A framework drafted for seats collapses on a deal priced in tokens
Enterprise AI is not a software category. It is five categories rolled into one budget line, and hyperscaler model platforms, SaaS embeds, agent platforms, fine tuning programmes, and internal hosting each carry distinct commercial mechanics, data terms, and exit paths. One discipline holds across all five: define the use case before the vendor, lock the data terms before the price, trade commitment for unit price rather than volume, and hold an exit lever at every renewal.
Prepared by Redress Compliance · August 10, 2026 · GenAI advisory. Based on 25 to 35 enterprise AI procurements, 2024 to 2025.
Executive summary
Weekly active use settled at 35 to 45 percent of licensed seats after ninety days. Most seats sat idle, which makes seat sizing the dominant commercial variable on every SaaS embed and the reason a ramp profile matters more than a discount point.
The corollary is a procurement rule rather than an adoption one: license against measured adoption on a ramp, typically 50 percent in year one, 75 in year two, and 100 in year three, and make the escalation conditional on the usage evidence rather than on the calendar.
Buyers who negotiated the AI add on separately from the suite renewal gave up 10 to 25 percent of available discount.
The AI line is leverage on the wider agreement and is worth far less negotiated alone, because the vendor wants the AI attach and will pay for it in the terms of the thing it is attached to.
Sequencing the two conversations into one event is close to free and it is the single most common procedural mistake in this category.
Data terms have to close before price, and default positions were accepted then reopened at material cost.
Four terms decide whether a deal is safe to sign: a training carve out stating that customer data, prompts, and completions are not used for foundation or shared model training; an abuse monitoring exception against the default thirty day retention of prompts and completions.
Output ownership with the vendor licence limited to service operation; and vendor indemnification on third party IP claims arising from foundation model output.
All four belong in the order before discount talks open.
Commitment buys unit price, not volume, and the term should be short on price. Trade a 12 to 36 month spend commitment for a 25 to 45 percent unit price reduction on hyperscaler platforms, never for additional licences.
Because AI vendor pricing has been falling 20 to 40 percent annually, lock no more than twelve months on unit price and take the longer term only on the minimum commitment.
Hold an exit lever at every renewal: termination for convenience or model deprecation rights, data export in plain text formats, and prompt and completion portability.
Five categories, five commercial models
| Category | Commercial unit | Typical share of AI budget | What it negotiates |
|---|---|---|---|
| Hyperscaler model platforms | Token usage, 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 licence | 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 licence terms, GPU procurement, support model |
The common failure is a single framework drafted for one category and applied to all five.
A per user per month template with a three year term and a ten percent annual uplift cap is a reasonable instrument for a SaaS embed and the wrong instrument entirely for a hyperscaler platform deal, where the commercial unit is tokens and the right play is a twelve month committed spend with a unit price reset clause.
The framework does not merely underperform on the wrong category, it collapses, because the clauses have no purchase on the unit being sold. Decide the category before drafting the paper. The platform economics sit in the AI platform TCO comparison.
Use case discipline before any vendor conversation
- Start six to nine months before signature, with the first ninety days fixing the use cases, the success metrics, and the data perimeter rather than shortlisting vendors.
- Require three documented use cases before selection opens, each with an adoption target, a measurable outcome, and a defined data perimeter.
- The volume play: high frequency, low complexity work such as drafting, summarisation, and search, with an adoption target above 60 percent of the target user base, measured on time saved per user per week.
- The judgment play: medium frequency, high impact work such as analysis, recommendation, and code generation, target above 30 percent, measured on an outcome metric rather than on usage.
- The differentiator: low frequency, high differentiation work such as customer facing agents and regulatory drafting, target above 10 percent, measured on revenue impact or risk reduction.
The enterprise AI procurement strategy brief
End to end sourcing, contracting, and renewal across the five categories, with the data terms, the commitment arithmetic, and the exit levers.
Get the white paper →The four data terms, before any price talk
| Data term | Default vendor position | Buyer side counter |
|---|---|---|
| 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 broad licence | Customer owns outputs, vendor licence limited to service operation |
| Indemnification | Limited, customer responsible for output | Vendor indemnification on third party IP claims from model output |
The commercial maths then runs differently per category, and applying the wrong one is how buyers overpay on three categories while under governing the other two.
On hyperscaler platforms the unit is input and output tokens plus reserved throughput for predictable workloads, the discount lever is committed spend through the cloud commitment vehicle.
And the trade is a 12 to 36 month commitment for a 25 to 45 percent unit price reduction with credit pool flexibility across services.
On SaaS embeds the unit is per user per month, the levers are the enterprise agreement discount, the volume tier, the elimination of deactivation lag, and the true up cadence, and the trade is a commitment with a ramp profile tied to adoption for a 10 to 25 percent unit reduction.
On agent platforms the unit is per agent action or per agent licence, the levers are action volume commitment, licence bundling, and idle session policy, and the trade is a six to twelve month pilot followed by a 24 month commitment at a 30 to 50 percent unit reduction.
Alongside all three sits portfolio discipline: run one hyperscaler, one SaaS embed, and one open weight model in parallel, because single vendor lock in on a platform layer this new is a bet on a market that has repriced every year.
Data residency belongs in the same conversation, with sovereign options documented per region rather than assumed. The seat side of the Microsoft embed sits in the Copilot pricing guide and the platform side in the Azure OpenAI guide.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across enterprise AI procurements, 2024 to 2025
Across roughly 25 to 35 enterprise AI procurements we ran between 2024 and 2025, seat utilisation and data terms decided value far more than the per seat price did:
Active weekly usage after ninety days, which is why a ramp tied to adoption evidence beats a discount point on the full population.
Available discount surrendered by buyers who negotiated the AI add on separately from the suite renewal instead of as one event.
Three patterns recurred: active weekly use settling at 35 to 45 percent of licensed seats after ninety days, buyers negotiating the AI add on separately from the suite renewal giving up 10 to 25 percent of available discount.
And default data residency and training boundaries accepted in most first drafts then reopened later at material cost.
The buyer side move is procedural rather than clever.
Scope three use cases with adoption targets before any vendor conversation, close the four data terms before price, pick the commercial framework that matches the category's unit, trade commitment for unit price rather than volume, keep the price term to twelve months while AI pricing keeps falling.
And hold an exit lever at every renewal point.
Your first five moves
- Classify the spend into the five categories before drafting any paper, because a framework built for one commercial unit collapses on another rather than merely underperforming.
- Scope three use cases with adoption targets and KPIs before vendor selection opens, starting six to nine months before signature with the first ninety days on scope rather than shortlists.
- Close the four data terms before discount talks: training carve out, retention exception, output ownership, and indemnification, all written into the order.
- Trade commitment for unit price, never for additional licences, and size SaaS embed seats on a ramp tied to adoption evidence, since only 35 to 45 percent are in weekly use at ninety days.
- Negotiate the AI line inside the suite renewal, not beside it, and keep the price term to twelve months with an exit lever at every renewal. The GenAI practice runs the procurement with you.
Frequently asked questions
Why treat enterprise AI as five categories rather than one?
Because hyperscaler model platforms, SaaS embeds, agent platforms, fine tuning programmes, and internal hosting each carry a different commercial unit, different data terms, and different exit paths.
A single framework drafted for one of them collapses on another: a per user per month template has no purchase on a deal priced in tokens, where the right play is a short committed spend with a unit price reset.
How many licensed AI seats actually get used?
In our file, active weekly use settled at 35 to 45 percent of licensed seats after ninety days, leaving most seats idle.
That is why seat sizing is the dominant commercial variable on SaaS embeds, and why a ramp profile tied to measured adoption, typically 50 percent in year one rising to 100 by year three, is worth more than an extra discount point.
Which data terms need to close before price?
Four. A training carve out stating that customer data, prompts, and completions are not used for foundation or shared model training. An exception to the default thirty day abuse monitoring retention. Output ownership with the vendor licence limited to service operation.
And vendor indemnification on third party IP claims arising from foundation model output.
What should a commitment buy?
Unit price, never additional licences. On hyperscaler platforms a 12 to 36 month spend commitment buys a 25 to 45 percent unit price reduction plus credit pool flexibility. On SaaS embeds the same trade runs at 10 to 25 percent with a ramp profile.
On agent platforms, a six to twelve month pilot followed by a 24 month commitment reaches 30 to 50 percent.
How long should an AI price term run?
No more than twelve months on unit price, with the longer term taken only on the minimum commitment. AI vendor pricing has been falling 20 to 40 percent annually, so a three year unit price lock is a bet against the direction of the whole market.
Structure the commitment long and the price short rather than the reverse.
Should the AI add on be negotiated separately from the suite?
No. Buyers who negotiated it separately gave up 10 to 25 percent of available discount in our file, because the AI line is leverage on the wider agreement and the vendor will pay for the attach in the terms of the thing it attaches to.
Sequencing both conversations into one event costs nothing and is the most common procedural mistake in the category.
What exit levers should an AI contract carry?
Termination for convenience or model deprecation rights, data export rights in plain text formats, and prompt and completion portability, at every renewal point rather than only at term end.
Alongside them, run a three vendor portfolio, one hyperscaler, one SaaS embed, and one open weight model, so single vendor lock in never forms on a platform layer this new.
