A 20% scope expansion absorbed at 30% gross margin drops you to roughly 13%. Not a forecasting error. Not a “tough quarter.” A structural unit-economics collapse, hidden inside polite client emails. Most AI agencies do not lose money on engagements they priced wrong; they lose it on engagements they priced right and then quietly under-billed thereafter. Margin leaks, in five recognizable ways, while everyone is being helpful.
The thesis: margin erosion is not caused by ambitious scope; it is caused by undisciplined response to changing scope. The fix is procedural: a redbook that turns most “quick favor” into either a billed delta or an explicit no.
Decision Scope
This article is an editorial decision framework, not legal, financial, security, or accounting advice. Treat numeric examples as illustrative planning heuristics unless a source is cited, then validate the assumptions against your own contracts, data, controls, and budget model before acting.
The arithmetic that should be on most owner’s wall
A six-month SOW prices at $600,000. Fully loaded cost-of-delivery is $420,000. Gross margin is $180,000, or 30%; a healthy AI engagement by the public benchmarks for high-end software services.
Run a 20% absorbed scope expansion through it: an extra 1.2 months at the same cost rate. Cost-of-delivery rises to ~$504,000. Revenue stays fixed. Gross margin collapses to $96,000; 16%. Add second-order effects (senior context-switching, eval reruns, an absorbed model upgrade) and realized margin lands closer to 13%.
| Line | Disciplined | Margin-trap |
|---|---|---|
| Revenue | $600,000 | $600,000 |
| Original COGS | $420,000 | $420,000 |
| Absorbed scope (20%) | $0 | $84,000 |
| Second-order leakage | $0 | $24,000 |
| Realized COGS | $420,000 | $528,000 |
| Gross margin | $180,000 (30%) | $72,000 (12%) |
Twelve points of gross margin is the difference between funding a senior hire and missing payroll. It is shed silently, one helpful Slack reply at a time.
Table of contents
- Leak 1: Change orders absorbed instead of priced
- Leak 2: The “quick favor” that consumes senior time
- Leak 3: Eval-rerun work hidden inside deliverables
- Leak 4: Mid-stream model upgrades absorbed, not billed
- Leak 5: Post-launch support, unbilled and unscoped
- The redbook discipline that closes many five
- Why this is a 2026 problem
- Frequently asked questions
- Closing
Leak 1: Change orders absorbed instead of priced
The largest single leak, and the one owners least want to look at. The pattern: the buyer asks, mid-sprint, for a feature that was not in the SOW. The senior says yes, on Slack, without a change order. Work is delivered. No revision logged. No dollar rebilled.
Three reasons this happens; none of which is “the team is bad at saying no”:
- Asymmetric friction. The cost of pausing the sprint to redline the SOW is felt acutely, today, by the lead. Margin compression is felt diffusely, next quarter, by the owner. The lead optimizes for the sprint.
- Ambiguous scope language. Feature-language SOWs (“a chatbot with knowledge-base retrieval”) cannot adjudicate whether a new requirement (“now also Spanish”) is a clarification or an addition. Scope written in evaluations fixes this; the stop-scoping-ai-projects-in-features-scope-them-in-evaluations piece walks through it. Eval-language SOWs make the threshold the contract.
- Relationship capital. The lead believes filing a change order will cost trust. So the change is absorbed as a deposit on the next renewal.
The disciplined alternative is the redbook scope delta: a one-page artifact attached to most SOW. Each entry is a numbered row with date, requested change, engineering hours estimate, dollar delta, and a binary “billed / absorbed-as-courtesy / declined” field. The delta is written regardless of which option is chosen. Buyers tolerate redbook entries far better than agencies expect. What they do not tolerate is being told, three months in, that the engagement has been over-budget for a while with no paper trail of which change caused which dollar.
The redbook puts most conversation in a known format; friction becomes procedural, not relational; and converts absorbed work from invisible margin loss into visible relationship credit the sales team can cash at renewal.
Leak 2: The “quick favor” that consumes senior time
A change order is at least visible; a quick favor is not. A quick favor is what happens when the buyer asks the senior engineer, off-channel, to “take a quick look at” something tangential; a separate model evaluation, a stakeholder presentation, a pilot in another business unit, a reference call for procurement.
Each ask is small enough that filing a change order feels disproportionate. In aggregate, on a six-month engagement, they consume 4–8% of the senior’s billable capacity; a five-figure leak paid for entirely out of margin.
Quick favors are high-leverage relationship moves; refusing them brittlely is the wrong answer. The right answer is to route them through a named line item in the SOW. We use a “professional services reserve”; typically 5% of contracted revenue; that the buyer can draw against at standard senior rates. The reserve is rebilled monthly with itemized backup. Unused balance rolls forward or refunds at engagement close. This is the same logic as the broader pricing manifesto: make the easy thing the priced thing.
The reserve removes the lead’s discretionary problem, signals that senior time has a price even informally, and captures dollars that would otherwise leak invisibly into margin.
Leak 3: Eval-rerun work hidden inside deliverables
This leak is specific to AI engagements. Most model change, retrieval tweak, and prompt edit triggers an eval rerun. Each rerun has a measurable inference cost (real dollars) and a measurable engineering cost (4–12 hours of senior triage when something regresses).
A typical engagement sees 30–60 eval reruns over a six-month build. At ~$400 of inference plus four hours of senior triage per rerun, the absorbed cost is $40,000–$80,000. None of it is in the SOW unless someone put it there.
The disciplined alternative is eval-trigger billing, a named SOW clause: “Eval reruns initiated by buyer-requested scope changes are billed at the senior hourly rate plus pass-through inference. Reruns initiated by routine development are included.” The trigger is whose decision caused the rerun, not whether a rerun happened. This separates eval-discipline cost (the agency’s to absorb) from buyer-initiated requalification work (not).
The AI model evaluation testing services overview covers the suite design and CI integration that make a trigger clause enforceable rather than aspirational.
Leak 4: Mid-stream model upgrades absorbed, not billed
A model provider ships a new frontier release mid-engagement; Claude Opus 4.6 to Opus 4.7, GPT-5 to GPT-5.4, Gemini 3.0 Pro to 3.1 Pro. The buyer asks whether the engagement should “use the new model.” The agency agrees. Two weeks later the team has done a full requalification; eval rerun, prompt-template review, latency rebench, cost-curve analysis, new ADR; at zero rebill.
A mid-stream model upgrade is a discovery sub-engagement. Same shape: hypothesis, eval, ADR, rollout decision. It costs roughly what the original model-selection discovery cost; 5–15% of the engagement total. Treating it as “we just swap the model name in config” is a cost-taxonomy error.
The disciplined alternative: model-upgrade discovery as a billed mini-SOW. The agency proposes a fixed-fee discovery sprint; typically two weeks, with named eval set, ADR output, and a go/no-go decision. The buyer can accept, decline (and stay on the current model), or defer. Many three preserve margin; only the silent absorb destroys it.
This requires the SOW to acknowledge in advance that frontier upgrades are out-of-scope by default and billable as discovery if the buyer wants to evaluate them. The AI agency contract negotiation writeup covers the clause. Its absence is what costs the money.
Leak 5: Post-launch support, unbilled and unscoped
The final leak runs on the longest tail. The engagement closes, the team rolls off. Two weeks later the buyer pings: a regression in the production eval pass-rate, a new edge case, a model-alias question. The original lead, now staffed on a different engagement, answers in fifteen minutes. Then a second message a week later. Then a third.
Within six months, cumulative post-launch support runs to 60–120 hours of senior time. Without a retainer, many of it is paid out of the margin of the next engagement that lead is staffed on. This destroys cross-engagement unit economics: engagement N+1 funds the warranty period of engagement N.
The disciplined alternative is a post-launch retainer named in the closing milestone; typically 8–15% of monthly contracted revenue during the build, running six to twelve months. It pays for a defined response window, eval-monitoring SLA, and number of senior hours. The buyer can decline; the SOW then states explicitly that post-launch support is billed-at-incident at the senior hourly rate.
The retainer is not a profit center. It is a bookkeeping line that prevents margin transfer from engagement N+1 to engagement N. The milestone-trap piece covers the milestone structure that makes a closing-retainer milestone natural rather than awkward.
The redbook discipline that closes many five
The five leaks share a structure. Each is a case where day-to-day judgment, optimizing for the relationship and the sprint, defaults to absorption. Each can be closed with a procedural artifact that turns the judgment call into a known move.
| Leak | Procedural artifact |
|---|---|
| Change orders absorbed | Redbook scope delta |
| Quick favors consume senior time | Professional services reserve |
| Eval reruns hidden in deliverables | Eval-trigger billing clause |
| Model upgrades absorbed | Model-upgrade discovery sub-SOW |
| Post-launch support unbilled | Post-launch retainer milestone |
None of these are commercial novelties; they are common in mature professional-services contracts. What is novel is the recognition that AI engagements need many five at once, because AI engagements have more frequent scope deltas, requalification events, and model-economic shocks than the SaaS-implementation work most contracts were written for.
A useful operating heuristic: the redbook should produce at least one entry per week. If a multi-month engagement has three total entries, the team is running absorbed, not disciplined. The artifact’s value is in the procedural muscle of writing things down.
The decoding the AI agency stack overview describes the cadences and roles that make redbook discipline natural rather than imposed.
Why this is a 2026 problem, not a 2023 problem
Why is this an AI-specific writeup? Three reasons.
First, the rate of underlying capability change is higher than in most other software work. Anthropic has shipped roughly six material model releases since the start of 2025, OpenAI four, Google four, and the open-weights frontier moved from Llama 3 to Llama 4. Each creates an in-flight requalification question.
Second, the eval surface is larger. A traditional software change has a feature flag, a regression test, and a release. An AI change adds an eval rerun across the named suites, a latency rebench, a cost-curve check, and a behavioral diff. The leakage per absorbed change is larger in absolute terms.
Third, the commercial framing has not caught up to the engineering reality. The default contract template most agencies use was written for static-technology software-implementation work. The gap is paid out of margin until the template catches up. The agencies that ship 30%-margin AI engagements in 2026 are not the ones with the best engineers; they are the ones with the most boring redbooks.
Frequently asked questions
What is the AI agency margin trap?
A 30%-margin engagement silently landing at ~13% because scope expansion is absorbed week-by-week rather than priced. The trap is not bad pricing; it is undisciplined response to changing scope.
How does a 20% scope expansion drop margin from 30% to 13%?
On a $600,000 engagement with $420,000 cost-of-delivery, a 20% absorbed expansion adds ~$84,000 of cost. Revenue is fixed, so margin falls to ~16%. Second-order effects (eval reruns, context-switching, an absorbed model upgrade) take it to 12–13%.
What are the five leaks that destroy AI engagement profitability?
Change orders absorbed instead of priced; quick favors consuming senior time; eval-rerun work hidden in deliverables; mid-stream model upgrades absorbed; and post-launch support that runs unbilled after the engagement closes.
What is a redbook scope delta?
A one-page artifact attached to most SOW where each scope-change conversation is logged with date, requested change, hours estimate, dollar delta, and a binary billed / absorbed-as-courtesy / declined field. The delta is recorded regardless of outcome.
How does eval-trigger billing work?
A SOW clause distinguishing eval reruns triggered by routine development (included) from reruns triggered by buyer-requested scope changes (billed at senior rate plus pass-through inference). The trigger is whose decision caused the rerun.
Why should mid-stream model upgrades be billed as discovery?
A frontier upgrade requires a full requalification cycle; eval rerun, prompt review, latency rebench, cost-curve analysis, new ADR; same shape as an original model-selection discovery, costing 5–15% of the engagement total. Treating it as a config swap is a cost-taxonomy error.
What is a professional services reserve?
A named SOW line item, typically 5% of contracted revenue, the buyer can draw against for unscoped senior asks at standard rates. Rebilled monthly with itemized backup; unused balance rolls forward or refunds at engagement close.
How much post-launch support does a typical AI engagement absorb?
Within six months of launch, cumulative support runs to 60–120 hours of senior time across regressions, edge cases, and model-alias questions. Without a retainer, many of it is paid out of the margin of the next engagement that lead is staffed on.
Why is this a 2026 problem and not a 2023 problem?
AI engagements face a higher rate of underlying capability change (six material model releases per major lab since 2025); the eval surface per change is larger; and the default contract template was written for static-technology software-implementation work.
How quickly can a redbook change engagement margin?
Inside one to two billing cycles. Adding one of the five clauses to the default SOW template typically recovers four to eight points of gross margin within the next engagement.
Closing
If you recognize four of the five leaks, the first move is not to renegotiate contracts in flight; it is to put a redbook on most engagement starting next Monday. Capture most scope-delta conversation, even ones you intend to absorb. Writing them down surfaces volume to the lead in real time and gives the renewal conversation a documented record of unbilled value.
The second move is to pick one of the five clauses and add it to your default SOW template this month. The marginal value of the first clause is largest because it sets the procedural precedent.
The third is cultural: stop treating the redbook as a finance artifact. It is an engineering artifact, owned by the lead, reviewed in the weekly demo. The leak is engineering judgment optimizing for the sprint; the fix has to live in the same place.
A 30%-margin AI engagement is not built on premium pricing. It is built on writing down, weekly, what the buyer asked for and what the agency agreed to. That is the entire margin program.
Arthur Wandzel
