AI development cost: how to scope, quote, and price an AI build

Why AI development cost is different from classic dev work

AI development cost behaves differently from normal software for three structural reasons: the work is part R&D so you cannot scope it accurately until you have seen the data, compute is part of the product so quality scales with spend, and every query re-runs the model so the marginal cost never trends to zero. Those three forces are why a flat fixed bid that worked for a CRUD app breaks down on an AI build, and why the headline number is best read as a range tied to a phase rather than a single price.

The first force is uncertainty. With AI you are often building something whose feasibility depends on data you have not inspected yet, so an upfront fixed bid prices a scope you cannot reliably specify. The second is the cost structure itself. a16z's analysis of AI economics argues that AI is "compute-bound": adding more compute directly produces a better product, so the marginal cost does not collapse toward zero the way it does for traditional software.¹ The third is ongoing inference. There is no build-once, sell-infinitely. Every transaction consumes GPU time, which is why AI gross margins sit below the SaaS norm. Bessemer's State of AI 2025 work puts AI-native gross margins around 50% to 65%, against the 70% to 85% that classic SaaS businesses enjoy.²

The spend is real and the failure rate is too, which is why getting the cost structure right matters. Gartner forecasts worldwide generative-AI spending to reach $644 billion in 2025, yet also expects at least 30% of generative-AI projects to be abandoned after the proof-of-concept stage by the end of 2025, frequently on escalating cost and unclear business value.⁶ The practical takeaway sits underneath all three forces: price the value and the run-cost as two separate things, and never bury recurring inference inside a one-time build fee. The rest of this guide turns that principle into models, a scoping method, and concrete billing structures. If you build AI for clients under your own brand, the same logic flows into how we run white-label development behind agencies.

The six pricing models and where each one fits

There are six pricing models for AI work: fixed-bid, time-and-materials, milestone or phased, retainer, value or outcome-based, and hybrid. No single one fits every engagement. For most AI builds the practical combination is phased pricing to contain uncertainty, with a hybrid base-fee-plus-usage structure layered on top, because roughly 92% of AI software companies now price with a usage component.³

Each model trades predictability against flexibility differently, and each handles the recurring run-cost differently. The comparison below maps where each one fits and how to handle inference inside it.

A few notes on the extremes. Fixed-bid is the riskiest for AI: directional industry figures put generative-AI overruns at 60% to 150% over budget when there are no hard scope gates, so reserve it for genuinely well-defined work and only after a discovery phase has shrunk the unknowns.⁴ Pure outcome pricing sits at the other end. It aligns price to value better than anything else, and Bessemer's playbook points to consumption, workflow, or outcome pricing for AI-enabled services, citing examples like a charge per completed legal document or Intercom's Fin at $0.99 per resolution.³ Yet pure outcome pricing stays rare in practice because clients want a predictable bill, which is exactly why the hybrid base-plus-usage shape has become the common ground.

Scoping and estimating under uncertainty

The reusable pattern is to refuse to bid the build before a paid discovery and POC reduces the unknowns. Spend real time in discovery to assess the data and define go or no-go KPIs, run a time-boxed proof of concept to prove feasibility, then quote the build phase from what you measured. Quoting the full build on day one prices a scope you do not yet understand.

Discovery earns its keep. Directional industry work suggests teams that invest at least a quarter of the POC timeline into discovery, meaning data assessment, a hypothesis framework, and go or no-go KPIs set before any code, see materially better outcomes. A focused POC then answers the feasibility question in roughly 8 to 12 weeks: does the technology work, are the data requirements understood, and does the business case justify the build. Keep the POC distinct from an MVP. A proof of concept is a feasibility test and not yet a product, and conflating the two is a frequent cause of AI projects stalling.

Budget for the cliff between pilot and production. Directional figures suggest hardening a pilot for production is its own major cost: teams often spend 60% to 80% of the production budget rewriting POC-grade code instead of extending it, once data pipelines, security hardening, and integration are in scope, so the build quote should come after the POC and never before it.⁴ The real cost drivers are rarely the model itself: they are data quality, integration complexity, accuracy requirements, and inference volume. Estimate those, and the representative ranges below are useful for sizing a conversation, never for fixing a price.

Billing the recurring AI run-cost

The central agency question is who pays for the model, API, and GPU spend after launch, and the margin-safe answer is the client, billed transparently. Three mechanisms do this: a pass-through cost line with an optional defined markup, a markup on metered usage, or a usage tier with a committed minimum and overage. All three push the variable inference cost back to the customer instead of leaving it on your books.

Pass-through is the simplest. You bill the actual model, API, or GPU spend as its own line, optionally with a stated markup, which keeps the agency off the hook for inference volatility. Metered usage resells tokens, calls, or inference at a margin. Usage tiers bundle an inference allowance into a price band and bill overage above it, which protects the client's predictability while capping your exposure. This is how the major platforms shipped 2026 pricing: Salesforce Agentforce, Intercom Fin, and ServiceNow Now Assist all moved cost-of-goods exposure off the vendor balance sheet and onto consumption.³

The representative ranges below help size a first conversation. Treat every figure as directional, because the real number moves with data quality, integration depth, and accuracy needs.

The build capability behind most of these engagements is the same: see AI application development for what production AI work actually involves.

Protecting margin over the life of the engagement

Margin protection on AI work comes down to five disciplines: pass inference COGS through instead of eating it, build unit economics from day one, tie price to a hard measurable result, gate scope phase by phase, and set price from client value rather than from your cost-plus. Each one addresses a specific way AI engagements lose money after the contract is signed.

Start with the run-cost, because that is the leak unique to AI. Every transaction has a variable cost, so a flat one-time or per-seat fee silently compresses margin as usage grows. Outcome, consumption, and hybrid models exist precisely to push that variable cost back to the customer, which is the structural lesson from both a16z and Bessemer.¹²

• Build unit economics from day one. If the math does not work at 10 customers, it will not work at 1,000. Fold founder, PM, and support time into COGS alongside the API bill.³
• Avoid soft-ROI positioning. Bessemer warns that soft ROI positioning kills willingness to pay, so tie the price to a hard, measurable outcome wherever the result can be measured.³
• Gate scope by phase. Re-price each phase as uncertainty resolves, which is the mechanism that contains the 60% to 150% overrun risk on generative-AI work.⁴
• Price from value, not cost-plus. Set the ceiling by the value the client gets, find the friction point where they hesitate, then step back one notch.

None of this requires exotic contracts. It requires separating the value from the run-cost, gating the work so a bad assumption surfaces in a $90k phase instead of a $900k build, and writing the inference billing into the statement of work before the first query is ever served.

Kanika Mathur

Head of Service Delivery, Resourcifi

Kanika Mathur runs service delivery at Resourcifi, where pricing an AI build is a weekly conversation with agency partners who resell our engineering under their own brand. She has watched flat one-time fees quietly turn profitable projects into break-even ones once inference volume climbed, and she wrote this guide to keep delivery teams out of that trap.

Resourcifi on LinkedIn →

Keep reading

Related guides worth your time

← →

Agency & white-label White-Label AI Services White label AI lets agencies resell custom AI builds under their own brand. Learn what can be white-labeled, how it works... Read guide → Hiring AI Engineering Team Structure AI engineering team structure explained: core AI team roles, AI engineer vs ML engineer, the minimum viable team, and how... Read guide → Hiring AI-ready development teams: what they are and how to vet one How to hire AI developers: what AI-ready means, a 6-point vetting checklist, and warning signs. Clutch 4.9-rated AI exper... Read guide → Hiring How to hire a dedicated development team: cost, process, red flags How to hire a dedicated development team: when it beats staff augmentation, what it costs, a step-by-step vetting checkli... Read guide → Models & sourcing Outsourcing to India Guide Learn how to outsource software development to India: engagement models, rate ranges by region, risk controls, and partne... Read guide → Models & sourcing Staff Augmentation Guide What is staff augmentation and when should you use it? This guide covers IT staff augmentation models, rates, and how to... Read guide → Models & sourcing Staff augmentation vs outsourcing Staff augmentation vs outsourcing: control, cost, IP, and risk compared. Find out which model fits your team, plus when a... Read guide → Agents & RAG Agentic RAG: When to Use It and How to Build It Agentic RAG explained: how it differs from naive and advanced RAG, the key patterns like corrective RAG and self-RAG, the... Read guide → Agents & RAG AI Agent for Fintech: Risk, Compliance, Ops, Customer AI agents in finance: fraud, AML, KYC and servicing use cases, how to build with money-movement guardrails and human appr... Read guide →