more thing: a small, specialized model, fine-tuned on the firm’s own cases, that handles routine classification at a fortieth of the frontier model’s inference cost, with betteraccuracy on the firm’s specific claim types, because it has seen ten years of them and the frontier model never will. Frontier intelligence gets reserved for the genuinely hard cases. Cost per trusted decision falls again.

The acid test

Mid-year two, the company swaps its primary model provider after a pricing dispute. The migration takes nine days, because every capability re-certifies automatically against the evaluation harness. Nothing evaporates. The taxonomy, tools, traces, rubrics, and controls all transfer, because they were built as the firm’s assets rather than the vendor’s features.

That nine-day migration is the moment token capital stops being a concept from an essay and becomes a fact on the ground.

Strip everything above to actions and the list is short, which is how you know it is real.

1. 01Reframe the question, out loud, in your next AI review. Not “what did we spend and what did we save”, but “what do we own that we did not own last quarter, and what is its autonomy stage?” The room will not have an answer. That silence is the starting line.
2. 02Commission the ledger. One owner, one quarter. Every capability inventoried with its evidence file and stage, the five metrics baselined, the budget split into consumption and formation. Expect the first trace capture number to be embarrassing, and publish it internally anyway.
3. 03Run the sovereignty audit. Where do our traces live, what transfers if we switch providers, what have we been donating? Fix the architecture before scaling the workload, because leakage compounds exactly as fast as capital does.
4. 04Take one workflow up the ladder deliberately. High volume, real money, available experts. Shadow, Propose, Execute Under Threshold — evidence at every rung. Do not run twelve pilots. Run one workflow to genuine autonomy and let the second prove the reuse economics. Full rotations beat broad experimentation, every single time.
5. 05Put judgment codification into the culture.Only leadership can do this. Your best people’s decision patterns are the scarcest input to the whole system. Recognize, reward, and staff the work of writing them down. The experts who encode their judgment should be your most celebrated employees, not your most threatened — and whether that is true in your company is a leadership choice, not a technology one.

Step back to where we started: eleven minutes on a Tuesday afternoon, and the question of what the company owns on Wednesday morning.

Every era of business has one asset that separates the compounders from the crowd, and in every era that asset is invisible to the accounting of its time.

Intelligence itself will not be the asset of this era, precisely because everyone will have it — cheaply and equally, the way everyone has electricity. The asset of this era is the firm-specific machinery that turns that universal intelligence into trusted, autonomous, compounding execution: the context, the skills, the traces, the evaluations, the codified judgment, and the earned autonomy that binds them into something no competitor can buy.

The sentence for your board is this:

In the age of abundant intelligence, the scarce asset is trust made executable — and it can be built, measured, governed, and compounded, starting now, while it is still invisible to everyone else.

The models will keep getting better, and they will get better for your competitors at exactly the moment they get better for you. Token capital is the only part of the story where that is not true. It is the part you own.

Build it deliberately, defend it seriously, and measure it before the accountants force you to — because by then the compounding will already have picked its winners.