Contents
Key takeawaysThe five categoriesScoping use casesThe four data termsCommercial terms by categorySizing AI seatsAI add on and suite renewalWhat we have seenClauses to requestWhat to do nextFAQEnterprise AI spend falls into five categories, each priced in a different unit. Scope use cases first, close the data terms before price, license seats on a ramp, and trade commitment for unit price on a short price term.
- Classify before you draft. Model platforms, SaaS embeds, agent platforms, fine tuning and internal hosting each need their own contract terms.
- Use cases come before vendors. Document three use cases with adoption targets and data perimeters six to nine months before signature.
- Data terms come before price. The training carve out, retention exception, output ownership and IP indemnity belong in the order form first.
- Buy seats on a ramp. Most licensed AI seats sit idle at ninety days, so each step up should depend on measured use.
- Keep the AI line inside the renewal. An AI add on negotiated on its own gives away discount the suite renewal would have captured.
- Commit long, price short. Lock unit price for twelve months at most while AI prices keep falling, and hold an exit at every renewal.
What does enterprise AI procurement actually cover?
Enterprise AI procurement covers five categories of spend that usually share one budget line: hyperscaler model platforms, AI features embedded in SaaS suites, agent platforms, fine tuning and training programs, and internal model hosting. Each is sold in a different commercial unit, with different data terms and a different exit path.
| Category | Commercial unit | Typical share of AI budget | What gets negotiated |
|---|---|---|---|
| Hyperscaler model platforms | Token usage, provisioned throughput units (PTU), committed spend | 40 to 60 percent | Rate cards, reservation discounts, committed spend credits |
| SaaS embeds | Per user per month | 20 to 35 percent | True up cadence, deactivation lag, model substitution |
| Agent platforms | Per agent action or license | 5 to 15 percent | Action rate cards, idle session billing, audit trail |
| Fine tuning and training | GPU hour, training tokens | 5 to 15 percent | Output rights, residency during training, deprecation horizon |
| Internal model hosting | GPU hour, instance hour | 5 to 20 percent | Open weight license terms, GPU procurement, support model |
Why does one contract template fail across all five?
The failure we see most often is a single template drafted for one category and applied to all five. A per user per month paper with a three year term and a 10 percent annual uplift cap suits a SaaS embed. On a hyperscaler platform deal, priced in tokens, those clauses have nothing to attach to.
The right instrument for the platform deal is a twelve month committed spend with a unit price reset clause. So decide the category before anyone drafts paper, starting with a pass through the last twelve months of cloud bills and suite order forms. Our AI platform TCO comparison sets out the platform economics side by side.
What You Are Actually Buying
How should you scope AI use cases before talking to vendors?
Start six to nine months before signature and spend the first ninety days fixing the use cases, the success metrics and the data perimeter. Vendor shortlists come after that. We ask clients to document three use cases before selection opens, each with an adoption target, a measurable outcome and a defined data perimeter.
| Use case type | Typical work | Adoption target | How success is measured |
|---|---|---|---|
| Volume | High frequency, low complexity: drafting, summarization, search | Above 60 percent of the target user base | Time saved per user per week |
| Judgment | Medium frequency, high impact: analysis, recommendation, code generation | Above 30 percent | An outcome metric, since usage alone proves little here |
| Differentiator | Low frequency, high differentiation: customer facing agents, regulatory drafting | Above 10 percent | Revenue impact or risk reduction |
Why do adoption targets belong in the contract?
The targets become the evidence gates for seat growth. If the volume use case needs 60 percent adoption to justify full deployment, the contract can require that level of measured use before the next tranche is billed. Without written targets, the vendor's rollout plan tends to set the seat count.
The data perimeter works the same way. A use case that touches customer records needs different retention and residency terms than one that drafts internal email, and you want that settled before the vendor sends its standard paper.
Enterprise AI procurement strategy brief
Sourcing, contracting and renewal across the five AI spend categories, in one download.
Get the white paper →Which AI data terms need to close before price?
Four terms decide whether an AI contract is safe to sign: a training carve out, an exception to abuse monitoring retention, output ownership and IP indemnification. Write all four into the order form before discount talks open, while the vendor still wants the deal more than you do.
| Data term | Default vendor position | What to ask for |
|---|---|---|
| Training carve out | Customer data may be used for model improvement | No use of customer data, prompts or completions for foundation or shared model training |
| Abuse monitoring retention | 30 day default retention of prompts and completions | Zero retention, or modified monitoring with documented eligibility |
| Output ownership | Customer owns outputs, vendor retains a broad license | Customer owns outputs, with the vendor license limited to operating the service |
| Indemnification | Limited, customer responsible for output | Vendor indemnification on third party IP claims arising from foundation model output |
What do the major AI vendors offer by default?
As of 2026, the largest model vendors publish enterprise defaults close to what you would ask for, which makes the four terms easier to get in writing. Each comes with conditions.
- OpenAI. It states that it does not train on ChatGPT Enterprise or API data by default and that customers own inputs and outputs. API data is kept for up to 30 days to identify abuse, and zero data retention is limited to eligible endpoints and qualifying use cases.
- Azure OpenAI. Prompts and completions are not used to train foundation models without permission. Removing stored data from human abuse review requires an application for modified abuse monitoring, and approval is not automatic.
- Microsoft Customer Copyright Commitment. The IP defense for Azure OpenAI output holds only if you ran the required mitigations, including a system message against reproducing copyrighted material, documented testing, and the protected material and Prompt Shields filters.
- Google Cloud. Output indemnity covers the services on Google's published list of Generative AI Indemnified Services, so check each model you plan to use against it.
These are policy pages the vendor can revise. Copy the position you rely on into the agreement, with a clause that later policy changes cannot reduce it during the term. Our guide to AI data governance and IP terms has the clause detail.
Where does data residency fit?
Residency belongs in the same conversation. Document the sovereign options per region instead of assuming them, including where data sits during fine tuning. On Azure, provisioned throughput reservations are bought separately for Global, Data Zone and Regional deployments and cannot be swapped, so the residency choice also fixes which commitment you can buy.
How do AI commercial terms differ by category?
Each category has its own unit, discount mechanism and standard trade. Applying the wrong one is how buyers overpay on three categories while under governing the other two.
| Category | Unit priced | Where the discount comes from | Usual trade | Unit price reduction |
|---|---|---|---|---|
| Hyperscaler platforms | Input and output tokens, plus reserved throughput for predictable workloads | Committed spend through the cloud commitment vehicle | 12 to 36 month spend commitment, with credit pool flexibility across services | 25 to 45 percent |
| SaaS embeds | Per user per month | Enterprise agreement discount, volume tier, no deactivation lag, true up cadence | Commitment with a ramp profile tied to adoption | 10 to 25 percent |
| Agent platforms | Per agent action or per agent license | Action volume commitment, license bundling, idle session policy | Six to twelve month pilot, then a 24 month commitment | 30 to 50 percent |
Fine tuning and internal hosting are billed in training tokens or GPU hours, usually through your model platform or cloud contract. The terms that matter there are output rights on tuned models, residency during training and the base model's deprecation horizon. For the Microsoft stack, see the Copilot pricing guide for seats and the Azure OpenAI guide for the platform.
What should a spend commitment buy?
A commitment should buy a lower unit price. Traded for extra licenses or bonus credits, it gives the vendor volume and leaves your cost per token, seat or action where it was.
Keep the price term short. AI vendor pricing has been falling 20 to 40 percent a year, so lock unit price for twelve months at most and take a longer term only on the minimum commitment.
- Amazon Bedrock. Provisioned throughput is sold with no commitment, one month or six months, and the longer term earns the lower hourly rate.
- Azure OpenAI. Provisioned throughput reservations run for one month or one year, so even Microsoft's reserved capacity is not priced over three years.
How many AI seats should you license?
License against measured adoption, on a ramp. In our file, weekly active use settled at 35 to 45 percent of licensed seats after ninety days, so a full population purchase pays for a majority of idle seats. On a SaaS embed, the seat count changes total cost more than any discount point.
The ramp we ask for is typically 50 percent of the target population in year one, 75 percent in year two and 100 percent in year three. Each step should depend on usage evidence, such as an agreed active user rate, and never arrive automatically on the anniversary date.
What does a seat ramp save in a worked example?
Say you plan an AI assistant for 4,000 employees at a hypothetical $30 per user per month, and the vendor offers 10 percent off for licensing everyone from day one.
| Option | Year 1 seats | Year 2 seats | Year 3 seats | Three year cost |
|---|---|---|---|---|
| Full population at $27 (10 percent off) | 4,000 | 4,000 | 4,000 | $3,888,000 |
| Ramp at $30 | 2,000 | 3,000 | 4,000 | $3,240,000 |
| Ramp at $30, year three gate not met | 2,000 | 3,000 | 3,000 | $2,880,000 |
The ramp costs $648,000 less over three years at the undiscounted price, and $1,008,000 less when adoption does not justify the final step. The discounted offer comes out ahead only if close to all 4,000 employees use the tool from the first year. Once the ramp is agreed, ask for the discount on top of it.
How do you check real usage before sizing the order?
- Microsoft 365 Copilot. In the Microsoft 365 admin center, open Reports, Usage, then Microsoft Copilot. It shows enabled users, active users and the active users rate over 7, 28, 90 or 180 days.
- Other SaaS embeds. Ask for a per user last activity export during the pilot, and write the right to that report into the order form.
- Model platforms. Azure Cost Management and AWS Cost Explorer show model spend by meter or usage type, and by tag. Tag each deployment with its use case so spend maps back to the business case.
Our note on Copilot monthly active users explains how to read the Microsoft figures before a true up.
Should you negotiate the AI add on with the suite renewal?
Yes. Buyers who negotiated the AI add on separately from the suite renewal gave up 10 to 25 percent of available discount. The vendor wants the AI attach and will pay for it in the terms of the agreement it attaches to, so the AI line is worth far less negotiated alone.
Combining the two conversations costs close to nothing, and splitting them is the most common procedural mistake we see in this category. If the account team offers the add on early, keep it open and fold it into the renewal timetable.
What will the account team say, and how should you answer?
- "This AI pricing is only available this quarter." Ask for the offer in writing with its expiry and say you will evaluate it with the suite renewal.
- "Licensing everyone earns the best discount." Send your seat model and usage gates, and ask them to price the ramp.
- "Our standard terms already cover data use." Then they can go into the order form, with the policy version date and protection against later changes.
- "A three year price lock protects you." With AI prices falling, it protects the vendor's revenue. Offer a longer minimum commitment for a twelve month price term that resets each year.
What have we seen in recent enterprise AI procurements?
Across roughly 25 to 35 enterprise AI procurements we ran between 2024 and 2025, seat usage and data terms decided value far more than the per seat price did. Three patterns recurred.
- Idle seats. Weekly active use settled well below the licensed count within ninety days, at the level described in the seat sizing section.
- Separate AI deals. Buyers who ran the AI add on as its own negotiation left discount behind that a combined renewal would have captured.
- Default data terms. Default data residency and training boundaries were accepted in most first drafts, then reopened later at material cost.
The buyers who did well worked in a fixed order: use cases first, then data terms, then price and term length.
Why we advise against consolidating on one AI vendor
A common recommendation is to standardize on one AI platform for simplicity and a larger volume discount. We advise against it for now. The platform layer is new and the market has repriced every year, so a single vendor commitment is a bet that today's choice keeps both the best models and a competitive price.
Run one hyperscaler platform, one SaaS embed and one open weight model in parallel. The open weight model gives you a working alternative to price against. Our GenAI vendor lock in assessment measures the switching cost you carry today.
A default data term accepted in the first draft costs far more to reopen after signature than to refuse before price talks start.
What should an enterprise AI contract include?
Beyond the four data terms, ask for the commercial and exit clauses below, and make the exits available at every renewal point. A right that applies only at term end arrives after the market has already repriced.
| Clause | Why it matters |
|---|---|
| Unit price reset | The rate follows falling market prices after the first twelve months. |
| Seat ramp with usage gates | Seats grow with measured adoption, never with the calendar alone. |
| No deactivation lag, flexible true up | Billing stops the day a seat is removed, and counts can fall as well as rise. |
| Action and idle session definitions | Agent platforms bill per action, so define what counts. See our agent metering redline. |
| Model deprecation rights | If the vendor retires the model you built on, you can exit or move committed spend. |
| Termination for convenience | A defined exit at renewal points, covered in our termination for convenience guide. |
| Data export and portability | Export in plain text formats, plus prompt and completion portability, so workloads can move. |
What should you ask the vendor before you sign?
- Which of our data, prompts and completions do you store, where, and for how long?
- Which models in this order carry your IP indemnity, and what must we configure to keep it?
- How much notice do you give before retiring a model, and can committed spend move to its replacement?
- Which usage report will we receive, and is it written into the order form?
Our checklist of 20 questions before signing covers the rest.
What to do next
- Classify the spend. Tag the last twelve months of AI invoices into the five categories before anyone drafts contract paper.
- Scope three use cases. Start six to nine months before signature, with an adoption target, a success metric and a data perimeter for each.
- Send your data terms first. Put all four into the draft order before discount talks open.
- Size seats on a ramp. Use pilot activity data, gate each step on usage, and trade commitment for unit price, never for extra licenses.
- Combine the negotiations. Settle the AI line inside the suite renewal and hold the unit price term to twelve months.
- Write in the exits. Add deprecation rights, termination for convenience and data portability. Our GenAI advisory practice can run the procurement with you.
Frequently asked questions
Why treat enterprise AI as five categories instead of one?
Because the unit you pay for changes the terms that protect you. Tokens, seats, agent actions and GPU hours each need their own price clauses, data terms and exits. A seat template applied to a token deal has no clause that touches the real cost driver, so classify the spend before drafting anything.
How many licensed AI seats actually get used?
Fewer than most business cases assume. In the procurements we ran, 35 to 45 percent of licensed seats were in weekly use ninety days after rollout. Check your own pilot data in the vendor's admin reports and budget for that level until your figures prove otherwise.
Which AI data terms need to be agreed before price?
Four: no training on your data, prompts or completions; an exception to the default 30 day abuse monitoring retention; output ownership with the vendor's license limited to running the service; and indemnity for third party IP claims over model output. Agreeing them first stops the vendor from charging for them later.
What should an AI spend commitment buy?
A lower unit price. On hyperscaler platforms, a 12 to 36 month commitment typically earns 25 to 45 percent off the unit rate, with the credit pool usable across services. On agent platforms, a pilot of six to twelve months followed by a 24 month commitment reached 30 to 50 percent.
How long should an AI price term run?
Twelve months or less on unit price. Vendor pricing for AI has dropped 20 to 40 percent a year, so a three year rate lock usually leaves you above market by year two. A longer term is acceptable for the minimum commitment, provided the rate resets each year.
Should the AI add on be negotiated separately from the suite?
No. In our work, buyers who split them gave up 10 to 25 percent of available discount. Tell the account team early that AI pricing will be settled inside the renewal, so any early AI quote feeds that negotiation.
What exit rights should an AI contract carry?
Termination for convenience or model deprecation rights, data export in plain text formats, and portability of prompts and completions, available at each renewal point. Pair them with a portfolio of one hyperscaler platform, one SaaS embed and one open weight model, so an exit is something you can actually use.