The defensible floor for a paid AI pilot in 2026 is $15,000. Below that, the eval set gets cut, then the capability probe, then the PRD — what is left is a demo with a fee. A defensible $15K pilot is two weeks of one senior engineer, produces five named artifacts, and hands off cleanly either to the founder building DIY with Cursor and Claude Code or to a full engagement where the pilot deliverables fold into week one. This article names the five artifacts, defends the floor on unit-economics grounds, and lists the deliverables that quietly disappear at $5K or $8K.
It builds on the founder-AI-partner operating manual and the broader idea-to-product manifesto. Companion reading: the three engagement entry shapes explained, why your AI agency should run a paid pilot, and anatomy of an AI agency engagement: the first 14 days.
Why $15K is the floor, not the average
Most boutique AI partners in 2026 quote paid pilots in a band of roughly $5,000 to $60,000. The band is wide because the word “pilot” is doing too much work. A demo with a slide deck and a hardcoded prompt can be priced at $5,000. A four-week capability study with two engineers and a built eval suite can be priced at $50,000. They are not the same procurement instrument.
The defensible floor is the price at which a pilot can produce evidence a third party could review and reach the same go/no-go conclusion. In 2026 the unit economics put that floor at $15,000 — what a senior contract AI engineer costs for two weeks at US market rates, with the eval set included rather than cut.
If a partner quotes $15K and structures the pilot around five named deliverables, you have a defensible procurement instrument. If a partner quotes $5K or $8K and waves at “a working prototype”, you have a sales demo with a fee.
The five artifacts $15K buys
A defensible $15K pilot in 2026 produces five named, falsifiable artifacts in two calendar weeks of focused work by one senior engineer. Each artifact is independently verifiable.
| # | Artifact | Effort | Falsifiability test |
|---|---|---|---|
| 1 | PRD draft (v0.1) | 2 days | Does it name the user, the job-to-be-done, the success metric, and three explicit non-goals? |
| 2 | Eval set v1 (50 items) | 2 days | Are 50 input/expected-output pairs versioned in a repo, with a stated launch threshold? |
| 3 | Capability probe on real data | 5 days | Is there code that runs end-to-end on a representative sample of the founder’s actual data? |
| 4 | Go/no-go memo with three-option handoff | 1 day | Does it name the eval baseline pass rate and propose a defensible next step (DIY, full engagement, walk)? |
| 5 | Handoff package | 0.5 day | Repo access, one-page runbook, cost-per-query estimate. Can the founder reproduce the run? |
The total is 10.5 engineer-days, which fits in two calendar weeks for one senior. Every artifact maps to a procurement question — what to build, how to know it works, can the model do it on real data, what the next step is, what you take with you. A pilot that ships fewer than five artifacts is cheaper because it is doing less. That is not a deal; it is a smaller purchase.
The unit economics behind the number
Senior AI engineers in the US contract market in 2026 charge $200-$400 per hour, concentrated in the $250-$300 range for two-to-four-week engagements. Stack Overflow’s 2024 Developer Survey reports 76% of developers using or planning to use AI tools — the productivity multiplier from Cursor and Claude Code is baked into senior rates, not offered as a discount.
At $250/hour and 60 billable hours over two weeks, the engineer cost alone is $15,000. That math is the actual reason $15K is the floor:
- Cheaper than $250/hour means a mid-level engineer or offshore contractor. The work product is worse on what matters — eval design, data understanding, model-class judgement. Writing a defensible eval set requires senior taste.
- Fewer than 60 hours means the eval set or capability probe gets cut. PRD and memo are quick; eval set and probe are the expensive line items.
- Less than two calendar weeks means the partner does not see real data. The probe degrades to a demo on a public dataset.
This is not a luxury number. It is the floor at which the five artifacts can actually exist.
What $15K is NOT
A $15K pilot is a procurement instrument, not a product. The non-deliverables are as load-bearing as the deliverables:
- Not an MVP. No user-facing UI beyond what is needed to demonstrate the capability. No auth, no billing, no production database.
- Not a beta. No real users. The capability probe runs on a sample of the founder’s data; it does not serve external traffic.
- Not a launch. The artifacts support a go/no-go decision. They do not support paying customers.
- Not a full PRD. The draft is v0.1 — a forced-structure 3-5 page document. A production PRD is 30-60 pages.
- Not a polished design. No Figma file. No design system. The pilot’s UI is a thin shell over the model call.
- Not a hosted demo. The handoff package lets the founder run the artifacts locally or in their own cloud. The partner does not commit to hosting.
When a founder reads a $15K pilot SOW and expects “a working AI MVP”, the partner has over-promised or the founder has misread the document. A defensible engagement starts with both sides agreeing on the non-goals before the first invoice.
What gets cut below $15K
A pilot priced under $15K is achievable. What gets cut is predictable, and the order is consistent across the boutique market:
- First to go: the eval set. Eval design takes senior time and produces no immediate-feeling output. Below ~$12K, the eval set shrinks from 50 items to “we ran some examples” or disappears. Anthropic’s Building effective agents treats the eval as the load-bearing artifact; cutting it removes the falsifiability.
- Second: the capability probe on real data. Below ~$10K, the partner runs the probe on a public dataset because data access takes calendar time the budget does not have. The probe still produces a working prototype, but it has not tested the founder’s data — where the real risk lives.
- Third: the PRD draft. Below ~$7K, the PRD reduces to a paragraph. No forced exercise of naming non-goals.
- Fourth: the go/no-go memo. Below ~$5K, the partner sends a recap email instead of a structured memo. The recap is sales material.
What remains at the $3K-$5K floor is a demo: working code, a public dataset, a recap email. The founder has spent money and has no defensible artifact. The hidden cost of cheap is procurement opacity.
A useful test: ask any partner quoting under $15K to list the five artifacts and effort allocation. If two or more are missing or vague, the pilot is structurally different.
The two-way handoff
A defensible $15K pilot is designed to hand off cleanly in two directions. This is what separates a real procurement instrument from a sales funnel.
Direction one: DIY exit. The founder takes the five artifacts and builds the MVP themselves using Cursor, Claude Code, or a similar AI development environment. The artifacts give them what they need: a PRD draft to scope from, an eval set to test against, a capability probe to start from. Roughly 25-35% of pilots end this way. The partner accepts this as a legitimate outcome. See can I build an AI app with Claude Code?.
Direction two: full engagement upgrade. The founder signs a six-to-twelve-week, $130K-$200K fixed-fee build, and the pilot’s artifacts fold directly into week one — the PRD draft becomes the engagement PRD, the eval set v1 becomes the engagement’s eval suite v1, the capability probe becomes the seed code, the memo becomes the risk register. The pilot fee is sometimes credited 50-100% against the first invoice; the SOW should say which.
A pilot designed only to upgrade is a sales funnel. A pilot designed to hand off in either direction is procurement. The test is whether the SOW names the DIY-exit option as a legitimate outcome. If it does not, part of the $15K is a customer-acquisition cost being charged to the founder.
The runway-relative frame
Pre-seed non-technical-founder teams in 2026 raise a median of $400K-$700K (Carta’s State of Private Markets 2025). At that runway, $15K is 2-4% of the round. The relevant comparison is not $15K vs $0 — it is $15K on a procurement instrument vs $150K on a full engagement with the wrong capability assumption.
A failed full engagement at the 60% mark — common when capability or data assumptions are wrong and untested — costs roughly $90K of sunk fee and 6-8 weeks of calendar. Spending $15K to test the assumption first is roughly a 6x return if it surfaces a “no” that would otherwise surface at the 60% mark. The math holds on a “yes” too: the artifacts fold into week one, the engagement starts 1-2 weeks ahead of schedule, the fee credits back. Net cost is at most $5K-$10K.
$15K is not the cheapest pilot. It is the cheapest defensible pilot. For broader cost math, see AI MVP cost comparison: idea-to-product service vs dev shop vs solo developer.
Three conditions move the price up legitimately: hard data access (HIPAA, SOC 2, VPC-deployed model calls) adds $5K-$8K; a genuinely novel capability adds $5K-$15K; a two-engineer pairing to compress calendar time adds $10K-$15K. Outside those, a $25K-$40K “pilot” is usually a small full engagement under a different label.
How to write the $15K pilot SOW
A defensible $15K pilot SOW is short — typically 3-5 pages — and names eight items in this order:
- The five artifacts with effort estimates per artifact (matching the table above).
- The capability hypothesis the pilot tests, as a falsifiable statement (“the model can extract X from Y at ≥75% accuracy on the eval set”).
- The launch threshold — the eval pass rate the founder will accept as evidence the full engagement is worth signing.
- The data access path — what data the founder will provide, in what format, by what date.
- The two handoff paths — what happens on DIY exit, what happens on engagement upgrade. The DIY path must be a legitimate exit, not a penalty case.
- The pilot-fee credit policy — typically 50-100% credited against the first engagement invoice if signed within 4 weeks.
- The non-goals — the explicit “this is not an MVP / not a beta / not a launch” list.
- The IP terms — the defensible answer: the founder owns all artifacts at delivery.
If the SOW is silent on any of these, ask for it in writing before signing. Silence is not an oversight; it is a procurement risk. For the deeper rubric, see the founder-friendly AI partner checklist: 11 must-haves.
FAQ
Why is $15K the floor and not $10K?
Because below $15K, the senior engineer time required to build a 50-item eval set on the founder’s real data does not fit. The $250-$300/hour senior rate times 60 billable hours equals roughly $15,000. That is the unit-economics floor in the US contract market in 2026.
Can I get a defensible pilot for less in an offshore market?
Sometimes. Offshore senior AI engineers can be 30-50% cheaper, moving the floor toward $9K-$11K. The trade-offs are time-zone friction in a two-week sprint and the harder-to-verify question of whether the engineer has seen enough US-founder context to write a useful eval set. The math usually requires a founder who can review the work substantively.
What if the partner offers a “free pilot”?
A free pilot is a sales demo with no procurement teeth. The partner spends 5-10 hours on a slide deck and a hardcoded prompt to win the full engagement. There is no real capability probe, no eval set, no falsifiable artifact. Treat a free pilot as the partner’s customer-acquisition cost, not as evidence of capability.
How much of the $15K is credited against the full engagement?
Defensible partners credit 50-100% of the pilot fee against the first engagement invoice if the engagement is signed within four weeks. Some credit 100% if the engagement value exceeds $100K. The credit terms should be in the SOW; if not written down, assume 0%.
What if the pilot fails the eval threshold?
That is the point of the pilot. Two defensible responses: the partner scopes the eval-failure work into the full engagement with a higher fee and a revised threshold, or both sides walk away cleanly — the founder has the artifacts, the partner has the fee, the failure is the procurement evidence. The wrong outcome is declaring the pilot a success despite the gap.
Is a $15K pilot enough to validate a B2B SaaS AI feature?
Yes for capability and data feasibility — the two biggest pre-build risks. No for go-to-market validation, pricing, or user adoption. The pilot answers “can we build this?” and “will it work on our data?”. It does not answer “will users pay?”. For the latter, you need a paying-customer letter-of-intent.
How does a $15K pilot differ from an “AI strategy sprint”?
A strategy sprint produces a slide deck and a roadmap — no code, no eval set, no capability probe. It is consulting, priced at $20K-$50K. A $15K pilot produces a runnable prototype and a measured eval baseline on the founder’s data. Sprints suit organisations choosing among 3-10 AI bets; pilots suit founders with a single bet to test.
Why not run two $15K pilots in parallel?
Pilots are short and intense; running two in parallel halves the partner’s attention on each. The defensible pattern is one pilot at a time, with a back-pocket second-choice partner pre-qualified through discovery calls. If the first pilot fails, the second runs in the following two weeks.
What is the SFAI Labs default pilot price?
$15,000 for a two-week pilot covering the five artifacts. The fee credits 100% against the first engagement invoice if signed within four weeks. The DIY exit is a legitimate outcome in the SOW.
Ready to scope a defensible $15K pilot? A 30-minute discovery call with SFAI Labs produces a scoping memo within 48 hours and a $15K pilot SOW naming the five artifacts. Book a discovery call.
Arthur Wandzel
