Born-Structured Exhaust: Mine for Friction, Not Guilt

AI-assisted work manufactures the organisation's causal layer at birth. Mine it for system friction and invisible work — never for individual guilt.

Scott Farrell · LeverageAI · July 2026 · Full ebook series companion to articles 189–190

Here is the scene every senior manager already knows. A board paper is due. The dashboards are green. Tickets closed. Documents published. Milestones marked complete. Then someone asks the only question that matters: what actually happened, and where are we stuck?

What happens next is archaeology. A senior person is assigned to “pull the story together.” They sample email threads, skim meeting minutes, re-open SharePoint folders, ask around. Days later they produce a narrative that is sampled, slow, unrepeatable — and gone by next quarter. The warehouse held the outcomes. The why lived in soft exhaust nobody was structured to read.

That archaeology is exactly the premise of BI for Soft Data: the interview already happened, every day, for ten years, and you dig it up after the fact. This article inverts that premise. When people work with AI, the interview is happening now — prompted, in real time — and the conversation transcript is the record as it is created. That inversion changes acquisition cost, freshness, and the structure of the causal layer all at once. It strengthens the parent framework; it does not compete with it.

The interview is happening now — prompted, in real time — and the transcript is the record.

The thesis, held hard

AI-assisted work manufactures the organisation's causal layer at birth — because getting useful help requires the worker to state intent, alternatives and blockers explicitly — which turns friction sensing from an inference problem into a reading problem, and raises the governance stakes in exact proportion.

Two load-bearing assumptions sit outside this piece. That the deliberation is worth treating as source at all is the argument of The Deliberation Is Source. That sessions must be joined to the artefacts they touch is the argument of Provenance-Coupled Work. Both are assumed here. This book owns what a manager can read from a portfolio of that exhaust — and the contract that keeps the reading legitimate.

Why the exhaust is born structured

Hard data got forty years of BI because it was born structured: the schema existed at write time. Soft data — the emails, minutes, threads and documents where the why lives — stayed dark because activation required comprehension, and comprehension had no unit price. IDC’s measurement that 90% of data organisations generated in 2022 was unstructured is the estate-scale backdrop for that darkness.¹ Traditional systems record outcomes. The soft layer holds the decisions, objections, trade-offs and workarounds that produced them.

AI conversation changes the write path. To get a useful answer, a person almost always has to say, in substance:

• Here is what I am trying to achieve.
• That approach failed because…
• Compare these three alternatives.
• Use this assumption rather than that one.
• This is still unresolved.
• I need an answer before Thursday’s steering committee.

Historically an email thread contained that reasoning by accident. Now it is close to a declared field. The model interaction rewards articulation. The finished Word document, spreadsheet or board pack still optimises for its audience — and still suppresses the messy deliberation. The conversation keeps the intent, the rejected paths, the open questions and the friction.

That is the same distinction developed for software: the code is the what; the transcript is the why. For knowledge work the line is: the document is the what; the deliberation is the why. The join between them is article 190’s job. The management read of a body of joined exhaust is this article’s job.

Be precise about what “born structured” does and does not mean. It does not mean every prompt is a clean database row. Sessions are still noisy. People paste. People ramble. People perform for the tool. What it does mean is that the interaction pulls semantic fields that conventional corporate writing actively suppresses. A board paper is optimised to look settled. A status email is optimised to look green. A model interaction is optimised to get help — and help requires the worker to externalise purpose, contrast, correction and uncertainty. That is a write-path change, not a better search over the same old residue.

Compare the final artefact to the session on purpose. The finished document tells you what survived publication. The deliberation tells you what the person was trying to achieve; what sources they relied on; what they changed and why; what they rejected; what remains open; what they intended but never completed; and where they got stuck. Those are the fields a manager needs to unstick a portfolio. They are also the fields a surveillance product wants to misuse. Design for the first use; refuse the second.

Company-owned surfaces, not retail exhaust

Today many knowledge workers use nothing — blocked on privacy grounds. Others use retail tools, and the conversation is trapped where the organisation cannot govern it. Others use custom agents whose logs are technically “owned” but practically inaccessible: wrong format, wrong retention, wrong product owner. None of those states produces a management view. They produce either silence or dark exhaust.

The architecture this piece assumes is simple enough to state and hard enough to implement: approved AI surfaces operate through a company-owned gateway; every surface emits the same class of structured work events; raw sessions are retained as bronze; a distillation pass produces session briefs focused on intent, decisions, rationale, rejected alternatives, blockers, open questions, commitments, artefacts changed and reusable learning. Most briefs stay project history. A minority feed promotion nominations. Managers receive compiled views, not prompt dumps.

That is deliberately not “archive all chats in a lake and hope retrieval fixes meaning.” Retrieval without structure is how you rebuild the board-story archaeology with better search. The product is a compile — continuous, role-shaped, and constrained by the legitimacy contract.

The study that should be run (and the number you must not invent)

The natural claim is that AI-conversation exhaust is denser in explicit intent, alternatives and blockers than matched email and meeting minutes on the same work. That claim is plausible. It has not been measured as a published comparison in this programme. Fabricating a multiplier — “3× more explicit intent” — would be the most damaging failure mode for this piece.

So here is the study, specified so a reader could commission it next quarter:

• Sample: matched work episodes where both an AI-assisted session and conventional email/minutes exist for the same task.
• Unit: explicit statements of intent, alternatives considered or rejected, and blockers or dependencies — coded per 1,000 words and per episode.
• Method: dual independent coding against fixed definitions; inter-rater agreement reported; copy-paste and performative prompting treated as failure modes.
• Expected shape (hypothesis, not result): AI sessions should show higher rates of explicit intent/alternative/blocker language because the interaction rewards articulation; conventional exhaust should show more implicit, context-dependent residue.

Until that study ships, speak the shape. Do not invent the rate. The rest of this article proceeds on qualitative economics: AI interaction rewards articulation of the fields managers need. That is enough to design the product. It is not enough to publish a fake precision statistic.

Also specify what the study must not become. It must not code “productivity.” It must not rank individuals. It must not treat session length as a proxy for contribution. Its only job is semantic-field density comparison on matched work — intent, alternatives, blockers — so the organisation can decide whether the born-structured claim is strong enough on its estate to justify the capture investment.

The four-part weekly view

A senior manager should not receive a dump of everyone’s prompts. The useful product is a compiled management view in four parts.

1. What moved

Workstreams advanced. Important artefacts created or materially changed. Decisions made and their rationale. Assumptions overturned. Deliverables approaching review or approval.

2. Where the organisation is sticking

Recurring blockers across teams. Unresolved dependencies. Repeated demand for the same scarce specialist. The same problem being solved independently in several places. Decisions repeatedly reopened. Work waiting on authority rather than execution.

3. Where management can help

Additional resources. Priority or scope clarity. A cross-team decision. Access to a missing system or data source. A reusable platform capability. De-scoping or deadline renegotiation. Escalation to the appropriate owner.

4. What the organisation learned

New reusable methods. Rejected approaches worth remembering. Emerging client or market patterns. Undocumented constraints discovered. Canon entries proposed. Related prior work resurrected.

That is not traditional activity reporting. It is a continuously compiled account of the organisation’s current cognitive state.

Every row in that view is system-addressable. “Same specialist pulled into every issue” is a capacity and knowledge-distribution problem. “Work waiting on authority” is a decision-rights problem. “Same problem solved independently in three places” is a reuse and communication problem. None of those rows is a person score. If your implementation adds a leaderboard of who prompted most, you have left the doctrine.

Walk the four parts as a weekly ritual, not a slide template. On Monday a head of function should be able to answer: what actually advanced; where the same friction appeared more than once; which interventions are management’s job rather than the team’s; and which learnings must not die in a project folder. If the compile cannot answer those four without reopening raw prompts, the product is incomplete. If it answers them by naming who was “slow,” the product is illegitimate.

Part one — what moved — is deliberately not a timesheet. Artefacts changed, decisions made with rationale, assumptions overturned: those are cognitive events. Part two — sticking — is where organisational behavioural telemetry and born-structured exhaust meet: recurring blockers and reopened decisions are pathway signals, not character assessments. Part three — help — is the legitimacy payload: every sticky signal should propose a support action a manager can take. Part four — learned — is the bridge to canon: without it you only fight fires; with it you stop re-lighting the same ones.

A compiled week (composite, labelled)

What follows is a composite drawn from several knowledge-work patterns of the kind that appear when AI sessions are retained and distilled — not a named employer and not a single person’s surveillance log. It is the artefact a manager should actually receive.

Read every row. The product is not a dashboard tile. It is a support response attached to a system problem, with enough semantic detail that a manager can act without reopening every raw prompt.

Notice what is missing from the table on purpose: names ranked by stuckness; a productivity index; “AI adoption score” per person; time-on-tool heatmaps used as performance evidence. Those omissions are the design. A works council reading this artefact should be able to see the unit of analysis without a lawyer translating. The legal regimes still matter and still sit outside this book; the product doctrine must already be readable as system-level help.

Also notice the grain. Signals are counted across workstreams and teams. “Four workstreams asked the same policy question” is organisational. “Person X asked the model four times” is the field you refuse to emit into the management product. Raw bronze may still need identity for security and audit under local law — that is a different plane. The management compile must not promote identity into a score.

Invisible work becomes creditable

Much valuable knowledge work produces little visible output. A worker might spend two days establishing that a proposal cannot work; reconciling conflicting interpretations; locating an undocumented dependency; discovering that three datasets cannot be safely joined; persuading several groups not to take the obvious path; narrowing a problem until the eventual document becomes simple.

The final artefact may be one paragraph: “Option B is recommended.” The AI conversation may contain the actual contribution — the alternatives walked, the constraints found, the people unstuck.

This is the recognition argument, and it is distinct from the capacity argument. Green by Heroics already owns the obligation-to-capacity question and the hard boundary against individual scoring: where has obligation increased without funded capacity, and how do you replenish the team rather than heroics? What that body of work does not own is credit — making invisible cognitive work visible so it can be recognised, not merely so capacity can be restored. That gap is this piece’s distinct positive claim.

Organisations already know, vaguely, that some of their best people “make hard things look easy.” They rarely have a receipt. Promotion packets fill with visible deliverables. Investigations that prevented a bad path leave almost nothing to point at. AI deliberation, if retained and joined, is the first scalable receipt for that class of work — provided the institution uses it for recognition rather than for a gotcha archive of who struggled aloud with the model.

Recognition still must stay non-scored. Credit a contribution in a human process: a manager cites the investigation in a narrative; a team records a “saved us from Option A” entry; a promotion case includes a deliberation-backed example with consent where required. Do not emit a continuous “contribution score” derived from session volume. The fence is the same fence: system learning and human credit, not algorithmic ranking of persons.

When signals stop being weak

Institutional Failure Radar fuses weak behavioural signals — discussion volume, compressed dissent, abnormal effort, what disappears between thread and pack — into nominations about pathways, not people. That is the right moral geometry, and it remains the right geometry here.

Born-structured exhaust changes the detection problem. Intent, alternatives and blockers arrive closer to declared fields. Detection moves from inference over weak residue toward reading explicit structure. That is a stronger sensing layer. It is also a more dangerous one if misused.

When the signal stops being weak, the governance stakes rise in exact proportion.

Say that explicitly. Stronger exhaust does not license looser ethics. It demands tighter unit-of-analysis rules, tighter access, and a default response path that cannot silently become individual discipline.

Conventional workplace monitoring already carries a wellbeing cost. The American Psychological Association’s Work in America work found that employees in monitored environments are more likely to feel burnt out, emotionally exhausted and less motivated than those who are not monitored.² Gartner estimated that technology-based worker monitoring in large businesses rose from about 30% before the pandemic to roughly 70% afterward — a shift reported in secondary synthesis of workplace surveillance studies.³ Experimental work published in Harvard Business Review found employees more likely to break rules when they knew they were monitored than when they were not.⁴ Those findings are not a reason to abandon organisational sensing. They are a reason to refuse the individual-scoring path that produces the harm.

System-level friction signals

The soft exhaust can reveal where:

• routine tasks require abnormal cognitive effort;
• the same specialists are pulled into every issue;
• discussion volume grows around supposedly settled work;
• review depth is compressing;
• work moves outside ordinary hours (team aggregate);
• preventative work is displaced by urgent work;
• obligations have expanded without additional capacity.

Each signal is a capacity hypothesis or a control-environment hypothesis. None is a loyalty score. The obligation-to-capacity framing from Green by Heroics is the systems question underneath: where has the ratio changed without a corresponding change in resources? The intended output is a team-level support response.

Fence every mechanism: if a product manager proposes “per-person friction score,” kill the feature. If a vendor demo ranks individuals by “stuckness,” walk away. If HR asks for individual AI-usage percentiles for performance season, the answer is no — and the architecture must make that answer easy by never emitting the field.

The canon promotion path

Most sessions stay project history. Some patterns deserve promotion:

• a recurring client objection;
• a repeated operational constraint;
• a common spreadsheet mistake;
• a policy ambiguity repeatedly explained;
• a dependency that surprises every new team;
• an approach tried and rejected several times;
• a workaround quietly becoming standard practice.

The system can nominate. A human disposes. That is the same promotion discipline Institutional Memory describes for the staging buffer and the governed record: raw sessions are bronze; durable claims earn a reviewed path into canon. Without that path, you have only a larger pile of chats. With it, the organisation stops re-paying the same learning cost every quarter.

Second-order learning already taught software teams to ask three questions after work: where did we stick; what surprised us; what should we never do again. Born-structured exhaust makes those questions answerable from the work itself rather than from a heroic retro that never gets scheduled. The nomination pass is the automated version of that ritual at estate scale — still human-gated before anything becomes “how we work here.”

Promotion is also where privacy and legitimacy meet. Not every session should enter canon. Not every friction should become a permanent institutional claim about a team. The weekly view can be ephemeral. Canon is durable. Different retention, different access, different review bar. If you collapse those layers, you either lose learning or over-retain raw first-person material under a management gaze it never earned.

Why now

McKinsey’s 2025 State of AI survey finds that nearly two-thirds of organisations have not yet begun scaling AI across the enterprise, while roughly one-third report that scaling has begun.⁵ Adoption is wide enough that privacy and gateway decisions are being made; scale is incomplete enough that schemas are still soft. Nielsen Norman Group’s synthesis of three studies found generative AI tools increased business users’ throughput by 66% on realistic tasks — one reason knowledge-work AI will keep expanding where it is allowed.⁶ Gartner’s 2024 digital-worker survey found only 23% of digital workers completely satisfied with work applications — pressure for better corporate surfaces continues.⁷

Meanwhile organisations are blocking AI on privacy grounds, which means the exhaust is being lost rather than governed; or workers use retail tools, and the conversation lands where the company cannot read it for help; or custom agents write to log files nobody treats as a management product. The argument that the by-product is more valuable than the tool needs to arrive before the policy hardens around either total blockage or individual surveillance.

The window is also a product window. Gateway vendors and internal platform teams are choosing which fields to keep. Token counts and latency look like “observability.” Resource identifiers, intent fields and blocker tags look like “noise” until someone names the management product those fields enable. Article 190 makes that case for join keys. This article makes the parallel case for semantic exhaust aimed at friction and learning. Drop the fields now and you will not reconstruct them later from a pile of final PDFs.

The legitimacy contract

The slogan is load-bearing, not decorative:

Mine for friction, not guilt.

The management question is never “who is slow.” It is “where are capable people repeatedly encountering an operating-system problem?” The default response to a signal is: provide help; clarify authority; add resources; redesign the process; remove repeated friction; reduce obligation; promote reusable knowledge. Monitoring is legitimate precisely because the observed party receives a direct benefit from being seen.

Name the legal boundary and stop. Detailed HR regimes, works-council procedures, consent mechanics and jurisdictional regimes are real and necessary. They are not this book’s subject. What this book specifies is the product doctrine a works council could actually read as intent: system-level only; help default; no individual score field; retention and access as organisational design choices with human review on promotion to canon.

A minimal policy sketch — not legal advice, not a jurisdiction — that a design team can put on one page:

• Purpose: compile system-level friction, help opportunities, invisible-work recognition and learning promotion — not performance management.
• Unit: team, workstream, control or pathway. No individual ranking product.
• Access: managers see compiles at their span of control; raw sessions restricted; security/audit planes separate.
• Retention: briefs and weekly views shorter-lived by default; canon longer-lived after human promotion; raw bronze retention set by risk class, not by “keep everything forever for HR.”
• Response obligation: a surfaced signal without a support response is a governance defect.
• Prohibited outputs: individual AI-usage percentiles; personal “stuckness” scores; automated discipline triggers from session content.

If your vendor cannot support those constraints, you do not have a fair implementation. You have a monitoring product wearing insight language.

Owning the optimism

One honesty requirement remains. The pro-worker framing in this piece was not a neutral finding that emerged at the end of analysis. It was a deliberate starting constraint given by the author: show the fair implementation and the upside; do not spend the page nitpicking problems into paralysis. Presenting that optimism as if it were an emergent, value-free conclusion of the analysis would be dishonest.

We wrote this piece to the brief: show the fair implementation and the upside. That optimism was a constraint we accepted at the start, not a finding we discovered at the end.

The constraints are still real — privacy, misuse, power imbalance. The design answer is to build the fair path first-class: born-structured exhaust, four-part weekly view, invisible-work credit, canon promotion, and a hard ban on individual performance measurement. That is the product. The pessimist’s path — score people with AI chat logs — is the failure mode we refuse to ship. Owning the optimism does not mean pretending the failure mode is impossible. It means designing so the failure mode is not the default product.

What you can specify on Monday

1. Company-owned AI surfaces that emit the same structured work events (intent, alternatives, blockers, artefacts touched — join details per article 190).
2. Raw sessions retained as bronze; automatic distillation into session briefs; no manager dump of raw prompts by default.
3. A weekly compile into the four-part view, aggregated to team/workstream grain.
4. Friction signals with attached support responses, not person ranks.
5. A recognition path for investigations that leave thin artefacts.
6. A human-gated canon promotion queue for recurring patterns.
7. An explicit product rule: no individual performance field derived from AI exhaust.

If you want a two-week pilot rather than a platform programme, pick one workstream with approved AI use, retain sessions with consent and purpose notice appropriate to your jurisdiction, distill briefs by hand if you must, and produce a single four-part weekly view for that workstream only. Attach support responses. Show the team the view before you show their manager’s manager. Ask whether being seen produced help. If it did not, stop — the legitimacy contract failed early, which is cheaper than failing at estate scale. If it did, you have a specimen, not a slide deck, and you can argue for the capture fields that make the specimen automatic.

The full enterprise picture — Cognitive Git as a metaphor — can wait until the weekly view works. Conversations as working sessions; knowledge-work commits as semantic waypoints; engagement worlds as project branches; institutional memory as the governed main branch; promotion as reviewed merge. That architecture is real and larger than this article. It is not a substitute for the manager’s artefact or the help-not-judgment fence. Ship the artefact and the fence first.

The deeper corporate opportunity is not capturing more documents. It is capturing the meaning-making process that produced them — and reading it for friction, recognition and learning, under a contract that makes being seen a form of being helped.

One more counterfactual, because it is the cost of inaction in human form. Without this compile, next quarter’s board narrative is again a senior person sampling threads. The scarce specialist is again re-taxed for free. Option A is rejected again by a new team that never saw last quarter’s dead end. The person who spent two days proving a join was unsafe is again invisible because the artefact is four sentences. And somewhere a vendor is demoing an individual AI-productivity score as if that were the innovation. The doctrine in this article is the alternative: born-structured exhaust, read for friction not guilt, with governance stakes named in proportion to how strong the signal has become.