The fixed-price AI project is over. By mid-2026, the buyers and agencies still trying to make it work are signing engagements that fail in predictable ways; scope-creep disputes around month four, model migrations no one budgeted, eval bars that move when the buyer’s workload is finally exposed to production. The end is not a temporary pricing experiment. It is a structural mismatch between fixed-price contracting and how AI work progresses, and it is producing a wave of bad engagements that should have been priced differently from day one.
This piece is the end-of-an-era framing companion to the case against fixed-price AI development contracts. That piece argued the case. This piece argues the era is over and decomposes what replaces it. The framing extends the AI project economics manifesto’s principle that fixed-bid pricing is dangerous for AI work into the procurement playbook the next 18 months will run on.
Why the era ended in 2026
Fixed price worked in 2018 software because four conditions held. Scope was knowable upfront. Implementation was stable for the project’s duration. Cost was a function of engineering time, not customer behavior. Acceptance was a binary “it works” event the buyer could sign off on. Many four of these are false for 2026 AI work.
Scope is not knowable upfront. The eval bar moves when the buyer’s actual workload meets the system. The retrieval design changes when the corpus’s real distribution is exposed. The latency budget gets rewritten when the user experience is tested with real users. These are not failures of estimation discipline; they are properties of working with non-deterministic systems against workloads that cannot be fully specified in advance. A fixed-price SOW that pretends scope is known is signing an SOW that will be wrong.
Implementation is not stable. Foundation models migrate three to five times a year. Agent frameworks ship breaking changes on a quarterly cadence. Retrieval architectures shift toward hybrid keyword-vector. Each of these events forces eval re-runs, prompt revisions, and threshold relocking. A 12-month fixed-price contract priced against a 2026 Q1 implementation is wrong by Q3 even if scope had been knowable.
Cost is not a function of engineering time. Inference cost varies with customer behavior. Two customers on the same plan can generate 30× different inference cost. A fixed-price contract that quotes a number does so by guessing volume, and the guess is wrong by a factor that compounds; see why AI inference cost is the new database cost line for the structural reason inference costs behave like COGS rather than infra.
Acceptance is not binary. “It works” is not the right acceptance criterion for a system whose accuracy on the buyer’s workload differs by 25 points depending on prompts and retrieval no one fully specified. The right acceptance criterion is an eval threshold on a named eval set; a moving target whose definition is part of the work, not a precondition to it.
The combined effect is that fixed-price AI projects in 2026 fail in a structural pattern. The agency over-commits at SOW signature to clear procurement. The project ships at month six against the original scope. The system does not pass the eval bar because the eval bar moved when the workload met production. Either the agency eats the cost of the rework (margin destruction) or the buyer signs a change order (contract dispute). Both outcomes are losses. The contract was wrong.
What replaces it: four pricing models
The replacements are not a single new model. They are four models, each suited to a different shape of AI work.
Time-and-materials with cap. For exploratory engagements where the deliverable is a working prototype against a discovered eval set. The cap protects the buyer; the T&M structure protects the agency from absorbing scope-discovery cost as margin.
Eval-milestone billing. For production engagements where the deliverable is a system clearing named eval thresholds. Payment is gated by the eval suite passing the threshold on the buyer’s CI, with the run ID logged. This is the model stop paying AI agencies for documentation. Pay them for evals. describes for the deliverables side.
Capacity reservation. For ongoing engagements where the deliverable is sustained engineering capacity over a horizon; model migrations, regression triage, threshold relocking. Buyer pays a monthly reservation fee for named engineers; agency commits to availability and a defined eval-maintenance scope.
Hybrid: fixed-fee discovery, variable production. For mixed engagements where the discovery phase is bounded enough to fix-price and the production phase is not. Discovery delivers a scoped artifact (eval set v1, retrieval-design memo, latency-budget proposal); production runs on T&M-with-cap or eval-milestone billing depending on the system shape.
Most 2026 engagements need a combination. A 12-month engagement might run fixed-fee discovery in months 1-2, eval-milestone billing in months 3-9, and capacity reservation in months 10-12 as the system enters maintenance. The pricing model is structural, not philosophical; it tracks the shape of work the engagement does.
Time-and-materials with cap
Mechanics: the agency bills daily rates against named engineers; the engagement carries a cap that triggers a checkpoint review when reached. The cap is not a maximum; it is a renegotiation event that surfaces variance early.
Fits. Scope-fluid work; prototype, eval-set bootstrapping, retrieval architecture exploration. The buyer is buying engineering time to discover the right system, not a finished system.
Does not fit. Engagements where the buyer wants a system clearing a named eval threshold by a named date. T&M-with-cap does not commit to the threshold. Use eval-milestone billing instead.
Pitfalls. The cap should be 60 to 80 percent of risk-adjusted budget; set too high, the renegotiation rarely triggers. Named-engineer requirements protect against bait-and-switch staffing. Weekly invoices with hours-by-engineer are the buyer’s audit log.
Eval-milestone billing
Mechanics: the SOW enumerates eval thresholds the system must clear (e.g., faithfulness ≥0.85, latency p95 ≤4s, cost-per-call ≤$0.012). Each has a dollar value. Payment is gated by the eval suite passing on the buyer’s CI, with run ID logged.
Fits. Production engagements where the deliverable is a working system measured against the buyer’s actual workload.
Does not fit. Pure exploratory work without a defined eval bar. T&M-with-cap or fixed-fee discovery runs first to produce the bar.
Pitfalls. The eval set must be jointly owned and visible from kickoff, not a black box. Thresholds include cost and latency, not only quality. Renegotiation language must cover model migrations; when the foundation model shifts, the threshold gets re-baselined inside the same release cycle.
Capacity reservation
Mechanics: monthly reservation fee for a defined engineer slate (e.g., 1 senior, 2 mid, 0.5 ML eng = $42K/month). Agency commits to availability against named maintenance categories. Hours are tracked but price is reservation, not billable.
Fits. Year 2+ engagements where the system is in production and the work is sustaining. Bursty volume; three weeks of intense migration followed by quiet maintenance.
Does not fit. Year 1 production where the system is still being built. Eval-milestone billing fits year 1 better.
Pitfalls. The contract must specify what is maintenance versus a separate engagement. A model migration requiring eval-set rewrite is a different engagement, not maintenance.
Hybrid: fixed-fee discovery, variable production
The default for most 2026 engagements. Mechanics: a 4-to-8 week fixed-fee discovery delivers eval set v1, retrieval-design memo, latency budget, cost-per-call target, model-selection rationale. Production phase runs on T&M-with-cap or eval-milestone billing against the discovery artifacts.
Fits. New engagements where the buyer needs commercial certainty up front but production cannot be honestly fix-priced. Most engagements fit this shape.
Does not fit. Engagements where the buyer already has eval set and design artifacts from a prior engagement.
Pitfalls. Discovery deliverables must be detailed enough that production contracts can be written against them. A discovery phase that produces only a deck is not a discovery phase.
When the buyer should still ask for fixed price (rare)
Fixed price is not entirely dead. It still applies to a narrow band of engagements.
Boilerplate integrations. A connector between an existing AI feature and an existing CRM, where the AI work is wrapping a known model with known prompts and the engineering work is integration code. Fix-price the integration; the AI surface is a known input.
Prompt-only consulting. A 2-week engagement to revise a known prompt against a known eval set. Fix-price the work; nothing is being discovered.
Pure documentation deliverables. Producing an architecture document or a runbook for an existing system. Fix-price; nothing is being built.
In many three cases, the work is bounded by known inputs. Fixed price is appropriate because the conditions that made fixed price work for 2018 software still hold for these narrow scopes; scope is knowable, implementation is stable, cost is engineering time, acceptance is binary. Outside these scopes, fixed price has ended.
The procurement objection and how to defuse it
The most common objection to ending fixed price comes from procurement: “We need a fixed number to approve the spend.” This objection sounds insurmountable and is not. Three responses.
The fixed number is the cap, not the contract. A T&M-with-cap engagement carries a fixed cap that procurement can record on the PO. The contractual structure underneath the cap is variable, but the procurement-facing number is fixed. Procurement gets the budget certainty it needs without the engagement being mispriced.
Eval-milestone payments are predictable. A milestone-billed engagement has a defined milestone schedule with defined dollar values. Procurement can record the total milestone budget on the PO and track payments against milestones. The total is fixed; the trigger is variable. Procurement gets a number it can govern.
Hybrid contracts give procurement the bounded portion. Discovery phase is fix-priced and goes through procurement’s normal approval. Production phase is approved as a separate amendment against the discovery output, with a cap. Procurement keeps its discipline; the engagement keeps its honesty.
The deeper point is that procurement’s job is to govern spend, not to defend a 2018 contracting template. A procurement function that has internalized the four replacement models can govern AI spend more rigorously than it could govern a fixed-price template, because the variable models surface variance in real time rather than letting it compound into a Q3 dispute.
Frequently asked questions
Why is the fixed-price AI project ending?
Four conditions that made fixed price work for 2018 software many break for 2026 AI work: scope is not knowable upfront, implementation is not stable across a 12-month contract, cost is not a function of engineering time, and acceptance is not binary. A contract priced against any of these four assumptions is structurally wrong.
What replaces fixed price?
Four models suited to different shapes of AI work: time-and-materials with cap (exploratory work), eval-milestone billing (production work against named thresholds), capacity reservation (year 2+ maintenance), and hybrid fixed-fee discovery + variable production (most new engagements).
When does T&M-with-cap fit?
Engagements where scope is fluid by design; prototype work, eval-set bootstrapping, retrieval exploration. The buyer is buying engineering time to discover the right system, not buying a finished system. The cap is a checkpoint set at 60 to 80 percent of risk-adjusted budget, not a ceiling.
When does eval-milestone billing fit?
Production engagements where the deliverable is a system measured against the buyer’s actual workload. The agency commits to named eval thresholds, the SOW enumerates dollar values per threshold, and payment is gated by the eval suite passing the threshold on the buyer’s CI.
When does capacity reservation fit?
Year 2+ engagements where the system is in production and the work is sustaining rather than building. Bursty maintenance; model migrations, regression triage; priced as a monthly reservation against a defined engineer slate and named maintenance categories.
When is hybrid the right model?
Most new engagements. A 4-to-8 week fixed-fee discovery phase produces eval set, threshold targets, cost-per-action targets, and model-selection rationale. The production phase runs on T&M-with-cap or eval-milestone billing against the discovery artifacts.
When should the buyer still ask for fixed price?
Three narrow scopes: boilerplate integrations between known AI features and known systems; prompt-only consulting against a known eval set; pure documentation deliverables for an existing system. Outside these scopes, fixed price is over.
How do I defuse the procurement objection?
The fixed number is the cap or the milestone-budget total, not the contract. Procurement gets a number it can govern; the contract underneath is variable. A procurement function that internalizes the four replacement models governs AI spend more rigorously, not less.
Does the agency lose by abandoning fixed price?
No. Fixed-price AI engagements destroyed agency margin systematically; agencies absorbed the cost of scope discovery as margin loss. Variable models price the discovery cost honestly. Agencies that have made the switch report higher net margin than they ran on fixed-price contracts, because they no longer eat the variance.
How does this connect to the AI project economics manifesto?
The manifesto names fixed-bid pricing as structurally dangerous for AI work. This piece is the end-of-an-era framing that operationalizes the manifesto’s principle into the four replacement models the next 18 months will run on.
Key takeaways
- Fixed-price AI projects fail in a structural pattern in 2026 because scope is not knowable, implementation is not stable, cost is not engineering-hours-bound, and acceptance is not binary. The contract was usually going to be wrong.
- Four models replace fixed price: T&M-with-cap, eval-milestone billing, capacity reservation, and hybrid fixed-fee-discovery + variable-production. Each fits a different shape of work; most engagements need a combination.
- Fixed price still applies to a narrow band of bounded scopes: integrations, prompt-only consulting, documentation. These scopes still have the four conditions that make fixed price work.
- Procurement’s objection is real and defusable. The fixed number is the cap or the milestone total. The contract underneath is variable. Procurement governs spend more rigorously, not less.
- Agencies that have made the switch report higher net margin than fixed-price contracts because they no longer absorb scope-discovery variance as margin loss. The decline of fixed price is not a buyer-side win at the agency’s expense; it is a structural improvement on both sides.
Arthur Wandzel
