Many AI proposals describe the budget more clearly than the work. The seller knows what they’ll build; the buyer often doesn’t, because “an AI integration” can mean many systems. Projects go wrong in that gap, before any code exists.
We scope and build AI projects through our AI integration service, so weigh that. Every fact below about Anthropic’s models, prices and terms, and about copyright and the EU AI Act, was checked against the official pages themselves in October 2026.
Short answer: Negotiate the definition of the work before the price. Get in writing: the outcome, the data the system touches, where a person decides, the exact model, a test set for acceptance, what counts as a change order, and an exit after each phase. Then cover what’s specific to AI: model retirement, whose API account it runs on, and who owns the outputs. Have a lawyer review the contract; this isn’t legal advice.
What the Official Sources Say
As of October 2026, from Anthropic’s documentation and terms, the US Copyright Office, UK legislation and the European Commission:
| Topic | What the source says |
|---|---|
| Model retirement | At least 60 days’ notice; calls to a retired model fail |
| Model IDs | Every Claude model ID is a pinned snapshot |
| Prices | Claude Sonnet 5.5: $2 input, $10 output per million tokens |
| Tokenizer | Claude 4.7 and later: about 30% more tokens for the same text |
| Spend caps | Standard API tiers have a monthly cap; usage pauses at it |
| Outputs | The Customer owns its Outputs; no training on Customer Content |
| US copyright | Prompts alone don’t make AI output protectable |
| UK copyright | A computer-generated work’s author is whoever made the arrangements |
| EU AI Act | Chatbot transparency rules apply from August 2026 |
Sources: Anthropic model deprecations, Claude models overview, Anthropic pricing, Claude API rate and spend limits, Anthropic Commercial Terms, US Copyright Office, AI report Part 2, UK Copyright, Designs and Patents Act 1988, section 9, European Commission, AI Act.
Why AI Scope Is Harder to Pin Down Than a Normal Web Project
A website has a finish line most clients can picture. AI projects rarely do: one request can mean very different systems, and behaviour can shift when the model underneath changes.
The “Chatbot” Problem
Ask several suppliers to quote “a chatbot” and the quotes can differ a lot, because they aren’t quoting the same thing. One matches questions to written answers. Another reads your documents, remembers the conversation, writes to your CRM and hands over to a person when unsure. Same word, different systems. Ask any proposal to list what the system reads, writes, does when it doesn’t know and who it hands over to.
The Gap Between Pilot and Production
A pilot shows something works on clean examples. Production means it handles your messy real inputs every day, with logging, error handling and someone watching. If a proposal prices only the pilot, get production written down as a separate phase with its own deliverables.
Before You Negotiate: Know What You’re Actually Buying
You can’t negotiate a scope you haven’t defined. Leave the supplier to define it and it will fit what they already build.
Three Decisions That Define Every AI Project
Before any negotiation, get three things in writing:
- The business outcome. Not “an AI tool”, but a change you can measure, such as less staff time per invoice, against a baseline you record before the project starts.
- The approved data sources. Which systems the AI reads from, which it writes to and who controls access. If nobody has asked about these, scoping hasn’t started.
- What human action follows the output. An AI that flags anomalies is a different system from one that acts on them. Where the AI stops and a person decides is one of the biggest design choices in any integration.
How to Define “Done” Before Money Changes Hands
Ask for written acceptance criteria before you sign. “System is live” isn’t a condition. “Processes [X]% of the agreed test set without manual correction, and holds that rate on live inputs over [N] days” is. You set X and N together.
A rate alone can hide the cases that matter most, such as high-value invoices or exceptions. So agree a test set of real examples, including the hard ones. Our post on how to define done for an AI integration goes deeper.
The Three Red Flags in Any AI Proposal
Any one of these deserves questions. All three together is a reason to step back.
A Price Quoted Before a Deliverable Is Named
If a proposal leads with a total and describes what it buys later, vaguely, the number is an anchor, not a scope. Ask for each deliverable by name, then its price.
A Tool Recommendation Before Any Workflow Review
Naming a model before mapping your workflow fits you to an existing solution. Tool choice should follow requirements.
Once the model is chosen, the contract should name it exactly. Every Claude model ID is a pinned snapshot (source), so the SOW can say which ID the system runs on. That makes model changes visible.
No Exit Clause in the Engagement Letter
Every AI engagement needs a point where you can stop without paying for undelivered work. If ending early means paying all remaining fees, you’re committed however the work goes. Ask for a phased exit, with fees tied to accepted deliverables rather than the calendar.
Specific Contract Language to Demand (and Language to Strike)
The proposal sells; the contract binds. These clauses are a starting point for your lawyer, not a replacement for one.
Performance Baselines and Acceptance Criteria, What to Require
Require a clause that names the baseline before work starts and the threshold for acceptance. For example: “The system is accepted when it meets the criteria in Appendix A, on the agreed test set and then over the agreed period of live operation.” Appendix A holds metrics such as accuracy and error rate, agreed before signing.
The “Vendor May Adjust Deliverables” Clause, Strike It Every Time
Some contracts include wording like: “Vendor reserves the right to adjust deliverables, timelines, or technical approach as project requirements evolve.” It lets one side change what’s being built. Replace it with: “Any change to deliverables, technical approach, or timeline requires written agreement from both parties before it is implemented.”
Change Order Triggers, What Counts as Scope Change, What Doesn’t
Put a list of examples in the contract:
- Scope change: a new data source, integration endpoint, user role or output format.
- Not a scope change: fixing defects in delivered work, missing the agreed acceptance criteria, clarifying requirements already agreed.
- Needs its own clause: a model migration forced by the provider.
Model Retirement and Running Costs
Anthropic gives at least 60 days’ notice before retiring a publicly released model, and requests to a retired model fail (source). Claude Sonnet 4.5, for example, was deprecated on 30 September 2026 and retires on 30 November 2026. Amazon Bedrock and Google Cloud set their own dates. So the contract should say who migrates to the replacement, how soon and who pays. See what happens to your integration when the model changes.
The API bills per token: Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens (source). A model change can move that cost, because Claude 4.7 and later models produce about 30% more tokens for the same text (source). Standard API tiers also have monthly spend caps, and usage pauses at the cap (source). Agree whose account it runs on and who watches usage.
Ownership of Code, Prompts and Outputs
Name who owns the code, prompts, test sets, documentation and outputs, and when they transfer. Under Anthropic’s Commercial Terms, the Customer owns its Outputs, and Anthropic may not train models on Customer Content (source). The Customer is whoever accepted those terms. If the supplier holds the API account, your contract has to pass the rights to you.
Copyright adds a layer. The US Copyright Office concluded that AI output is protected only where a human determined enough of its expressive elements, and that prompts alone aren’t enough (source). UK law says the author of a computer-generated work is the person who made the arrangements needed to create it (source). Ask a lawyer what that means for your outputs.
Compliance Duties Nobody Owns by Default
The European Commission says the AI Act’s transparency rules, such as telling people they’re talking to a machine, take effect in August 2026, and rules for high-risk uses like employment apply from 2 December 2027 (source). If the system serves people in the EU, name who builds the disclosure and who checks the risk category.
Phased Contracts: Protection or Trap?
Phased engagements are often pitched as the low-risk start: discovery first, then build. Whether they are depends on what’s written into the move between phases.
When Phased Scoping Is Legitimate
Discovery makes sense when there’s real uncertainty: an undocumented workflow, data quality to audit, or an integration whose feasibility needs testing. What makes it legitimate is the deliverable. Good discovery produces a scope document, architecture map or data audit that any qualified supplier could build from. If the only output is a presentation just the seller can act on, you’ve paid for a sales process.
How to Negotiate Phase-Exit Conditions
Every phase needs a measurable exit condition and a decision point. For example: “At the end of Phase 1, the client receives [specific deliverable]. The client may proceed to Phase 2, commission Phase 2 from a different supplier, or stop. Proceeding to Phase 2 with this supplier requires a separate signed SOW.”
That gives you a real choice at each step. Each phase also needs its own acceptance criteria, to catch weak work inside it. To test before committing, see how to run an AI pilot without committing to a full build.
Related Posts
- How to Define Done for an AI Integration Project: acceptance criteria you can test
- Ownership vs Dependency: What Your AI Integration Contract Should Say: code, keys and handover
- AI Model Updates Break Integrations: What Businesses Miss: planning for model retirement
- How to Run an AI Pilot Without Committing to a Full Build: testing before the big contract
Frequently Asked Questions
What should an AI project SOW include that a regular web project SOW wouldn’t?
A data source inventory, acceptance criteria tied to a test set, the exact model, and what happens when that model is retired. Anthropic gives at least 60 days’ notice, then requests to the retired model fail, so agree who migrates and who pays.
How do I evaluate whether a paid discovery phase is legitimate or just a revenue grab?
Ask what the deliverable is and who can act on it. Legitimate discovery produces a scope document, architecture map or data audit any qualified supplier could use. If it’s “a strategic recommendation”, you’re paying for a sales process.
What’s a fair change order threshold for AI projects?
Define change orders by type, not amount. A new endpoint, data source, user role or output format is a scope change. Fixing defects, missing agreed acceptance criteria and clarifying agreed requirements are not.
How do I define success metrics for an AI integration before the project starts?
Work back from the business outcome. For invoice processing: time per invoice, error rate and the share processed without correction. Record today’s baseline, build a test set of real examples, and agree both in writing.
What exit clause language should I require in an AI consulting contract?
The right to end on written notice, with payment limited to work completed and accepted under the SOW, not projected fees. On termination, all code, prompts, test sets and documentation transfer to you within an agreed number of business days. Have a lawyer review the wording for your jurisdiction.
Who should hold the API account for an AI project?
Usually you. Under Anthropic’s Commercial Terms, whoever accepts the terms is the Customer, and the Customer owns the Outputs. You also see usage and keep access if you change supplier.
If a proposal lacks a named deliverable, a baseline and an exit clause, it isn’t a scope document yet.