The decision is about who owns the unknown
Buyers usually argue this as a cost question, and it is not one. The same team doing the same work bills roughly the same amount either way. What changes is who absorbs the difference when the work turns out larger than the estimate. FAR 16.202-1 states the fixed-price position without softening it: a firm-fixed-price contract "places upon the contractor maximum risk and full responsibility for all costs and resulting profit or loss." FAR 16.601 states the other end just as plainly, warning that a time-and-materials contract provides "no positive profit incentive to the contractor for cost control or labor efficiency" and therefore requires government surveillance. Everything between those two sentences is an argument about which party is better positioned to eat a surprise.
That framing matters because the honest answer changes partway through most AI projects. Before anyone has touched the data, the seller cannot price the build, and a fixed price at that moment is either padded or reckless. After two weeks inside the data, the same build is often fully specifiable and a fixed price becomes the cheaper instrument for both sides. A contract written as if the uncertainty were constant across the whole period mismatches the structure to the risk for half the work.
The useful question is not "fixed price or time and materials." It is: which specific unknown am I trying to survive, when does it resolve, and who can act on it once it does.
Four unknowns, not one
"Scope is uncertain" is a summary, not a diagnosis. On AI work the uncertainty decomposes into four things that resolve at different times, sit with different parties, and call for different treatment.
Data uncertainty. Nobody knows what is actually in the tables until someone competent looks. Column meanings drift, join keys collide, the labeled set turns out to be labeled by three people who disagreed, and half the historical records came from a system retired in 2014. This is the unknown that most often blows a fixed price, and the one the buyer is best positioned to retire, because the buyer owns the data and controls access to the people who understand it.
Performance uncertainty. No engineer can promise an accuracy number on data they have not seen, and any who does is guessing or has done the work already. This is a physical limit, not vendor caution. It resolves only after a baseline exists on the buyer's real data, and until then a fixed price carrying a performance guarantee is priced with a risk premium the buyer pays whether or not the risk materializes.
Environment uncertainty. Access approval, credentials, network paths, a place to run the code, the security review that clears before a container touches anything real. This sits entirely inside the buyer's organization, routinely consumes more calendar than the modeling, and is the least defensible thing to push onto a fixed price. A seller who accepts schedule risk on access they do not control has agreed to be punished for someone else's queue.
Definition uncertainty. Nobody has written the sentence that says what "done" means. This one differs from the other three: it is not a fact waiting to be discovered but a decision waiting to be made, and no contract structure can absorb it. Time-and-materials will happily bill against an undefined finish line forever. A fixed price with an undefined finish line becomes a dispute. Definition uncertainty has to be removed, not allocated.
What moves a segment of AI work toward a defensible fixed price
Editorial weighting from the regulation and practitioner reading. Illustrative, not a measured statistic.
What the regulation actually requires
Federal buyers do not get to pick a structure on preference alone, and the constraints are worth knowing even for commercial buyers, because they encode decades of expensive lessons.
FAR 16.103(a) treats contract type as "a matter for negotiation" and adds that negotiating the type and the price "should be considered together." FAR 16.103(b) sets the default: a firm-fixed-price contract "shall be used when the risk involved is minimal or can be predicted with an acceptable degree of certainty," but "when a reasonable basis for firm pricing does not exist, other contract types should be considered." That second clause is the one AI buyers keep forgetting exists.
FAR 16.104 lists the factors, and several land on AI work. Under (d), "complex requirements, particularly those unique to the Government, usually result in greater risk assumption by the Government." Under (l), "contractor risk usually decreases as the requirement is repetitively acquired," which is why a fifth deployment of a known pipeline prices differently from a first. Under (i), before agreeing to anything other than firm-fixed-price the contracting officer must confirm the contractor's accounting system can produce the cost data on time.
Time-and-materials sits under FAR 16.601 and carries its own admission price. It "may be used only when it is not possible at the time of placing the contract to estimate accurately the extent or duration of the work or to anticipate costs with any reasonable degree of confidence." The contracting officer must prepare a determination and findings that no other contract type is suitable, and the contract must include "a ceiling price that the contractor exceeds at its own risk." Where the base period plus options exceeds three years, that determination has to be approved by the head of the contracting activity before the base period is executed. Ordering off a Federal Supply Schedule carries the same shape under FAR 8.404(h).
For commercial products and services the door is narrower still. FAR 12.207(a) says agencies "shall use firm-fixed-price contracts or fixed-price contracts with economic price adjustment." Paragraph (b) allows time-and-materials or labor-hour for commercial services only when competitive procedures were used, and the determination must contain "sufficient facts and rationale to justify that no other contract type ... is suitable," document the market research, and describe how the acquisition will move toward firm-fixed-price on future buys.
Two more anchors matter. FAR 37.102(a) makes performance-based acquisition the preferred method for services and sets an order of precedence: a firm-fixed-price performance-based contract or task order first, then a performance-based contract that is not firm-fixed-price, then a contract that is not performance-based. And FAR 35.002 concedes what every research buyer already knows, that "unlike contracts for supplies and services, most R&D contracts are directed toward objectives for which the work or methods cannot be precisely described in advance," with FAR 35.006 pointing basic and applied research toward cost-reimbursement.
One currency note for 2026. The Revolutionary FAR Overhaul is live and moving, with model deviation text, practitioner materials and a FAR Companion published on acquisition.gov, and agencies operating in places under class deviations rather than the codified text. The Part 16 text quoted above carries FAC 2026-01, effective March 13, 2026. Confirm the deviation text your contracting activity uses before quoting a paragraph number at a contracting officer. The underlying economics have not moved.
The two structures, side by side
| Dimension | Firm-fixed-price | Time-and-materials or labor-hour |
|---|---|---|
| Who absorbs an overrun | The seller, entirely. FAR 16.202-1: maximum risk and full responsibility for all costs and resulting profit or loss | The buyer, up to the ceiling. Past the ceiling the seller continues at its own risk or stops |
| What the buyer must produce first | A finish line specific enough to test, and the inputs needed to reach it | A ceiling, a labor-category rate table, and someone with time to supervise |
| What the contract must contain | Acceptance criteria and a deliverable list. Payment follows acceptance, not effort | Fixed hourly rates covering wages, overhead, G&A and profit; materials at actual cost |
| Federal paperwork | None beyond the ordinary. It is the default under FAR 16.103(b) and 12.207(a) | A determination and findings that no other type is suitable, plus HCA approval when base and options exceed three years |
| Where it fails on AI work | Priced before anyone has seen the data, so the premium is either padding or a loss recovered through scope arguments | Effort continues without a finish line. Burn is visible, progress is not, and the ceiling arrives before the outcome does |
| Best fit | Assessment, integration, hardening, documentation, transition, and any build repeated from a known baseline | Genuine discovery, first-of-kind modeling on unseen data, and work gated by access the seller does not control |
The paperwork is a signal, not an obstacle
Sellers read the determination-and-findings requirement as friction. It is better understood as a message about where the burden of proof sits. The default is fixed price; time-and-materials requires a government official to write down, on the record, why no fixed-price arrangement would work. Asking for that structure means asking a contracting officer to author that document, and the fastest way to get it is to hand over the facts it needs: what specifically cannot be estimated, what evidence supports that, and what will be known at the end of the period that is not known now.
The payment clause is worth reading before signing rather than after. Under FAR 52.232-7, hourly amounts are the contract rates times direct labor hours performed, materials are reimbursed at allowable cost with no profit or fee to the prime on materials, and the government "will not be obligated to pay the Contractor any amount in excess of the ceiling price." The clause also permits the contracting officer to withhold five percent of amounts due, capped at $50,000 for the contract. That withhold is working capital sitting on someone else's balance sheet, and it belongs in the seller's cash plan from day one.
A third structure sits between the two poles and gets overlooked. A firm-fixed-price level-of-effort term contract buys a specified level of effort over a stated period for a fixed dollar amount, on work "that can be stated only in general terms." FAR 16.207-3 confines it to cases where the work cannot otherwise be clearly defined, the effort is agreed in advance, there is reasonable assurance the result cannot be reached with less, and the price is at or below the simplified acquisition threshold unless the chief of the contracting office approves more. That threshold is $350,000 under FAC 2026-01. For a bounded research question with an uncertain answer, it is often the cleanest instrument on the shelf.
What makes a fixed price safe to write
A fixed price is priceable when three conditions hold at once. The input is bounded: a named dataset, system or corpus, with a size that will not triple after signature. The output is a named artifact: a pipeline, a model with a model card, a service with an interface contract, a report with a defined shape. And the acceptance test is computable by both parties on evidence both can see. If any of the three is missing, the price on the page is a guess wearing a suit.
The trap specific to AI is the second and third conditions colliding. A buyer wants an accuracy number in the acceptance criteria, which is reasonable, because that number is what makes the deliverable meaningful. But a number promised before anyone has measured a baseline on the real data is set by negotiation rather than by physics. Sellers respond in one of two ways, both bad: they price in a premium the buyer pays regardless, or they accept the number and then spend the engagement arguing about which records should count.
The way out is sequencing, not cleverness. Pay a fixed price for the measurement. A short assessment on the real data produces a baseline, an error taxonomy, and a stated confidence interval. Then set the build's performance target against that baseline, in a second priced segment, with an acceptance test naming the held-out set, the metric, the threshold, and who computes it. The buyer has purchased the right to write a defensible fixed price, which is cheaper than buying a risk premium.
Split the project rather than the difference
The regulation anticipates this, and buyers rarely use the provision. FAR 16.102 permits negotiated contracts to be "of any type or combination of types that will promote the Government's interest," subject to the prohibition on cost-plus-a-percentage-of-cost. FAR 16.104(e) is more pointed: "if the entire contract cannot be firm-fixed-price, the contracting officer shall consider whether a portion can be established on a firm-fixed-price basis." That is the answer to almost every AI scoping argument, and it is a shall.
In practice this means separately priced segments: CLINs on a federal award, phases in a commercial statement of work, or options priced at signature and exercised later. The shape below survives contact most often.
A five-segment structure that matches price type to the live unknown
Segments one, two, four and five are fixed price without strain, because in each the finish line is visible before the work starts. Only segment three carries live performance uncertainty. A project that looked unpriceable turns out to be roughly two-thirds firm on day one, which beats a single blended arrangement for the buyer and one fixed price carrying every unknown at once for the seller.
One caution on segment three. Written as open-ended effort with a ceiling and nothing else, it will consume the ceiling. Give it a stop condition that is not the money: a decision point at a stated date where the measured result is reviewed against the criteria from segment two and the buyer chooses to continue, stop, or re-scope. A ceiling limits exposure. A decision point limits waste, which is the larger loss.
Running the flexible segment without losing control
Time-and-materials earns its bad reputation through administration, not economics. FAR 16.601 requires government surveillance precisely because the structure removes the seller's incentive to be efficient. Buyers who take the structure without the supervision get what the regulation predicts.
Treat the ceiling as exposure, not budget. A ceiling is the maximum the buyer can lose, not the amount the work should cost. If the ceiling and the estimate are the same number, the estimate is doing no work.
Require burn reporting on a fixed cadence, in the same units as the plan. Hours by labor category against the segment, weekly or biweekly, next to what was delivered. A burn report without a delivery column is an invoice with extra steps.
Demand a running artifact, not a status narrative. Something that executes: a notebook, a script, an evaluation run, an interface. Progress on modeling work is visible in artifacts and invisible in prose.
Fix the rate table and labor categories at signature, with named qualifications per category. Substitution without notice is where quality quietly leaves.
Name the person who can say stop. The most expensive time-and-materials engagements are the ones where nobody on the buyer's side had the authority to end a line of work that stopped paying for itself.
Performance-based terms, and the AI-specific clauses
OMB Memorandum M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, issued April 3, 2025, rescinded and replaced M-24-18 and applies to contracts awarded under solicitations issued 180 days or more after issuance, putting it in force from the end of September 2025. It does not cover AI acquired for a National Security System. For any buyer structuring an AI contract, federal or not, it names the terms that turn an outcome into something testable.
It encourages performance-based techniques by name: statements of objectives and performance work statements rather than over-specified statements of work, quality assurance surveillance plans as the place where metrics get negotiated rather than assumed, and incentives tied to metrics "correctly tied to desired business and mission outcomes."
The testing language is worth copying into commercial agreements verbatim. Agencies must use data they have defined for independent evaluation, and that data "should not be accessible to the vendor, and should be as similar as possible to the data used when the system is deployed." Contracts must detail the vendor's examination, testing and validation procedures and must not prohibit the agency from internally disclosing how testing was conducted or its results. A buyer who writes only that clause has removed most of the ways an accuracy claim can be true and useless at the same time.
The memo also directs terms on vendor lock-in, including knowledge transfer, data and model portability, and rights to code and models produced in performance; a permanent prohibition on using non-public agency data and outputs to train publicly or commercially available models absent explicit consent; monitoring on a stated cadence; and sunset criteria for when continued use should be reconsidered. None of these are about price type. All of them decide whether the thing being bought stays worth its price after the contract type stops mattering.
Commercial buyers face the same physics
Strip the citations away and a commercial buyer is solving an identical problem with fewer guardrails, which cuts both ways. Nobody will make a commercial buyer write a determination and findings, or stop them signing a fixed price for work that cannot be specified. Four structures cover most of the commercial ground. Milestone fixed price works wherever the artifact is nameable and testable, and should be used for every segment where that is true. Capped time and materials, sometimes called not-to-exceed, is the same instrument the federal side surrounds with paperwork, and it deserves the same discipline: a ceiling, a rate table, cadence reporting, and a decision point. A retainer suits sustainment and advisory work where the value is availability rather than a deliverable. Outcome or gain-share pricing is attractive and usually fails, because it requires both parties to agree on a counterfactual, and counterfactuals do not survive audit. When the measured business effect depends on a dozen decisions the seller does not control, the pricing model converts every performance conversation into an attribution argument.
Two commercial failure modes deserve naming. The first is the change-order gap: a fixed price signed against a scope document nobody re-reads, with changes agreed verbally in standups until the final invoice arrives and the two sides discover they had different projects. Written change control, even one paragraph per change, costs less than the argument. The second is the pilot-to-production cliff, where a production request is priced from the pilot's number but quietly includes reliability, security review, monitoring, documentation, on-call and support that were never in the pilot. Price production as production, and say so at the pilot's outset.
Questions that settle the structure in one meeting
These are the questions we work through before quoting. A buyer who arrives with the answers usually leaves with a price.
- Has anyone competent looked at the actual data, or only at a schema? These produce different projects.
- Does a measured baseline exist on the real records, and if so, who measured it and on which held-out set?
- Is environment access granted or promised? Name the date credentials exist and the person who grants them.
- Who signs acceptance, and what will they look at? If the answer is a committee, the schedule already contains an unbudgeted month.
- What is the smallest useful thing deliverable in four weeks that would still be worth having if the rest never happened?
- What happens to the data and the model at the end? Portability, rights to code and models produced under the contract, and whether anything trains anything else.
- What is the sustainment plan, and is it priced? A model with no monitoring plan has an expiry date nobody wrote down.
- If the flexible segment must stop early, what stops it? A named date, a named reviewer, a named criterion.
Bottom line
Contract structure is risk allocation with a price tag attached, and the mistake is treating one project as one risk. Split it. Pay a fixed price for the segments where the finish line is already visible, which on most AI work is assessment, integration, hardening, transition and sustainment. Use a bounded flexible arrangement for the one segment carrying live performance uncertainty, and put a decision point in it so the ceiling is not what ends the conversation. Write acceptance criteria against a measured baseline rather than a negotiated hope. FAR 16.104(e) said this long before anyone was buying models: look for the portion that can be firm-fixed-price. On a well-structured AI project that portion is larger than either party expects at the first meeting.
Frequently asked questions
When the input set is bounded and already available, the deliverable is a named artifact, and the acceptance test can be computed by both parties on evidence both can see. That describes assessment, integration, hardening, documentation, transition, sustainment, and any build repeating a baseline already measured on similar data. It does not describe first-of-kind modeling on data nobody has profiled, where a fixed price is either padded with a premium the buyer pays regardless or accepted at a loss the seller recovers through scope arguments.
Under FAR 16.601, a determination and findings that no other contract type is suitable, plus a ceiling price the contractor exceeds at its own risk. Where the base period plus options exceeds three years, the head of the contracting activity must approve that determination first. For commercial services, FAR 12.207(b) also requires competitive procedures, documented market research, and rationale showing how the acquisition will move toward firm-fixed-price on future buys.
Yes, and the regulation encourages it. FAR 16.102 allows negotiated contracts to be of any type or combination of types that promotes the government's interest, apart from the prohibited cost-plus-a-percentage-of-cost arrangement. FAR 16.104(e) directs the contracting officer to consider whether a portion can be firm-fixed-price when the whole contract cannot. That means separately priced segments: CLINs, phases, or priced options exercised later.
Against a measured baseline, never against a number set in negotiation before anyone has seen the data. Pay a fixed price for a short assessment that measures the baseline on the real records, then set the build target relative to that measurement. The criteria should name the held-out set, the metric, the threshold, and who computes the result. OMB Memorandum M-25-22 directs agencies to evaluate on data they define that is not accessible to the vendor, a standard worth adopting commercially.