Escalation in small steps

The first five chapters were about valuing flexibility: the right to stop at a gate, the right to buy information and the right to wait. All of them assume that the business will use those rights when the numbers say so. This chapter turns to the force that works against them. Projects rarely stop because of one large decision to continue. They grow through a series of small ones, each reasonable on its own, none of them seen as a decision about the whole project.

The chapter sets out what the research on escalation of commitment found and why it matters to a small business, follows SB36’s development stage through three top-ups between October and December 2025, applies the forward test that should have been run at each one, and ends with the re-gate that the business’s own rule required and nobody called. The surprise in the case is that continuing was the right answer every time. The failure was that the business never asked the question.

What escalation of commitment means

Escalation of commitment is the tendency to keep investing in a course of action after evidence has arrived that it is going badly, and to invest more than a fresh decision-maker would. The term comes from a line of experiments that began with Barry Staw’s 1976 study, “Knee-deep in the big muddy”. In a role-playing business investment exercise with 240 business school students, participants allocated research and development money to one of two divisions of a company, then learned how the chosen division had performed and made a second allocation. Those who had been personally responsible for the first choice, and saw it turn out badly, committed the most further money to the same division. People who inherited the choice from someone else did not show the same pull.

Staw’s 1981 review in the Academy of Management Review described escalation as a cycle rather than a single error. A decision-maker who chose a course of action has a reason to show that the choice was sound; a setback creates pressure to justify it, and the cheapest way to justify it is to spend a little more so that it can still come right. Staw called this self-justification, and listed it alongside other causes: the information arriving piecemeal, the project’s apparent closeness to completion, and the social cost of being seen to give up. Later attempts to replicate the original experiment have had mixed results, and laboratory studies with students are not the same as a family business deciding about its own product. The pattern they describe is still the one to look for, because it does not depend on anyone being careless. It depends only on the person who championed a project also being the person who approves its next instalment.

Chip Heath’s 1995 study adds the remedy that this chapter builds on. People who set a budget for an endeavour tended to stop investing once the budget was exhausted. Escalation was more likely when no budget had been set, or when spending was hard to track. A budget works as a pre-commitment: it records, before the bad news, how much the endeavour was thought to be worth. It works only if someone compares the running total with it.

Why small top-ups escape review

This is an illustration. Corella Plains Implements, the fictional Wimmera maker of trailed boomsprays and chaser bins that runs through this book, approved Stage 3 of the SB36 programme at Gate 3 on 4 August 2025. Stage 3 was the development of two prototype 36 m camera spot-spray booms, with a budget of A$560,000 excluding GST. All money in this chapter is in Australian dollars excluding GST, and in thousands (A$k) where that is clearer.

Two governance rules applied to Stage 3, and both were written down before it started:

• The engineering manager, who leads the programme and is its champion, may approve spending requests of up to A$75k each under delegated authority. Above that, the managing director approves.
• If a stage’s forecast cost rises above 115% of its budget, the stage goes back to a re-gate: a short gate meeting, with the same gatekeepers, that decides whether to continue, change or stop. For Stage 3 the cap was 1.15 × 560 = A$644k.

The first rule is about the size of a single request. The second is about the total. In practice, the business’s spending controls ran on the first rule: each request went through the approval system, which checked it against the approver’s limit. Nothing in the approval system added the request to the previous ones and compared the result with the cap. The finance manager’s monthly commitment register (Chapter 1) showed spent and committed amounts by stage, but not a forecast at completion, so a stage could be on its way past 115% without any line in the register saying so.

That is the structure in which escalation thrives. A request for A$70k is not a decision about whether a A$560k stage should become a A$705k stage. It is presented as a decision about a camera supplier, or a valve, or a test, and it is approved by the person who most wants the stage to succeed.

The Stage 3 record

This is an illustration. Stage 3 received three top-ups between October and December 2025. Each was requested and approved by the engineering manager, and each was below the manager’s A$75k limit.

Figure 1Stage 3 of SB36: budget, top-ups and the gate that moved

1. 4 Aug 2025Gate 3: Stage 3 approvedbudget A$560k; re-gate cap 115% = A$644k
2. 2 Oct 2025Top-up 1: +A$70ksecond camera supplier integration; forecast A$630k, 12.5% over
3. 6 Nov 2025Top-up 2: +A$45kboom section valve redesign; forecast A$675k, 20.5% over, above the cap
4. 26 Nov 2025Gate 4 as plannednot held; prototypes not ready
5. 1 Dec 2025Top-up 3: +A$30kextra bench validation on stubble; forecast A$705k, 25.9% over
6. 17 Dec 2025Gate 4 heldtechnical pass; finance manager finds the missed re-gate

The first top-up, on 2 October, paid for integrating a second camera supplier after the first warned of a long lead time on the module the prototypes used. The second, on 6 November, paid for redesigning the boom section valves, which did not switch fast enough for the nozzle timing at working speed. The third, on 1 December, paid for extra bench validation of detection on stubble, after the bench trial had been run mostly on bare soil.

None of these was frivolous. Each addressed a real technical problem, and each was cheaper than the problem it solved. Together they added A$145k, a 25.9% overrun on the stage budget, and they moved the programme’s dates. Gate 4, planned for 26 November 2025, was held on 17 December. The field season on growers’ farms, planned to start on 1 December, started on 22 December, three weeks late. Because one of the six growers who had agreed to host trials had already sprayed out their fallow by then, and another could offer only part of their area, the season ran on 5 farms and 7,400 ha instead of the planned 6 farms and 9,000 ha: 83.3% of the planned farms and 82.2% of the planned area. A smaller season gives less evidence for the market and field tests that Chapter 4 valued, which is a cost of the overrun that appears nowhere in the A$145k.

Figure 2Stage 3 forecast at completion against its budget and re-gate cap

Budget at Gate 3A$560k

Re-gate cap, 115%A$644k

After top-up 1, 2 OctA$630k12.5% over

After top-up 2, 6 NovA$675k20.5% over: re-gate due

After top-up 3, 1 DecA$705k25.9% over

The business had seen this before. Its own records for 2014 to 2024, set out in Chapter 1, show development budgets exceeded on six of eight products, with actual development spend averaging 1.157 times budget. Stage 3 finished at 705 ÷ 560 = 1.259 times budget, worse than the average but well inside the range of the business’s history, in which one product reached 1.60. A reference class of the business’s own projects would have predicted an overrun of this kind before Stage 3 began. Cutting the tail before approval in Reference Class Forecasting by the Numbers shows how such a history is used before approval: to find where the overrun tail comes from and to fence the riskiest work behind its own gate and walk-away conditions.

The forward test

The question each top-up should have prompted is not “is this request reasonable?” but “given what we now know, is the rest of this stage still worth doing?”. That is a forward-looking question, and it has a forward-looking answer that uses only numbers the business already had.

At Gate 3, Chapter 2 rolled back the decision tree to find the value of the programme at Gate 4, before the field season: V4 = A$1,177.5k. That figure is the expected value of everything after Stage 3: the A$320k field season, the launch decision that follows its signal, and the launch payoffs in the three market states. It is reached only if the prototypes pass the technical gate, which the gatekeepers judged at Gate 3 to have a probability p_T = 0.75. Money already spent on Stage 3 is gone whatever happens next and does not enter the test.

Figure 3The forward test for a top-up

forward value of continuing = p_T × V4 − cost to go

break-even cost to go = p_T × V4

p_T

probability that the prototypes pass the technical gate, reassessed on the day of the request

V4

value of the programme at Gate 4, A$1,177.5k, with the field season and launch decision inside it

cost to go

Stage 3 forecast at completion minus what has been spent to date

1. 2 Oct: 0.75 × 1,177.5 − (630 − 300) = 883.1 − 330 = A$553.1k
2. 6 Nov: 0.70 × 1,177.5 − (675 − 470) = 824.3 − 205 = A$619.3k
3. 1 Dec: 0.72 × 1,177.5 − (705 − 590) = 847.8 − 115 = A$732.8k

This is an illustration. The probabilities on each date are those the engineering manager and the independent board adviser recorded afterwards from the test logs, when the business reconstructed the stage at Gate 4.

• 2 October. Forecast A$630k, spent to date A$300k, so A$330k to go. The second camera supplier had not yet been proven, and p_T stayed at 0.75. Forward value of continuing: 0.75 × 1,177.5 − 330 = A$553.1k. Continuing would have stayed worthwhile up to a cost to go of A$883.1k.
• 6 November. Forecast A$675k, spent A$470k, A$205k to go. The valve problem lowered the technical outlook, and p_T fell to 0.70. Forward value: 0.70 × 1,177.5 − 205 = A$619.3k; break-even cost to go A$824.3k.
• 1 December. Forecast A$705k, spent A$590k, A$115k to go. The redesigned valves had passed their bench test and p_T recovered to 0.72. Forward value: 0.72 × 1,177.5 − 115 = A$732.8k; break-even cost to go A$847.8k.

Each time, continuing beat stopping by a wide margin. On 6 November, for example, continuing would still have been right unless the cost to go had been four times as large, or p_T had fallen below 205 ÷ 1,177.5 = 0.174. The forward value even rose from top-up to top-up, because the money still to be spent fell faster than the probability of success. That is the normal shape of a development stage that is going roughly to plan: the closer it gets to its gate, the less it costs to reach it, and so the more a fresh decision-maker would pay to finish.

Figure 4The forward test on 6 November 2025

Cost to go, A$205k

• Stage 3 forecast A$675k
• less spent to date A$470k
• Sunk spend: not on the scale

p_T × V4, A$824.3k

• p_T 0.70 after the valve problem
• V4 A$1,177.5k from the Gate 3 tree
• Stopping: valued at zero

The same test also shows where continuing would stop being right. The break-even cost to go is proportional to p_T. At p_T = 0.75 it is A$883.1k; at 0.5 it is A$588.8k; at 0.25 it is A$294.4k; and at 0.1 it is only A$117.8k. A stage that is cheap to finish but very unlikely to pass is not worth finishing, and the forward test says so in the same arithmetic that approved Stage 3’s top-ups. Escalation research describes the cases where people keep paying when p_T has collapsed; the forward test is the number that tells them it has.

Figure 5Break-even cost to go for Stage 3 at four technical probabilities

p_T 0.75A$883.1k

p_T 0.50A$588.8k

p_T 0.25A$294.4k

p_T 0.10A$117.8k

Cost to go on 6 NovA$205kp_T was 0.70

Why the process failed when the decisions did not

If every top-up was justified, it is fair to ask what went wrong. Three things, and none of them is visible in the outcome.

First, nobody ran the test. Each request was judged on its technical merits by the person who had asked for it. The arithmetic above was done in December, at Gate 4, as a reconstruction. Had p_T fallen to 0.15 in November, the same process would have approved the same top-up, because the process never looked at p_T at all. A good result produced by a procedure that cannot tell good from bad is luck, not control.

Second, the champion approved their own top-ups. The engineering manager wanted SB36 to succeed, presented it at every gate and had chosen the first camera supplier. That is exactly the position Staw’s experiments found most prone to escalation: personally responsible for a course of action that has hit a setback. None of this says the engineering manager acted in bad faith. It says that the business had placed the decision to continue with the person least able to see it neutrally, which a delegated limit per request does not correct. Sponsors who can stop the work and owners who must make it pay sets out why the authority to stop belongs with someone who does not carry the project’s delivery, and Shared goals, co-location and decision rights in Project Teams and Contracts by the Numbers shows how decision limits can be set by level and threshold so that speed does not cost oversight.

Third, no gate saw the total. The A$75k limit was applied to each request; the 115% cap was meant to be applied to the stage. Because the approval system checked only the first, the second rule existed only on paper. Heath’s finding is that a budget restrains escalation when spending is tracked against it. Corella had the budget and the cap, but no one tracked the forecast against either between gates.

The cost of the failure in this case was modest: three weeks and 1,600 ha of field season, and a A$145k overrun that the forward test would have approved anyway. The same habit is more dangerous later in a programme. At Gate 5 the next request would be the A$1.8 million launch, and Chapter 2’s tree values launching after a Weak field signal at −A$261.2k. A business that continues by default, in small approvals made by the champion, would approve that too. Chapters 7 and 8 return to the launch gate with that risk in mind.

The error: the missed re-gate

This is an illustration. On 6 November 2025 the second top-up took Stage 3’s forecast to A$675k, above the A$644k cap. Under the business’s own rule, Stage 3 should then have gone to a re-gate before more money was committed. No re-gate was called. The approval system recorded a A$45k request within the engineering manager’s limit, and the commitment register at the end of November showed spent and committed amounts for Stage 3 that were each unremarkable. The engineering manager later said they had regarded the cap as applying to the original scope, and the valve redesign as a correction within it.

The finance manager found the breach while preparing the Gate 4 pack on 15 December. Gate 4 papers include a stage close-out, and the close-out compares each stage’s actual cost with its budget. Stage 3 closed at A$705k, 125.9% of budget, and the finance manager traced the dates on which the cumulative forecast had crossed 115%. The first was 6 November.

The gatekeepers at Gate 4 dealt with the consequence and the cause separately. The consequence needed no reversal: the forward test, run on the record of each date, showed that continuing had been right each time, and the stage was closed at A$705k. The cause needed a rule change, adopted at the same meeting:

Figure 6Rules adopted at Gate 4 to stop escalation in small steps

• Do: Forecast at completion on every requestthe approval form shows the stage's new forecast and its percentage of budget
• Do: Cumulative re-gate ruleany request that takes a stage's forecast above 115% goes to a re-gate, whatever its own size
• Do: Top-ups approved by the managing directorthe engineering manager may request but not approve additions to their own programme
• Do: Forward test written on the requestp_T, V4 or the current equivalent, cost to go and the break-even cost to go
• Avoid: Judging each request only against the approver's limitA$75k each, A$145k in total, never seen as one decision
• Avoid: Treating a redesign as "within scope" to avoid the capthe cap applies to the stage's cost, not to its original scope

The second rule is the important one. It changes the trigger from the size of a request to the size of the stage, which is what the gate approved. The third removes the self-approval that Staw’s work warns about, at a cost of a short delay on each top-up. The fourth turns the forward test from a reconstruction into a habit, so that a top-up arrives with the question already answered.

The rule set has limits. It does not prevent a champion from proposing an optimistic p_T, and it does not help if V4 itself is stale. For that reason the business also decided that a re-gate must update V4 with whatever has been learned about the market since the last gate, and that the independent board adviser, who has no stake in SB36, signs off the p_T used in a re-gate paper. Chapter 7 takes the same idea further: the conditions that would stop the programme are written down before the evidence that tests them.

Exercise: a gloomier November

Suppose that on 6 November 2025 the valve problem had been judged more serious, so that p_T was 0.6 instead of 0.70. The forecast and spending are unchanged: Stage 3 at A$675k, A$470k spent, A$205k to go. V4 is still A$1,177.5k. What is the forward value of continuing, what is the break-even cost to go, and should the re-gate have continued, changed or stopped the stage?

Worked answer. The forward value of continuing is 0.6 × 1,177.5 − 205 = 706.5 − 205 = A$501.5k, and the break-even cost to go is 0.6 × 1,177.5 = A$706.5k. Continuing is still clearly better than stopping, by A$501.5k, so the re-gate should have continued the stage. The lower probability reduces the margin by A$117.8k against the actual case (A$619.3k), but the cost to go would have had to rise from A$205k to more than A$706.5k, or p_T fall below 0.174, before stopping became the better choice. The A$470k already spent plays no part in the answer. What the re-gate adds is not a different decision in this case, but a recorded one: the gatekeepers, not the champion, would have looked at p_T and the cost to go, and would have seen the 25.9% overrun coming while there was still time to protect the field season.