THE ADD-BACK · episode 06
SightingAn automation programme is approved. Eighteen months later everyone agrees the design was wrong, and nobody moves, because the correct alternative now costs three times what it would have before the build.
The consensus readSunk cost. Ignore what's spent, evaluate forward, make the right call now.
The mechanismSunk cost is the wrong frame and it is why the call never gets made. What's spent is gone and irrelevant. What matters is that the price of the alternative rose — integrations, retrained staff, a data model everything depends on, a vendor contract with term left. That is not sunk. It is a live and climbing liability.
The exposureA fund/kill decision is an option that decays, and by month twenty-four it can go negative.
The testOn one live automation, price the alternative today and price what it would have cost at approval. The gap is the number.
Lock-in cost
What the correct alternative costs more than it would have cost before you committed.
The test. Price the second-best topology today, and price it as of the approval date. If the gap is inside normal estimating error, you have no lock-in and this doesn't apply. If it's a multiple, the decision you're deferring is getting more expensive on a clock you're not watching.
Sizing is applied to the same composite programme as Episode 01. The Staw and Dietvorst findings are real, published and cited with their samples; the dry-powder figure is internal market research and flagged as such.
| Line | Moves | Running |
|---|---|---|
| Claimed — A fund/kill decision deferred at month 18 to 'gather more data', at no cost | $0 to defer | |
| 1. Deferral is priced at zero, and that is the error changes the story The decision to wait is almost never written down as a decision. It appears as a review scheduled for next quarter, an additional pilot, a request for more data. Nothing is approved, so nothing is priced. Meanwhile the programme keeps running — integrations keep being written against it, staff keep being trained on it, downstream processes keep taking dependencies on its outputs. Deferral is not a neutral act. It is an unpriced purchase of additional lock-in, and the clock runs whether or not anyone is looking at it. | −$0 | decision valued at $0 to defer |
| 2. What actually rises Four components, all estimable in an afternoon. Integration surface — every downstream system reading the automation's output, each a rewrite in the alternative topology; the one engineering will quote and usually the smallest. Retrained process — staff have adapted to the system's behaviour including its defects, so reverting is a second change programme, harder than the first because the first cost people something. Data model dependency — the most expensive and least visible: once reporting, forecasting or a KPI depends on how the system structures records, the alternative must reproduce that structure or every downstream number moves, and boards notice when numbers move. Contract term — vendor commitments with time left, the only component anyone thinks to check. Against a mid-market programme, eighteen months of accumulation runs $280,000–$450,000 of incremental cost to do the thing you'd have done for less at month zero. | −$280,000 to −$450,000 | −$280,000 to −$450,000 |
| 3. The decision quality degrades on the same clock The half nobody prices, and it has a measured basis. Staw's work on escalating commitment established that people commit the greatest additional resources to a failing course when they were personally responsible for the original decision — the marker being that roughly 75% of subjects responsible for a prior failure sought retrospective justifying information, against about 25% of those not responsible. Overlay that on Line 1: 'let's gather more data before we decide' is not a neutral governance instinct, it is the specific response of a decision-maker who is committed and looking for justification. So two curves run together — the cost of the alternative climbs, and the probability that the person who can authorise it will do so falls. That is why programmes don't get killed at month 18 and do get killed at month 36 by a new operating partner with no position in the original decision. | −$140,000 to −$190,000 | −$420,000 to −$640,000 |
| 4. Where I was wrong — I priced lock-in as technical debt where I was wrong My first model of lock-in cost was the first two components of Line 2: integration surface plus rebuild. Engineering can estimate both, they are defensible in a board paper, and I built the case on them. It produced a number that looked rigorous and was roughly half the real figure. What it missed is the Episode 05 cost: when a remediation visibly errs in front of operators they don't escalate, they quietly reinsert themselves and check everything, erasing the gain on the majority of volume that never had a defect. That is not technical debt — it is the withdrawal of trust from a system that was working, and there is no engineering line item for it. The mechanism is measured: Dietvorst, Simmons & Massey found participants who watched a model err were significantly less likely to bet on it, including the 83% who watched it beat the human, while seeing a human err did not reduce willingness to use the human. What it cost: Roughly half the true figure, understated in a board paper. Every month you run a known-wrong topology you are spending down operator trust in the parts that work — which is the component that makes lock-in cost accelerate rather than accumulate linearly. At the composite's scale, $310,000–$420,000 of annualised operating leverage. | −$310,000 to −$420,000 | −$730,000 to −$1.06M |
| 5. The option curve residual Put the three together and the fund/kill option has a shape worth drawing on one slide. At approval: worth the full delta between the right topology and the wrong one, at zero switching cost. Month 12: the delta is intact, switching cost has grown, the option is worth the delta minus lock-in. Month 24: switching cost approaches or exceeds the delta, and remediation risk means acting can destroy value that isn't currently at risk. Month 36: the decision is made for you, usually by a personnel change. | +$0 | −$730,000 to −$1.06M |
| Realised | A decision worth its full value at month 0 is worth $730,000 to $1.06M less by month 24 |
And the party who can authorise it is progressively less able to — the cost curve and the authority curve run in opposite directions on the same clock.
Value at risk · multiple 9x, illustrative
$730,000–$1.06M of accumulated lock-in and trust erosion at an illustrative 9x, against a decision that cost nothing to make at month zero and was never framed as a decision at all. Substitute your own comps.
$6.6M to $9.5M of enterprise value — against a decision nobody recorded making
When it surfaces. Never, cleanly — which is what makes this the hardest of the ten to act on. Lock-in cost doesn't appear as an event. It appears as a programme everyone privately agrees was a mistake and nobody proposes stopping, booked as ongoing operating cost until an exit process forces the question.
THE ADD-BACK · episode 06 · diligence pack
Artifact: Two estimates, same scope, same team.
The gap is lock-in cost and nobody has ever asked for it.
Artifact: A dependency list from data/BI, not from the programme team.
This is the component that surprises people.
Artifact: The meeting records.
Per Staw, post-threshold information-gathering is the escalation response. Read it as a diagnostic, not as diligence.
Artifact: Names.
If they're the same person, the review has a known and measured bias.
Artifact: One line in the approval memo.
Free, and the only control here that works.
Disqualifier
If nobody can produce estimate 1, the programme has never been compared to an alternative — which means the original approval was a yes/no on one option rather than a choice between two.
| Claim | Source | Sample | Class |
|---|---|---|---|
| ~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 |
| Participants who watched a 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 | measured |
| Lock-in accumulation of $280,000–$450,000 over 18 months; trust erosion of $310,000–$420,000 annualised | Applied to the same composite programme as Episode 01 | n/a — composite, our sizing | composite |
| Roughly $1.3 trillion of global buyout dry powder outstanding, concentrated in aging vintages under pressure to deploy | Internal market research | n/a — not a published measurement; held loosely, direction only | reported |
| 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.