The evidence pack that wins an AI renewal

If you still need convincing that renewals go better when you bring data, read how to negotiate AI renewals from evidence first. That post makes the argument. This one assumes you are sold and hands you the assembly instructions.

An evidence pack is not a slide deck and it is not a story. It is a short, dated, boring document that a vendor's account team can read in ten minutes and cannot easily dispute. It works because it turns a negotiation, where the better talker wins, into an audit, where the side holding the records wins. You hold the records.

Seven sections. Build them in this order.

1. The seat roster, with last activity and a dated pull

One row per paid seat, since a seat is a paid license for one person. Name or ID, the date it was assigned, the last activity date, and the date and time you pulled the export. Attach the raw export, not a summary of it.

The dated pull is not a formality. It makes every number downstream checkable, and it stops a three-week negotiation from turning into an argument about stale data. Pull it once, quote the timestamp everywhere, and refresh it the week you sign.

2. How many seats get used, with "active" defined in the same sentence

Two numbers: seats active, seats paid. One definition, stated where the numbers appear, not in a footnote.

Pick a definition you can defend to a doubtful account executive, then never move it. "At least one accepted suggestion in the last 30 days" is defensible. "Logged in at some point" is not, because it counts a developer who opened the tool once in March. If your definition is more generous than the vendor's, say so. They will find that out anyway, and it is much better coming from you. What good AI seat utilization actually looks like gives you a range to judge whether your number is bad or normal.

3. The spend trend, with the flat fee and the usage charge shown apart

Twelve months if you have it, six if you do not. Two lines on the same chart: the committed flat fee, and the usage charge stacked on top of it.

Combining them hides the problem. One blended total can conceal a flat fee that is shrinking and a usage charge that has tripled, which is a completely different renewal conversation. Vendors know their own usage charges better than you do. Showing that you have split the two apart signals that you will not accept a headline discount on the part of the bill that is not growing. The metered traps hiding in AI tool pricing covers the mechanics worth flagging here.

4. The duplication finding

If engineers hold seats on two overlapping tools, this section names them, counts the overlap, and shows activity on both. It is the single most uncomfortable slide for an account team, because it turns their renewal into a comparison rather than a foregone conclusion.

Be precise and be fair. Overlap is not automatically waste, and some teams have good reason to run an inline assistant and an agent-style tool for different work. Say which overlaps you think are justified. A pack claiming that all duplication is waste is easy to dismiss. One that concedes half is hard to argue with. Do you need both Copilot and Cursor? works through where the line usually falls.

5. The public-price benchmark, with the URL and the date you read it

List the vendor's published price, the published price of the nearest alternative, and your current real rate per seat per month. Include the URL for each and the date you loaded the page.

The date matters more than it sounds. AI tool pricing pages change often enough that an undated screenshot is worth nothing eight weeks later. GitHub, for instance, publishes its Copilot plan pricing openly, which makes it a stable public anchor. Use anchors like that rather than a number someone repeated to you on a call.

For a benchmark of negotiated rates rather than list prices, Vendr publishes pooled data from deals done through its marketplace. For Cursor, that data shows a median yearly contract value of $111,408 across 76 documented purchases, ranging from $19,200 to $395,040. Vendr reports 15–25% off through an annual commit, and 25–40% when you combine several terms. Attribute it exactly that way. It is a sample of buyers who chose to use a negotiation service, which skews toward larger and more deliberate buyers. It is a rough guide, not a market price. Cite it as what it is and the vendor cannot dismiss it. Overstate it as "the market rate" and they will, correctly.

6. The two asks, stated as two numbers

This is the section most packs get wrong. There are two asks. They do not overlap, and they must never be blended.

The seat cut — dropping your committed seat count to match the seats you actually use — is idle seats × price per seat. An idle seat is a paid seat nobody has used lately. You are entitled to this one however the talks go, because you are not asking for a gift. You are declining to buy something you do not use. State it as a hard number.

The discount is your real rate against the comparable benchmark, applied only to the seats you keep. That one is a negotiation.

Blend them and the vendor will happily "give" you the seat cut as though it were the discount. You will pay full rate on every seat you keep and feel like you won. Two numbers, two lines, two sign-offs.

The seat cut and the discount are two different asks. Blend them, and the vendor pays you once for both.

7. The walk-away, and what you'd actually do instead

One paragraph: the number above which you do not renew, and the specific alternative. Move everyone onto the other tool. Drop to a cheaper plan. Run a 90-day trial of a competitor. Or let it lapse for some teams.

A walk-away you have not costed is a bluff, and account teams have a good ear for bluffs. If the alternative costs three weeks of migration and a dip in output, write that down. The walk-away is still real. It is just priced. Get whoever would approve the alternative to sign off *before* the first call.

The calendar: start 60 days out

• Day 60 — pull the seat roster and 12 months of billing. Timestamp both.
• Day 55 — fix the definition of active and work out how many seats get used.
• Day 50 — run the duplication check across vendors.
• Day 45 — collect public prices and benchmarks, with URLs and dates.
• Day 40 — draft the two asks as separate numbers.
• Day 35 — get internal sign-off on the walk-away from whoever owns the alternative.
• Day 30 — send the pack to your account team. Not the day before renewal.
• Day 21–14 — first counter, second pass.
• Day 7 — refresh the roster pull and paper it.

What weakens a pack

Four things, in the order they usually show up.

Undated pulls. A roster with no timestamp is just a claim. Everything built on it becomes arguable.

An "active" you cannot defend. If your definition is soft, the usage number collapses under one question, and the collapse takes the credibility of the rest of the pack with it.

A blended ask. Covered above. It is the most expensive single mistake in the document.

A market rate you made up. One unsourced number lets a vendor re-argue your entire pack instead of your specific ask. Every benchmark gets a URL and a date, or it comes out.

The vendor's counter-moves

Expect three, and note that your pack already answers each one.

They will offer the discount in exchange for a multi-year commit. Section 3 answers that: a usage charge growing faster than the flat fee means a long commit prices the wrong risk.

They will accept the seat cut but attach a seat floor just above your current use. Section 1 answers that: the roster shows what use does month over month, not at a single point in time.

They will propose an upgrade to a higher plan that makes the per-seat math look better. Section 2 answers that: if a third of your seats sit idle, more features per seat is not the problem you have.

The honest caveats

An evidence pack improves your position. It does not guarantee an outcome. Vendors with real power — deep integration, a migration you cannot afford, a champion inside your company who will fight for them — can absorb a good pack and still hold price. Usage data has real blind spots. Some genuine use leaves no record at all, and low activity for a specialist team can be correct rather than wasteful. The Vendr figures describe one marketplace's completed deals, not the whole market. And a pack this precise takes a working day or two to build, so below roughly a $50K annual contract the honest answer is that the math may not clear the effort.

Find your recoverable AI spend

Sections 1 through 4 of this pack are exports Spendassay produces directly, dated and across every vendor, so the assembly work is mostly deciding what to ask for. Run it 60 days before your next renewal and you will walk in with the seat cut already priced.

`Start free` → /login?src=blog_ai-renewal-evidence-pack