The contract models used to buy AI engineering work, what each one actually rewards, and how to stop paying to build a platform that already exists
On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so you can deploy anywhere, from your own cloud to a fully air-gapped network.
Last updated:
Almost every AI infrastructure engagement is bought one of four ways: time-and-materials, fixed-price, labor-hour, or cost-reimbursement. Which one you pick changes who carries risk, what the vendor is rewarded for, and whether the programme has a ceiling.
Time-and-materials is the default, and the reason is honest: nobody can accurately size a system that has not been designed yet. FAR 16.601 permits T&M in federal contracting only when it is not possible to accurately estimate the extent or duration of the work, and states plainly that T&M provides no positive profit incentive for cost control or labor efficiency.
That is the sentence to sit with. The contract model is not the problem — it is a symptom of how much remains unbuilt when the engagement starts.
This guide covers the four models, where each genuinely fits, how to write a statement of work that bounds AI scope, and the option most comparisons omit: licensing a platform that already exists so the only variable left is integration.
You should be able to state what the AI system must accomplish and for whom. You do not need to have designed it — that is what makes the estimate hard, and it is the thing to fix.
The integration surface is the estimable part of AI work. List the SIS, EHR, CRM, ticketing, identity provider, and document stores the system must reach.
Whether data may be processed by a third party changes which vendors and which architectures are viable, and it is far cheaper to settle before an RFP than after.
Contract-model decisions are usually made by whoever must defend the spend in year two. Involve them before the model is chosen, not after the overrun.
Most AI scope is not specific to you. Retrieval, evaluation, guardrails, access control, audit logging, and model routing are the same in every organization. Your data model, workflows, and integrations are not. Splitting these two makes the estimate tractable.
Usually a small minority of the technical scope.
If a competitor would need the same thing, it is undifferentiated.
This single question separates platform vendors from services firms.
Pick the model that matches the actual uncertainty rather than the one that feels safest. Fixed-price on unknowable work buys a risk premium and change-order friction; T&M on knowable work gives away a ceiling for nothing.
Best when the platform exists and only integration is being priced.
If you cannot describe the deliverable, this is the honest structure.
The cost of the platform is known because someone else already incurred it.
AI statements of work fail in predictable ways: they specify outputs that depend on data quality nobody has assessed, and they leave evaluation undefined so nobody can say whether the work is done.
Agree the test data and the passing threshold before work begins.
Named systems, named APIs, named identity provider.
Delivery does not imply ownership. This is the clause most often discovered too late.
Rate cards and licences are not comparable line items. Normalise every bid to the same thing: what it costs to reach production, plus three years of running and maintaining it, plus what you hold at the end.
The controls that keep AI engagements from drifting are agreed at the beginning or not at all. All of them are easier when the platform underneath is finished.
Open-ended billing exists because the work is unestimable, and the work is unestimable because the platform is being built from scratch. Reduce what is unbuilt and every contract model prices better.
Retrieval, evaluation, guardrails, access control and audit are identical across organizations. Funding their construction is the single most common way AI budgets are consumed without producing differentiation.
Buying hours does not confer the right to operate or modify what was produced. This is governed by clauses that are routinely under-negotiated and discovered years later.
Every model release and protocol change is a funded backlog item for a bespoke system. Over three years this frequently exceeds the original build cost.
If data cannot leave your perimeter, no amount of contractual assurance substitutes. Decide this before the RFP, because it eliminates whole categories of bid.
Categorise the SOW's tasks into undifferentiated platform work versus your data model, workflows and integrations
Compare actual cost at first production workload against the signed estimate
Date from contract signature to first workload serving real users
Run an acceptance test in which internal engineers ship a modification unaided
Consequence: A 10% rate reduction is erased by a 12% overrun, which is well within the normal range for AI work.
Prevention: Bound the quantity of work by reducing what has to be built, then compare totals.
Consequence: The vendor prices the uncertainty in and defends the scope, so every discovery becomes a change order.
Prevention: Use fixed-price where the deliverable is genuinely stable — which it is once the platform already exists.
Consequence: A working system the organization cannot legally modify or operate independently.
Prevention: Make rights an explicit, negotiated term of the statement of work.
Consequence: Nobody can say whether the AI system works, so acceptance becomes a negotiation.
Prevention: Agree an evaluation set and a passing threshold from your own data before work starts.
ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.
Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
See how ibl.ai deploys AI agents you own and control—on your infrastructure, integrated with your systems.