Terminal Value · Discovery Method · AI Capital Allocation
The Cognition Scarcity Audit: Fund the Analysis You Never Do
Strategic AI capital should not ask what work to speed up. It should ask what valuable analysis, vigilance and synthesis never happens because human labour made it uneconomic — then fund that.
The short version
- Most AI roadmaps inventory existing workflows and ask how to make the old horse go faster. That is operational, not strategic.
- A cognition scarcity audit hunts four absences: analysis done only on samples, events investigated selectively, planning limited to a few scenarios, and synthesis that waits for someone to ask.
- Run the findings through the portfolio test from the Terminal Value Doctrine: Horse Optimisation (tolerate) versus Car Construction (fund as strategy). The prize is a persistent intelligence layer — machine-scale attention with human judgment at consequential decisions.
Here is what I keep seeing when serious organisations open an “AI opportunity” workshop.
They put current processes on the wall. Inspection cycles. Project reporting. Maintenance planning. Risk registers. Regulatory packs. Then someone asks the question that feels responsible and ends up small:
How can AI make the old horse go faster?
That is not a stupid question. It is the natural question of a labour-scarce institution. The work already has owners, systems and KPIs. Acceleration is easy to budget and easy to defend. It will not be on anyone’s radar that the more valuable move is often the opposite: fund analysis the organisation has never performed at all — not because it lacked value, but because no human team could afford the breadth, depth or frequency.1
This article is a discovery method for that second question. It extends the Terminal Value Doctrine; it does not restate the asset-class map. It borrows the Cognition Ladder’s higher rungs — batch and overnight work that was previously infeasible — without re-teaching the ladder.23
The method has a name: the cognition scarcity audit.
The wrong map looks complete
Human organisations do not examine everything. They sample assets, escalate some incidents, model a manageable set of futures, and synthesise when a person remembers to convene the right meeting. That is not laziness. It is economics. Expert attention is expensive, coordination is expensive, and deep reading of soft records — notes, exceptions, near-misses, contractor deviations, “we learned this before” comments — does not fit inside a quarterly planning cycle.
The trap is treating that rationed map as the list of valuable work. If you only ask “where can AI help the work we already do?”, every answer stays inside Horse Optimisation or Horse Replacement: faster drafts, copilots for existing roles, automation of approval steps that already exist.4
The Terminal Value Doctrine’s sharper filter is different:
Does this project apply AI to work the organisation already performs, or does it make a previously uneconomic cognitive capability possible?
That is the portfolio test. The audit’s job is to fill the second column with candidates that are real enough to fund — not slogans about “enterprise intelligence,” but named forms of attention, investigation, simulation and contradiction-hunting that human labour could never run continuously.
Four scarcity signatures
Open with one workshop question:
What important thinking do we currently not do — not because it lacks value, but because it would require too many experts, too much coordination, or too much time?
Then search in four places. Each signature is a place where scarcity was mistaken for strategy.
Analysis performed only on a sample
Shape: The organisation inspects, reviews or deeply assesses a subset because full coverage is impossible at human cost. Sampling schedules, risk tiers and “materiality” thresholds become the quiet architecture of attention.
Recognisable example: A national infrastructure operator’s asset team runs thorough condition reviews on the assets already flagged as high-risk or due this cycle. Hundreds of other assets receive checklist-level attention. Nobody claims the unflagged assets are free of early-stage problems — only that expert hours cannot follow every transformer, corridor segment, project package and contractor record with the same intensity. The sample becomes the world.
Audit probe: “If cognition were nearly free overnight, which population should receive a regenerated evidence-based argument — not just a score — that we currently touch only by sampling?”
Events investigated only when they escalate
Shape: Major failures and selected incidents get disciplined investigation. Near misses, temporary repairs, procedure divergences, protection anomalies and “odd but closed” tickets do not — because comprehensive investigation is expensive and most will conclude nothing significant occurred.
Recognisable example: An operations centre closes dozens of minor anomalies a week. Formal investigation capacity is reserved for outages that cross a threshold or attract external attention. Five individually unremarkable events can share a common precursor for months before a human team is authorised to connect them. The organisation is not ignoring weak signals on principle; it is rationing detective work.
Audit probe: “Which classes of weak signal would we investigate every time if the marginal investigation cost approached zero — and what systemic pattern might only appear when all of them are examined together?”
Planning limited to a handful of scenarios
Shape: Strategy and network planning can construct futures, but only a few can be staffed, interpreted and challenged. Combinatorial futures — demand shifts plus supply delays plus weather extremes plus community opposition plus cost escalation — exceed the practical scenario library.
Recognisable example: A planning team produces a flagship scenario pack for the board: base case, high demand, delayed build, and a stress case. Each pack is hard-won. What they cannot maintain is a living map of thousands of plausible futures, with decisions scored across them and leading indicators for which future is becoming real. External CEO research has already noted how much planning energy compresses into short horizons — a structural pressure that makes “one more polished base case” feel safer than a permanent counterfactual engine.5
Audit probe: “Which decisions would change if we could keep thousands of scored futures warm — and which signals would tell us which future is arriving?”
Synthesis that depends on someone remembering to ask
Shape: Cross-silo insight happens when a senior person convenes the right rooms, or when a crisis forces the documents onto one table. Between those moments, the organisation’s soft exhaust — field notes, standards exceptions, risk-register language, project assumptions, regulatory commitments — is not continuously joined.
Recognisable example: Engineering standards say one thing; repeated field practice does another. The risk register marks a control as effective; maintenance notes imply otherwise. A project assumption outlives the constraint that created it. Two teams quietly use inconsistent parameters. A lesson appears in three close-out reports and never enters the standard. Humans encounter these contradictions accidentally. Nobody’s job is to hunt them every night across the corpus.67
Audit probe: “What contradictions would we want exposed continuously — policy vs practice, model vs field, commitment vs delivery plan — without waiting for a human to nominate the question?”
What the four signatures share
Each is a place where the organisation’s current work map is a budget of attention, not a complete list of valuable cognition. The audit treats that budget as temporary. Cheap, parallel, multi-step analysis — the kind agent systems are built to run over many turns on open-ended problems — changes what can sit inside the permanent operating model.8
Worked sketch: asset-heavy operator (TNSP-shaped)
Apply the audit to a generalized transmission-network service provider — any asset-heavy operator with a large physical estate, contractor ecosystems, regulatory commitments and project pressure. The names of assets change; the scarcity signatures do not.
Opening workshop question
What important thinking does this operator currently not do — not because it lacks value, but because it would require too many engineers, too much coordination, or too much time?
Signature walk-through
- Sampled analysis. Condition and risk attention concentrates on known-risk tiers and scheduled inspections. Car-construction candidate: a regularly regenerated risk argument for every major asset class member — why it is believed safe, what evidence supports that belief, what evidence is missing, what changed, what would reverse the conclusion, which similar assets behaved differently, which inspection buys the most information.
- Selective investigation. Formal investigation capacity sits behind severity thresholds. Car-construction candidate: lightweight, disciplined investigation of every protection anomaly, unusual maintenance observation, contractor deviation, near miss, unexpected outage precursor and repeated temporary repair — most concluding “nothing,” occasionally connecting five unremarkable events into a systemic problem.
- Few scenarios. Planning packs hold a handful of board-ready futures. Car-construction candidate: continuous generation and scoring of large futures sets (electrification pace, renewable delays, extreme weather, supply-chain disruption, corridor opposition, simultaneous retirements, cost escalation and combinations) with a living map of decisions that perform across them and signals for which future is becoming real.
- Ask-dependent synthesis. Cross-reading of standards, field practice, risk registers, project assumptions and regulatory commitments happens in incidents and major reviews. Car-construction candidate: continuous contradiction exposure and unnominated-risk search across the soft corpus — the organisational source code, not only the structured dashboards.
What “persistent network cognition” means in this sketch
Translate the four absences into one operating picture. Not a dashboard that restates known condition scores — a system that continuously reads inspection reports, maintenance notes, outage records, weather and bushfire context, engineering standards, project changes, contractor reports, SCADA anomalies, regulatory commitments and previous failure investigations, and keeps asking questions no single team can afford every night:
- Which weak signals are appearing across otherwise unrelated assets?
- Which current risks resemble the early stages of historical failures?
- Which assumptions in the asset strategy are becoming stale?
- Where are field observations contradicting the formal asset model?
- Which apparently isolated problems may share a common cause?
That is not faster asset management. It is a new layer of persistent network cognition. Today, engineering attention is rationed by known risk, inspection schedules and escalation thresholds. With cheaper cognition, every transformer, tower, line, easement, project and significant component can receive a regularly regenerated risk brief: why the asset is believed safe, what evidence supports that belief, what is missing, what changed, what would reverse the conclusion, which similar assets behaved differently, and which inspection would buy the most information. Humans still make consequential decisions. The organisation no longer waits for a person to decide an asset deserves investigation before assembling the case.
Audit output sketch (not a build plan)
A one-page scarcity map for the operator might list:
- Continuous weak-signal synthesis across inspection notes, SCADA anomalies, weather and bushfire context, contractor reports and prior failure investigations.
- Evidence-based risk argument per asset — regenerated on a batch cadence, human decision retained for consequential calls.
- Every-anomaly investigation queue with cheap close-outs and rare high-value joins across protection anomalies, temporary repairs, near misses and procedure divergences.
- Thousands of counterfactual futures kept warm for planning and investment decisions — electrification pace, renewable delays, extreme weather, supply-chain disruption, corridor opposition, simultaneous retirements, cost escalation and combinations.
- Contradiction engine between as-designed and as-operated artefacts (standards vs field practice, risk register vs maintenance notes, commitments vs delivery plans).
- Unnominated-risk search — risks nobody put on the register yet; reconstruction of why historical network decisions were made and whether their constraints still exist.
That list is the beginning of a strategic portfolio. It is not an implementation architecture. How any of these capabilities is governed, reviewed, disclosed or staffed is owned by adjacent work — elastic assurance, institutional linting, failure-shape sensing, workforce reserve, barbell governance, intent compilation — not by this discovery article.91011121314
The portfolio test: horse optimisation vs car construction
Once candidates exist, sort them. The Terminal Value Doctrine already named the classes. This audit only forces honesty about which class each idea belongs to.
| Project idea | Classification | Capital treatment |
|---|---|---|
| Draft engineering and regulatory reports faster | Horse Optimisation | Tolerate as operations spend |
| Summarise standards and inspection packs for staff | Horse Optimisation | Tolerate; measure time saved, not strategy |
| Copilot for maintenance planners | Horse Optimisation | Operational improvement, not the strategic portfolio |
| Automate an existing approval step | Horse Optimisation or Horse Replacement | Ops / process; watch for automation trap |
| Continuously reassess every major asset from all available evidence | Car Construction | Strategic: new cognitive capability |
| Investigate every operational anomaly and near miss | Car Construction | Strategic: coverage humans cannot staff |
| Maintain thousands of scored network futures | Car Construction | Strategic: permanent counterfactual planning |
| Expose contradictions between policy, models and field reality | Car Construction | Strategic: continuous institutional synthesis |
| Search for risks nobody has nominated | Car Discovery → Construction | Strategic exploration becoming capability |
The optimisation projects are not useless. The doctrine’s language is precise: tolerate, don’t celebrate. Fund them as operational improvements. Do not let them consume the strategic AI portfolio because they are easier to inventory.
If your “AI strategy” is only the left side of the table, you do not have a strategy problem. You have a discovery problem. You never ran the audit that finds the right side.
The proposition the audit is aiming at
BI for Soft Data and related substrate work matter here as plumbing, not as the headline. The headline is not “better search over documents.” The headline is a persistent intelligence layer over the asset base and its institutional memory — capable of examining every asset, event, assumption and plausible future at a depth and frequency no human organisation could afford, while preserving human engineering judgement at consequential decision points.
Machine-scale attention everywhere. Human judgment where consequences concentrate.
That is a different operating model from “engineers work faster.” It is the AI-native answer to cognition scarcity: stop rationing expert attention as if the sample were the world; apply continuous attention to populations, weak signals and futures; keep people as decision-makers and designers of the questions, not as the only affordable sensors.
Batch and overnight cadences are the natural home of this work — Cognition Ladder rungs where 10–100× more analysis becomes possible because the system is not competing with a human in a real-time conversation.2 Nightly decision-build hygiene is one operating pattern for making that batch layer real; it is a how, not a substitute for choosing the right capability to fund.15
Mapping example capabilities onto batch / overnight rungs
| Capability from the audit | Why it is not real-time substitution | Ladder home (shape) |
|---|---|---|
| Regenerated risk argument per asset | Reads large evidence sets; output is a reviewable brief | Batch augmentation → overnight depth |
| Every-anomaly investigation | High volume, mostly negative results; joins emerge over time | Batch / queue transcendence of investigation scarcity |
| Thousands of futures scored | Combinatorial load humans cannot staff | Overnight / continuous transcendence |
| Contradiction and unnominated-risk search | Corpus-wide synthesis without a human prompt each time | Standing overnight inquiry |
How to run the audit in practice
Keep it short enough to finish and sharp enough to hurt.
- Lock the question. Write the opening line on the wall: valuable thinking we do not do because of cost, coordination or time — not “AI use cases for existing processes.”
- Collect absences, not systems. For each of the four signatures, demand one concrete population, event class, scenario gap or synthesis that today only happens when someone asks.
- Force a recognisable example per signature. If the room only produces abstractions (“better insight”), they have not finished. Names of asset classes, ticket types, planning packs and document genres are required.
- Translate each absence into a capability sentence. “Continuously X every Y using Z evidence, producing a reviewable artefact for human decision.”
- Run the portfolio test. Label each sentence Horse Optimisation, Horse Replacement, Car Discovery or Car Construction. Move strategic capital toward the car column.
- Separate discovery from design. Do not let the workshop collapse into governance architecture, model selection or vendor demos. Those are later articles and later weeks. The deliverable of this room is a scarcity map and a funded shortlist.
Boundary: This method does not tell you how to assure elastic disclosure, lint institutional artefacts, sense failure shapes in soft behaviour, measure human reserve under green dashboards, barbell project governance, or compile intent into deterministic fusion. Those are real problems — and they are the wrong problems to solve first if you still only fund faster horses.
What changes if this idea wins
For individuals: the AI conversation stops being “help me finish the pack I already write” and becomes “what pack should exist that we never staffed?”
For teams: workshops stop inventing copilots for every role and start inventing coverage for every population, weak signal and future the role could never reach.
For capital allocation: the strategic AI portfolio is judged by whether it builds previously uneconomic cognition — the Terminal Value test applied at discovery time, not only at retrospective reclassification.
The Cognition Dimension Ladder’s related warning applies in the background: deepening a saturating rung (slightly better copilots on the same workflows) is not the same as entering a new dimension of work.3 The scarcity audit is how you find the candidates that live in that new dimension.
Close
They will look at existing workflows. That is fine for the operations budget.
Strategic capital should point at the uneconomic thing — the analysis, vigilance and synthesis humans could not do at sufficient breadth, depth or frequency. The cognition scarcity audit is how you find it. The horse-versus-car portfolio test is how you stop lying to yourself about what you found. The persistent intelligence layer is what you are trying to fund: machine-scale attention across assets, events, assumptions and futures, with human judgment retained where it counts.
Fund the analysis you never do. That is the discovery method. Everything else is making the old horse go faster.
Run the audit before the use-case inventory
If you want a facilitated scarcity map for one asset class, one project portfolio or one regulated control family — sorted into horse versus car — start a conversation.
References
- Scott Farrell, LeverageAI. "The Terminal Value Doctrine — Stop Optimising the Horse." — Horse Optimisation vs Car Construction; fund what governs terminal value under cheap cognition. leverageai.com.au/wp-content/media/articles/61-terminal-value-doctrine.html
- Scott Farrell, LeverageAI. "Maximising AI Cognition and AI Value Creation." — Cognition Ladder: real-time substitution vs batch augmentation vs overnight transcendence. leverageai.com.au/wp-content/media/articles/27-maximising-ai-cognition.html
- Scott Farrell, LeverageAI. "The Cognition Dimension Ladder — Why Your AI Strategy Is One Rung Too Low." — Value climbs by entering a new dimension of work. leverageai.com.au/wp-content/media/articles/62-cognition-dimension-ladder.html
- Scott Farrell, LeverageAI. "Stop Automating, Start Replacing." — Automation trap; roadmaps that mirror process maps. leverageai.com.au/wp-content/media/articles/24-stop-automating-start-replacing.html
- Oliver Wyman Forum. "The CEO Agenda 2026." — Large share of CEO planning effort on sub-one-year horizons; long-horizon planners make different strategic choices. oliverwymanforum.com/ceo-agenda/how-ceos-navigate-geopolitics-trade-technology-people.html
- Scott Farrell, LeverageAI. "Your Organization Has Source Code (And You Can Finally Read It)." — Soft exhaust as organisational source code; as-designed vs as-operated. leverageai.com.au/wp-content/media/articles/86-your-organization-has-source-code.html
- Scott Farrell, LeverageAI. "BI Tells You Where, the Wiki Tells You Why." — Structured systems locate; soft layers explain. leverageai.com.au/wp-content/media/articles/106-bi-where-wiki-why.html
- Anthropic. "Building Effective Agents." — Agents operate for many turns on open-ended problems with unpredictable step counts. anthropic.com/research/building-effective-agents
- Scott Farrell, LeverageAI. "Elastic Assurance: Compute Broadly, Disclose Narrowly." — Assurance cadence; compute broadly, disclose narrowly. leverageai.com.au/wp-content/media/articles/136-elastic-assurance.html
- Scott Farrell, LeverageAI. "The Institutional Linter: Static Analysis for Your Organisation." — Codified-org static analysis. leverageai.com.au/wp-content/media/articles/137-institutional-linter.html
- Scott Farrell, LeverageAI. "The Institutional Failure Radar: Failure Changes Shape Before It Changes the Numbers." — Behavioural sensing; systems not people. leverageai.com.au/wp-content/media/articles/138-institutional-failure-radar.html
- Scott Farrell, LeverageAI. "Green by Heroics: The Safety Margin Your Dashboard Can't See." — Green sustained by hidden human reserve. leverageai.com.au/wp-content/media/articles/139-green-by-heroics.html
- Scott Farrell, LeverageAI. "The Governance Barbell: Run Projects Like Pull Requests." — Heavy verification at the ends; thin middle. leverageai.com.au/wp-content/media/articles/140-governance-barbell.html
- Scott Farrell, LeverageAI. "The Intent Compiler: Deterministic Fusion of Fuzzy Priors." — Intent compiled into runnable form. leverageai.com.au/wp-content/media/articles/141-intent-compiler.html
- Scott Farrell, LeverageAI. "Nightly AI Decision Builds." — Batch/overnight decision engines with regression and artefact packs. leverageai.com.au/wp-content/media/articles/45-nightly-ai-decision-builds.html