THE ADD-BACK · episode 09
SightingA fund has run AI operational work inside nine portfolio companies. Ask which of the nine worked and why, in comparable form, and nobody can answer — not because the work was bad, but because no two engagements produced the same kind of object.
The consensus readBandwidth. More operating partners, more coverage.
The mechanismBandwidth isn't the constraint. Nine engagements produced nine one-offs, so engagement ten starts at the same cost as engagement one. The fund has paid for nine instances of learning and banked none of it, because learning at fund level requires comparability and nothing enforced it.
The exposureThe thing you're building is a services team, not an asset.
The testAsk two operating partners for the decision output of their last engagement. If the two documents aren't the same shape, your learning rate is zero.
Portfolio learning rate
Whether engagement N+1 is cheaper, faster or better because engagements one through N happened.
The test. A comparison, not a survey. Put the decision outputs of two engagements side by side. If they share a decision type, an artifact type and a failure vocabulary, the rate is positive. If they share a font, it's zero — and zero is the default, because nothing in a normal operating model produces comparability.
The nine-engagement fund is a composite. Ostrom and the agent-market coordination study are real and cited with their samples; the per-engagement value is our estimate.
| Line | Moves | Running |
|---|---|---|
| Claimed — A value-creation function whose capability compounds across the portfolio | 9 engagements, 0 comparable pairs | |
| 1. What compounding would require changes the story Three things have to be common across engagements before any comparison is possible. The same decision type: every engagement terminates in the same class of answer — fund, structurally rewrite, or kill. Not a score, not a roadmap. If engagement A produced a maturity rating and B produced a go/no-go, there is nothing to compare. The same artifact type: a fixed shape — the corrected boundary, alternatives considered, failure modes, the decision, and what would have reversed it. A fixed shape is what makes engagement six's output legible to someone who only worked on engagement two. The same failure vocabulary: both engagements call the same phenomenon by the same name. This is the one that's invisible until you try to aggregate — if one team wrote 'adoption issues' and another wrote 'exception load in adjacent functions', they may have found the identical thing and no query will ever connect them. Most operating models supply none of these, because none is required to do good work on any single portfolio company. | −$0 | 9 engagements, 0 comparable pairs |
| 2. Why the internal team can't fix this by trying harder changes the story Operating partners are typically ex-operators with deep judgment and their own methods. That is why they are valuable, and it is the source of the problem: method heterogeneity is a hiring outcome, not a discipline failure. You recruited nine people for independent judgment and then asked why their outputs aren't standardised. Bandwidth is a budget question; consistency is a design question nobody owns. Each operating partner is measured on their portfolio company's performance, nobody is measured on whether their output is usable by the next partner, so nobody produces it. A straightforward incentive gap that does not resolve by exhortation. | −$0 | 9 engagements, 0 comparable pairs |
| 3. Where I was wrong — I proposed a central playbook where I was wrong Given method heterogeneity, my recommendation was the obvious one: write a central methodology, mandate it across the operating team, enforce it through the value-creation function. A playbook. Every firm that has confronted this has reached for the same lever. Ostrom's field work on institutions governing shared resources says that is the configuration that fails. Two design principles bear directly. Collective-choice arrangements: those affected by the rules participate in modifying them, and externally imposed rules routinely violate this because they don't fit local conditions and nobody who lives with them owns them. Monitoring: in the durable cases monitors are accountable to the users, not to an external authority. A centrally-written playbook mandated onto nine operating partners is externally imposed rules with external monitoring — complied with formally and routed around substantively, which is the standard outcome every firm reports. A matching result from the agent-market literature: comparing eight coordination mechanisms across 200 rounds, the winner was mediation, where participants proposed and voted on the mediator's rules themselves, at 1556 cumulative honest-agent utility against a 1209 do-nothing baseline. Top-down governance with an external monitor and escalating penalties came third at 1288, and binding contracts imposed on the parties finished fifth, below doing nothing, at 1130. What it cost: I proposed the thing that produces compliance instead of the thing that produces comparability. The correction: the fund supplies the machinery, the operating partners author the content. Fix the decision type and the artifact shape — structural and cheap to agree — and let the partners define the failure vocabulary collectively and revise it. That is a two-hour working session, not a methodology document, and it is the version that survives contact with nine people hired for judgment. | −$0 | 9 engagements, 0 comparable pairs |
| 4. What a positive rate is worth With comparable outputs, three things become available that currently aren't. Prior-based scoping: engagement ten opens with six comparable priors on similar operating shapes, so the diagnostic phase — usually the expensive half — shortens materially because the hypothesis space is pre-narrowed. Cross-portfolio pattern detection: if four of nine portfolio companies show exception load migrating to adjacent functions after automation, that is a fund-level finding about a class of vendor or process, and it cannot be seen from inside any single engagement. A transferable claim at exit: documented, comparable evidence that the same structural intervention improved outcomes at three portfolio companies is a different object in a sale process than an assertion that the sponsor is operationally strong. | +$180,000 to +$400,000 per subsequent engagement | $180,000 to $400,000 per subsequent engagement, before exit-narrative value |
| Realised | A learning rate of zero converts a fund-level asset into a services cost line |
Nine instances of learning paid for, none banked.
Value at risk · multiple 9x, illustrative — though the direct cost here is a spend line rather than a multiple
$180,000–$400,000 of avoidable diagnostic cost across, say, six further engagements over a hold cycle. The multiple is illustrative and the arithmetic is exposed; substitute your own.
$1.1M to $2.4M of value-creation spend buying capability the fund already paid for once — plus an unpriceable exposure: a fund with nine AI engagements and no comparable record cannot answer an LP asking which operational levers actually work — and cannot answer it for itself when allocating the next round
When it surfaces. During fundraising, when an operating-model claim meets a diligence question — and internally every time an operating partner starts an engagement from zero without knowing a colleague solved the same structural problem two years ago.
THE ADD-BACK · episode 09 · diligence pack
Artifact: The two documents.
Twenty minutes, and it is the whole diagnosis.
Artifact: Nine sentences.
The ones that can't be written are engagements that terminated nothing, per Episode 08.
Artifact: Two answers.
If they describe the same phenomenon in different words you have a vocabulary problem, not a knowledge problem — fixable in one session.
Artifact: A name, or its absence.
Absence is the finding, and it is the norm.
Artifact: A one-page schema with their names on it.
Per Line 3, authorship is the mechanism. A document they wrote survives; a document they were sent does not.
Disqualifier
If the response to item 5 is to commission a methodology from an external firm, the mechanism has been inverted and the output will be complied with rather than used.
| Claim | Source | Sample | Class |
|---|---|---|---|
| Design principles present in durable common-pool institutions and absent from most failed ones; collective-choice arrangements and user-accountable monitoring among them | Ostrom, Governing the Commons, 1990, and the design-principles literature | Decades of documented field cases, some institutions durable for centuries | measured |
| Mediation with participant-authored rules won at 1556 cumulative honest-agent utility vs 1209 do-nothing baseline; top-down governance third at 1288; binding contracts fifth at 1130, below doing nothing | Agent-market coordination study, eight mechanisms | 8 conditions × 200 rounds — note n=1 per condition; treat ordering as indicative, not the margins | measured |
| $180,000–$400,000 of avoided diagnostic cost per subsequent engagement | Our estimate, exposed so it can be substituted | n/a — illustrative | illustrative |
| A fund with nine AI engagements and no comparable pairs | Composite | n/a — composite | composite |
One standing caveat. Every number in this show is somebody else's measurement, and I'll tell you whose, with the sample. None of it is diligence on your deal. Do that yourself.