The fixed price vs. time and materials debate is older than software itself. AI engineering forces an honest answer to it. The probabilistic nature of LLMs, the open-endedness of evaluation, and the rapid pace of model releases mean that traditional fixed-price contracts often misprice the work, and traditional T&M arrangements often misalign incentives.
What's different about AI engineering pricing
Three things make AI work harder to price than conventional software. First, the unit of progress is an evaluation pass, not a feature ticket. Second, the underlying tools change every quarter — Anthropic released three major Claude versions in 12 months in 2024–2025, each changing the cost-performance tradeoffs significantly. Third, cost structure shifts: token spend, vector store fees, and inference latency budgets are real line items that don't exist in a CRUD app. Gartner estimates that AI infrastructure costs account for 35–55% of total AI project spend in production, a factor rarely included in traditional fixed-price estimates.
When fixed price works for AI projects
Fixed price is appropriate when scope is genuinely fixed: a contained AI feature, an integration with a defined API, or a discovery with a known set of deliverables. Our AI Discovery Mission is fixed price for exactly this reason, the artefacts are well-defined and the timebox is short.
- A scoped integration: 'Add LLM-powered search to the help centre'
- A discovery: 'Four-week AI discovery, eight named deliverables'
- A migration: 'Move classification from rules to model X with a defined eval pass-rate'
- A capped spike: 'Two-week feasibility build, fixed budget, single deliverable'
When T&M is the honest answer
T&M is the right model for open-ended product work, multi-agent systems, and long-running platform builds. The reality is that nobody can fix-price 'build us an agentic copilot' without padding it 2–3x to absorb risk, which neither side wants.
Fixed-price contracts work well for well-defined deliverables but poorly for AI work where scope evolves. The Standish Group's CHAOS Report 2024 found that 66% of software projects with fixed budgets experienced scope changes — for AI projects, that figure is significantly higher.
If a vendor offers you a fixed price for an unscoped agentic build, two things are true: they have priced in the risk, and they will protect that price by saying no to the changes you need.
The hybrid we recommend
We default to a hybrid: fixed-price discovery, fixed-price MVP with a defined eval pass-rate, then T&M with a monthly cap for the iteration phase. This aligns incentives at the points that matter, getting to a working baseline, and keeps flexibility where the work is genuinely uncertain.
- Phase 1: Fixed-price discovery (4 weeks)
- Phase 2: Fixed-price MVP keyed to an eval pass-rate (6–8 weeks)
- Phase 3: T&M with a monthly spend cap and a quarterly review
- Throughout: Transparent token and infra costs, billed at-cost
The true cost of AI engineering includes more than just development hours — McKinsey estimates that infrastructure, tooling, evaluation, and ongoing model maintenance account for 40–60% of total cost of ownership for a production AI system.
The right pricing model is the one that lets the team make the right engineering decisions at every stage. For AI work, that almost always means a hybrid. A 2025 Deloitte survey of 300 enterprise AI buyers found that 61% of projects priced under a pure fixed-price model ended with scope disputes, compared to 18% under hybrid milestone-based arrangements.