THE ADD-BACK · episode 05
SightingThe most expensive automation failure isn't the one that doesn't work. It's the remediation — aimed at the part that failed, landing on the part that was working.
The consensus readA failed build. Write off the capex and move on.
The mechanismThe write-off is the small number. A fix operators stop trusting doesn't get escalated — they quietly reinsert themselves and check everything, erasing the gain on the ninety percent that never had a problem. That regression is cycle time drifting back to baseline with no line item attached. And the person who must kill it approved it, which is measurably the worst possible evaluator.
The exposureThe reported number understates the damage by roughly 3.5 to 4.4 times.
The testFind a remediation approved in the last two years and pull the process metric before the original build, after it, and after the fix. If the third number is worse than the second, you have this.
The operating case is the same composite as Episode 01. The Dietvorst and Staw findings are real, published and cited with their samples; the two are kept separate throughout.
| Line | Moves | Running |
|---|---|---|
| Claimed — A $180,000 remediation to recover $260,000 of leaked automation savings, presented to the board as a nine-month payback | +$80,000 net claimed | |
| 1. The build cost, which is the number everyone tracks Six weeks of engineering, roughly $180,000 all-in against the composite. It ran, it shipped, and it did not fix the problem: the classifier had to fire before the system had resolved the fact that determined whether a transaction needed routing, and that fact lived in another system. Roughly 61% precision, producing a second queue of misroutes on top of the first. This is the number that gets written off and reported. It is the smallest number in the episode. | −$180,000 | −$180,000 |
| 2. The regression on the part that was working The cost nobody books. Operators stopped trusting the routing, and when people stop trusting an automated router they do not escalate — escalating is slow, political, and marks you as a complainer. They quietly begin checking everything. Six weeks after the remediation shipped, median cycle time had regressed most of the way to pre-automation baseline, across the clean majority of transactions that never had a defect. The mechanism has a measured basis and is not a culture problem: Dietvorst, Simmons and Massey ran five studies where people chose between a model's forecasts and a human's with money on it. The model won in every condition — humans produced 15–29% more error on one task and 90–97% more on another. Participants who watched the model perform, and therefore err, were significantly less likely to bet on it, including the 83% who watched it beat the human. Seeing the human err did not reduce willingness to use the human. Machine errors are not forgiven; human errors are. A remediation that visibly errs in front of operators does not cost you the remediation — it costs you their willingness to rely on the original system, which was working. | −$310,000 to −$420,000 | −$490,000 to −$600,000 |
| 3. Why it ran for eleven months after it was known to have failed The escalation cost, and this line belongs to the sponsor rather than the portfolio company. Staw's work on escalating commitment established the variable, and it is not stubbornness: people commit the greatest additional resources to a failing course of action when they were personally responsible for the earlier decision. Someone who inherits a failing position de-escalates; the person who chose it escalates, because stopping converts a recoverable situation into a permanent verdict on their judgment. The marker is information-seeking — roughly 75% of subjects responsible for a prior failure sought retrospective justifying information, against about 25% of those not responsible. Read that against how a remediation gets reviewed: the ops partner who approved the build presents the review, the board asks whether it is working, and the answer is 'early signal is mixed, we're gathering more data' — the measured escalation response, not diligence. Eleven months of a partially-trusted system running alongside restored manual checking. | −$140,000 to −$190,000 | −$630,000 to −$790,000 |
| 4. Where I was wrong, twice — and the second one is the useful one where I was wrong Both on the record because they compound. The first is the remediation itself, which was my recommendation: given a queue with no owner, put it on the org chart — classify at intake, route to a named person with an SLA. Wrong for a structural reason: I put the boundary where the work looked different instead of where the information actually was. Drawing a boundary around unresolved information doesn't resolve it; it forces a probabilistic guess at the point where you used to have a slow correct answer, converting a slow correct process into a fast wrong one. The second error is worse and it is about governance: after the failure I recommended tighter approval criteria for remediation spend — written business case, defined payback, board sign-off above a threshold. Sensible, and it addresses nothing, because every one of those controls is evaluated by the person who wants the spend approved. What it cost: Staw's finding is that the binding variable is who reads the criterion, not how well the criterion is written. The correction is unglamorous and structural: a remediation review has to be presented by someone who did not approve the original build. Not an advisor, who is chosen and can be unchosen — someone in the fund with no position in the decision. A governance change costing nothing, and the only intervention in this episode with a measured mechanism behind it. I recommended better paperwork when the problem was who signs it. | −$0 | −$630,000 to −$790,000 |
| 5. Total against $260,000 recovered residual Presented as a $180,000 spend to recover $260,000. It cost $630,000 to $790,000 and recovered a fraction of the leak, most of the damage landing on operating performance that was never at risk. | +$0 | −$630,000 to −$790,000 |
| Realised | −$630,000 to −$790,000 against a claimed +$80,000 net |
The reported number — the $180,000 capex write-off — understates the damage by roughly 3.5 to 4.4 times.
Value at risk · multiple 9x, illustrative
$630,000–$790,000 of destroyed run-rate performance at an illustrative 9x, against a project whose written-off capex was $180,000. Substitute your own comps.
$5.7M to $7.1M of enterprise value — against a reported write-off of $180,000
When it surfaces. The capex write-off surfaces immediately and is discussed. The regression surfaces as a slow drift in an operating metric with no owner and no project attached — typically noticed at the next VCP review or, worse, by a buyer comparing the operating metrics in the CIM against the ones two years earlier.
THE ADD-BACK · episode 05 · diligence pack
Artifact: Transaction-level data at all three points.
The third number against the second is the whole diagnosis.
Artifact: Sampled observation, not self-report.
Nobody reports that they've started checking everything.
Artifact: The approval and board records.
If those are the same person, the review has a known and measured bias and should be read accordingly.
Artifact: The list.
Per Staw, post-threshold information-gathering is the escalation response. Treat it as a diagnostic, not as diligence.
Artifact: A written line in the approval memo, before funding.
The one control that matters, and it is free.
Disqualifier
If the answer to question 3 is that the approver also presented the review, you do not need the rest of the pack to know how to read the review. Before approving any fix, establish what is currently working and price the cost of breaking it — remediations are scoped against the failure and evaluated against the failure, while the exposure sits on the volume that never had a problem.
| Claim | Source | Sample | Class |
|---|---|---|---|
| Model outperformed humans in every condition; humans produced 15–29% more error on one task and 90–97% more on another. Participants who watched the model err were significantly less likely to bet on it, including the 83% who watched it beat the human. Seeing a human err did not reduce willingness to use the human. | Dietvorst, Simmons & Massey, Journal of Experimental Psychology: General 144(1), 2015 | Five studies, incentivised choice with money at stake | measured |
| ~75% of subjects responsible for a prior failure sought retrospective justifying information, against ~25% of those not responsible | Staw, escalating commitment literature (1976 onward) | Experimental subjects; responsibility manipulated | measured |
| $180,000 build at ~61% classifier precision; regression across the clean 90% of volume; eleven months of duplicated effort | Composite, built to the shape of a real quote-to-cash automation (same case as Episode 01) | n/a — composite | composite |
| 9x entry multiple | Illustrative mid-market comp | n/a — illustrative, arithmetic exposed | illustrative |
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.