The default question
Most agency relationships in software services are time and materials — an hourly rate, a weekly burn, a meter that runs until the client says stop. The default works because it's easy to start: nobody has to commit to a scope.
The default also has a structural problem in AI work: the thing you're scoping has high uncertainty about what's possible. The buyer doesn't know what's in scope. The vendor often doesn't either. T&M lets both parties duck the hard scoping conversation, and the project drifts.
Fixed-fee forces the conversation. That's most of the value.
Why fixed-fee works for AI engagements
Three reasons it fits AI work specifically:
1. The buyer can budget
A buyer with a $50K AI feature has $50K to spend, not "between $20K and $200K depending on how things go." Fixed-fee gives the buyer a number their CFO can sign off on. T&M proposals have a way of growing 60-100% past their initial estimate; we've seen this enough times to know it's not anomaly.
2. The vendor is forced to scope
A fixed fee means we have to know what we're shipping before we sign. That forces the discovery conversation that the AI build checklist is designed for. If we can't answer the 23 questions, we can't price it. If we can't price it, we don't take the work yet.
This is uncomfortable when budgets are tight. It's the right kind of uncomfortable.
3. Both parties are aligned on shipping
Under T&M, more time = more revenue. The vendor is rewarded for slow progress. Under fixed-fee, more time = thinner margin. The vendor is rewarded for shipping. That's the alignment we want.
We've turned down T&M-only engagements specifically because the misalignment makes the work less satisfying — there's a small voice in the back of the head suggesting taking the long way. Fixed-fee removes the voice.
How we structure the price
Three engagement shapes, three price ranges:
- Audit / scoping (1-2 weeks): $5K-$15K fixed. Output is a written report, an architecture proposal, and a pre-priced build engagement. About a quarter of audits don't lead to a build because the audit reveals the build shouldn't happen — we cover this in when NOT to build with AI.
- Build engagement (4-12 weeks): $25K-$150K fixed depending on scope. Outcome is shipped working software, eval suite, runbook, handoff.
- Retained advisory (ongoing): $4K-$12K/mo fixed retainer. We're on Slack, in Linear, in the design reviews. Output is shipped product the client's team builds with our judgment in the loop. This is the only ongoing model we offer.
No T&M tier. No "we'll see how it goes." If we can't fit the work into one of those three shapes, we have a different conversation about whether we're the right vendor.
What it costs us when we get it wrong
The honest part. Fixed-fee means we wear the over-runs. Three examples we've absorbed in the last 18 months:
- A migration scope was deeper than discovery surfaced. We took the L on ~3 weeks of unbilled work to ship at the price we'd quoted. Margin halved on that engagement.
- A client's "labeled training data" turned out to be unlabeled. We rebuilt the labeling pipeline as part of the engagement at our cost because it was foundational to the deliverable.
- A compliance constraint surfaced in week 3 that required re-architecting the storage layer. We re-designed and re-shipped without a change order.
None of these wiped out the engagement. All of them hurt margin. The pricing model means we have to be sharper on discovery to avoid them. That sharpness is a skill we get better at engagement-by-engagement; T&M would have masked the lessons.
When T&M is actually right
Two narrow cases:
- Genuinely undefined R&D. The work is "explore whether AI can do X" with no commitment to ship. T&M with a weekly cap fits — we've taken three of these in 18 months, all with experienced research-shaped clients.
- Embedded staff augmentation. The client wants a contractor on their team for a specific role for a specific duration. This isn't really an engagement; it's a contractor relationship and pricing should follow that pattern. We rarely do this.
Outside those two, fixed-fee.
The objection we hear most
Buyers sometimes say "but if you finish in three weeks instead of six, you're getting paid double." Yes. That's the structure working as designed. The buyer agreed to a price for an outcome, not for our hours. If we found a faster way, that's a feature of the relationship, not a bug.
The reverse holds too: if it takes us nine weeks instead of six, we eat the difference. The buyer pays for the result. Both directions of variance live with us. That's the trade.
The summary for buyers
If you're shopping AI engagement work and the proposal is hourly with no fixed scope, ask why. The answer may be reasonable (genuine R&D). It may also be that nobody has done the scoping work, and the meter is going to run longer than the proposal suggests.
For us, fixed-fee is the discipline that keeps engagements shipping and forces the scoping conversations that produce real outcomes. It's not the only honest model. It's the one that aligns with how we want to work.