AI add on contracts signed in 2024 and 2025 are now reaching their first renewal anniversaries. The repricing is real, the bands are wide, and the buyers without caps at signing are paying the test in full. This report sizes the cliff and lays out the clauses that hold the line.
AI add on contracts signed in 2024 and 2025 are now reaching their first renewal anniversaries. The bills coming back are not the bills that were signed. This report reads the first wave of AI renewal cliffs, sizes the repricing band by vendor, and lays out the clauses that hold the line.
About this report
The AI Renewal Cliff Report is a buyer side read on what happens when AI add on contracts signed in 2024 and 2025 reach their first renewal in 2026 and 2027. It draws on three inputs.
We report bands and directions, not precise rates. Individual outcomes vary widely with estate size, seat count, term length, and whether a cap was negotiated at signing. Where a single number appears, treat it as the middle of a range rather than a guarantee.
The AI renewal cliff is the gap between the rate a buyer signed in 2024 or 2025 and the rate the same vendor now quotes. In our engagement file the opening ask lands in a 20 to 45 percent band, with outliers above. It is the first real test of how AI prices outside a promo window.
The cliff is hitting now because most of the early enterprise AI add ons carried promotional pricing, a discount stack, or both. Those promo windows are expiring as the contracts reach their twelve or twenty four month anniversaries. The promo year was the runway. The renewal is the price the vendor really wants.
The mechanic is straightforward. A vendor opens a new AI category with promotional pricing to drive adoption inside an installed base. The first wave of customers signs in the promo window. The vendor then watches realized usage and willingness to pay through the term.
At renewal the vendor moves the rate to the level the data suggests the market will accept. That level is rarely the signing rate. It is the test of how much pricing the category will bear, and the first batch of renewing customers funds the test.
This is not unique to AI. The same mechanic produced the subscription resets on traditional software in the early 2020s. What is unique is the speed and the size of the move, because the AI category has no settled pricing history to anchor against.
The Microsoft 365 Copilot add on at 30 dollars per user per month, the early Salesforce Agentforce conversation pricing, and the first wave of ServiceNow Now Assist seats were all signed inside a roughly eighteen month window between mid 2024 and late 2025. Twelve months later, those signings reach renewal. That is the cliff timing.
Each vendor has slightly different mechanics, but the calendar effect is the same. The first AI renewal cliff in 2026 sets the tone. The second wave in 2027 will be larger by volume, because the broader enterprise rollouts came later.
A normal platform renewal moves inside a relatively narrow band, typically a low single digit uplift if a cap was negotiated. An AI add on renewal does not. It is a new line item that vendors are still pricing into the market, and the band is wider by an order of magnitude.
That width is the whole reason the cliff matters. A 3 percent platform uplift is a budget exercise. A 30 percent AI uplift is a portfolio decision, and it changes the math on every AI business case that was built on the signing rate.
Agent products carry a wider cliff band because the unit they price is new and the vendor has more room to reset the minimum. A conversation or action unit can be repriced without a direct buyer comparable, which is the opposite of a seat priced product.
The minimum unit count also resets at renewal on most agent products. The buyer who bought a 100,000 conversation pack may find a 250,000 pack is now the smallest available tier. The headline rate can even fall while the contracted floor doubles.
The blended opening ask across the vendors we tracked lands in a 20 to 45 percent band on the first AI renewal. That is a wide spread, and the spread is the point. Vendors are testing how much the market will absorb, and the early signals are not consistent.
The realized uplift after a structured negotiation is materially lower. In our engagement file, the realized number typically lands at 40 to 60 percent of the vendor opening ask, with the capped uplift clause doing most of the work. The cliff is real, but it is not fully payable.
The cliff is not a normal distribution. It is a bimodal one. A first peak sits in the 5 to 12 percent band, where capped contracts land. A second peak sits in the 25 to 35 percent band, where uncapped contracts land. Very few outcomes sit in the middle.
That shape matters because averages mislead. The market average for the AI cliff in 2026 is around 18 percent, but almost no individual buyer actually pays 18 percent. They pay either the capped number near 8 or the uncapped number near 30. The cap clause is the variable that separates the two peaks.
Salesforce Agentforce and ServiceNow Now Assist are repricing hardest in our engagement file, with opening asks in the 20 to 45 percent band and a small number of outliers above that. Both products are early enough that vendors are still pricing into demand rather than measured value.
Microsoft 365 Copilot is repricing inside a narrower 10 to 25 percent band, but it is the largest absolute exposure for most enterprises because the seat count is the highest. A 15 percent uplift on a 30 dollar seat across thirty thousand users is a real number.
For context, a normal platform renewal uplift in 2026 sits at a low single digit number when a cap is in place, and at a high single digit number when one is not. The AI cliff is one to two orders of magnitude larger. It is a different category of repricing event, and it deserves a different category of preparation.
This is also why a single blended renewal figure obscures the AI cliff. If the platform line is 3 percent and the AI line is 30 percent, a blended 7 percent looks routine on the cover page and hides the real movement in the AI layer.
Across roughly 60 to 80 AI renewals in our file, three patterns recur. The opening ask was almost always higher than the contractual cap. The realized rate landed at 40 to 60 percent of the opening ask. Buyers who started nine months out paid materially less than buyers who started inside 90 days.
These patterns held across vendors and across AI categories. They also held across enterprise sizes, with the cliff sizes scaling rather than the underlying mechanics changing. The pattern is now well enough established to plan against.
The hardest repricers are the agent and assistant products built into existing platforms, where the vendor controls both the seat and the AI layer. Salesforce Agentforce and ServiceNow Now Assist sit at the top of the band. Their account teams have the most leverage and the least competitive pressure, because the seat is already locked.
The model API providers sit in the middle band. OpenAI and Anthropic enterprise commits are repricing through tier resets and overage rates rather than headline uplifts. The mechanic is different, but the realized cost can move materially.
Microsoft 365 Copilot is the largest of these by exposure. The list rate is documented on the Microsoft 365 Copilot pricing page, but enterprise renewal terms vary. The Copilot cliff is usually a tier mix shift inside the EA rather than a headline rate move, and can still add 15 to 20 percent.
Google Workspace AI sits in a similar seat priced model and has the narrowest cliff band in our file. The lower band reflects Google leaning into adoption rather than repricing, at least through the 2026 cycle.
Salesforce Agentforce is the cleanest example of the agent product cliff. The Salesforce Agentforce pricing page documents the conversation based unit, but the renewal mechanic is the bigger story. The minimum unit count resets, the conversation rate is requoted, and the bundled assumptions on which the original deal was sized often change.
ServiceNow Now Assist repricing follows a similar shape. The licensed seats reset, the AI per case rate is requoted, and the original Now Assist pricing tier may no longer exist by the time the renewal lands.
OpenAI enterprise contracts price largely on token volume tiers and minimum commits. The OpenAI enterprise page sets the public starting position, but enterprise repricing typically lands through new commit minimums and changed overage rates rather than a headline rate change.
Anthropic enterprise commits work similarly. The cliff mechanic is the tier reset rather than a single percent uplift, and it can be larger than the seat priced cliffs once the realized usage runs above the new floor.
Platform vendors with AI assistants, including the AI add ons embedded in core ERP and HCM platforms, sit in their own category. The cliff there tends to land through a platform tier upgrade rather than an AI line uplift. The buyer is moved to a higher tier of the platform that includes the AI feature as a bundled item.
The headline AI line may not move much, while the overall platform spend rises by 15 to 25 percent. This is the hardest cliff to detect because it is hidden inside a tier change rather than priced directly. Read every tier change at renewal as a potential AI cliff in disguise.
One clause matters more than the rest. A capped uplift on the AI line, written at signing, cuts the realized cliff roughly in half across our engagement file. Without that clause, the buyer pays whatever the vendor quotes. With it, the buyer pays the cap and walks into a real negotiation on the gap.
Two other clauses sit just behind it. A separate term for the AI add on, distinct from the platform renewal date, lets the buyer renegotiate the AI layer on its own clock. An explicit exit right on the AI SKU prevents the buyer from being locked into a bad AI deal because the platform is essential.
Cliff defense clauses, ranked by realized impact
| Clause | Where it sits | What it does | Realized impact |
|---|---|---|---|
| Capped uplift on the AI line | Order form, AI SKU | Bounds the renewal increase to a fixed percent | Largest single lever |
| Separate AI term | Master agreement | Lets the AI layer renew on its own clock | Prevents bundle hide |
| Exit right on the AI add on | AI SKU schedule | Allows mid term or renewal exit without penalty | Backstop on a bad renewal |
| Benchmarking reference | Order form | Pegs renewal to a comparable peer set | Holds the floor |
| Swap right between AI SKUs | Master agreement | Reallocates from one AI add on to another | Flexibility through the cycle |
A useful AI cap looks like a fixed percent uplift, written into the order form, applied to the AI SKU specifically rather than to the whole order. Five percent is strong, eight to ten percent is realistic, and anything above fifteen percent is essentially uncapped given the bands we are seeing.
The cap should bite at both renewal and any list price adjustment during the term. Vendors will often offer a renewal cap that is silent on mid term list adjustments, and that gap is where the cliff returns.
An exit right on the AI SKU is the buyer side backstop. If the AI add on fails to deliver the value that justified signing, an exit right lets the buyer drop the line without penalty at the next anniversary. Most first AI contracts did not include one.
The exit right is asymmetric. Vendors offer it reluctantly because it constrains them, but the request is reasonable on a category the vendor itself is still pricing into the market. Frame the exit right as the price of accepting an uncertain renewal trajectory.
Practical exit rights name the conditions cleanly. A drop in active users below a defined threshold, a missed feature commitment, or a list price increase above a stated cap are all standard triggers worth writing into the AI SKU schedule.
Rolling the AI add on into the platform renewal date makes the AI repricing invisible. A separate term forces the AI conversation onto its own calendar, with its own benchmarking and its own escalation path. That visibility alone changes the negotiation.
A separate term also lets the buyer time the AI renewal independently of the platform. A six month gap between the platform renewal and the AI renewal is enough to recover negotiating leverage on the AI line.
The standard advice from account teams is to roll the AI add on into the platform renewal so the bill is simpler. We disagree. In roughly 50 of the 60 to 80 AI renewals we have advised, the AI line rolled into the platform vanished inside a blended uplift, and the buyer lost the cap argument because the cap sat on the platform, not on the AI line. The buyer side move is the opposite. Treat every AI add on as a separately termed contract, cap the AI line at signing, and reopen it at every cycle. The bundle is the vendor play. The unbundled AI line is the buyer play.
Source: Redress Compliance advisory engagement file, 2024 to 2025.
The most reliable cliff defense begins nine months before the AI anniversary, not at the quote. By that point the buyer should already have the realized usage data, the contract clauses summarized, and a sense of which clauses are missing. The conversation with the vendor opens from a prepared position rather than a reactive one.
An early conversation also reframes the renewal. Instead of debating the vendor opening quote, the buyer leads with the realized usage data and the gap to the contracted minimum. The vendor responds to that data rather than introducing a number first. The whole sequence shifts the anchor of the negotiation.
Build a one page inventory for every AI line in the estate. Capture the contracted seats or units, the realized usage, the per unit rate, the term end date, and the cap status. The inventory is the single artifact the negotiation team reaches for at every step of the cycle.
The inventory also surfaces silent surprises. A line that auto renewed without notice, a unit count that crept above the contract floor, or a tier that was upgraded without consent all show up in the inventory and become negotiation items.
The AI renewal needs a clear owner inside the buyer, not a committee. Procurement leads the commercial conversation, the platform team feeds in the usage data, and finance signs off on the cap and reserve. Without a clear owner, the renewal drifts and the vendor sets the timing.
The owner also handles communication with the vendor. A single point of contact prevents the vendor from playing different stakeholders against each other, which is a common pattern on AI deals where the platform team and procurement do not share data.
The first vendor conversation is short. The buyer names the renewal anniversary, names the realized usage, and asks the vendor to bring a proposal that reflects the realized usage and the existing cap. The buyer does not invite an opening rate. The vendor returns to the table with a more grounded number.
The script also names the next milestone. A follow up in two weeks with a written proposal sets the tempo. AI renewals that drift past three months from the first conversation are usually drifting because the buyer has not enforced the calendar.
The first AI renewal is largely about damage control on a contract that was signed without the benefit of cliff data. The second AI renewal is the first one where the buyer has comparable peers, realized usage data, and a benchmarking reference. That changes the negotiation posture.
The second renewal also has to address the cumulative effect. A 30 percent uplift in year one followed by a 15 percent uplift in year two is not a 45 percent move. It is a 49.5 percent move compounded, and the cumulative number is what the budget actually carries.
The single strongest argument on the second AI renewal is the realized usage data from year one. Most enterprises overbought their first AI add on, and the actual seat or token usage is well below the contracted minimum. That gap is the negotiating room.
Bring the realized usage to the table before the vendor proposes the new rate. The conversation then becomes a right sizing exercise rather than a discount exercise, which is the conversation buyers tend to win.
On the first AI renewal, few buyers had a credible alternative because the AI category was too new. On the second renewal, that excuse is gone. Multiple credible providers exist in every AI category, and the buyer who has costed even a partial migration has more leverage than the buyer who has not.
The alternative does not have to be a full replacement. A partial workload move, a multi provider procurement, or a credible pilot with a competing vendor is enough to reset the conversation.
The second AI renewal is the right place to add the clauses the first one was missing. A per year cap, a separate AI term, an explicit exit right, and a benchmarking reference can all be inserted at this renewal even where the original contract did not carry them.
Vendors are usually willing to add the clauses at the second renewal because the alternative is the buyer reopening the entire deal. Frame it as an exchange. The buyer agrees to a longer term in return for the cap and exit rights on the AI line.
The second AI renewal is also the right moment to right size the AI scope. A first AI signing often overbought seats or tokens. The realized usage tells the story. Cut the contracted minimum back to where the realized usage actually sits, with a small headroom buffer.
Right sizing on the second renewal compounds for the rest of the contract life. Every year after, the renewal is applied to a smaller base. The savings are not just in this renewal. They are in every future renewal of the same line.
The cliff is not a single number. It splits cleanly by AI pricing model, and the defensive moves are different for each. Seat priced AI add ons reprice through list moves and tier mix shifts. Consumption priced models reprice through unit cost changes and overage. Agent products reprice through both at once.
Seat priced AI add ons are the simplest cliff to model and the easiest to cap. The renewal moves the rate or the tier mix, and the buyer can plan against a published list rate. The defense is a per seat cap, a tier protection clause, and a no upsell to a higher tier without buyer consent.
The risk on seat priced AI is the silent tier shift. Vendors will quietly retire the lower priced AI tier and migrate the seats to a higher tier at renewal. The buyer sees a small headline rate move and a large bill, because the seat mix has changed.
Consumption priced AI cliffs are harder to predict and harder to bound. The unit cost moves, the tier boundaries move, the overage rate changes, and the buyer cannot fully model the cliff without committing first. The defense is a tier lock for a defined term and a rate floor that survives consumption changes.
Consumption pricing also rewards aggressive right sizing. A buyer who has cut waste and predictable workload off the AI tier before the renewal pays a smaller cliff because the base is smaller. This is the same right sizing lever that works on platform renewals, applied to AI.
Agent products carry both cliff types at once. The seat or named user component reprices like a seat product, and the conversation or action component reprices like a consumption product. The defense is double the work, and the upside is double the protection if both clauses are in the contract.
Agent products are also the youngest AI category, so the cliff bands are still widening. We expect the agent cliff in 2027 to be larger than the agent cliff in 2026, because vendors are still calibrating the market.
Cliff risk by AI category
| Category | Pricing model | Cliff mechanic | Typical band |
|---|---|---|---|
| Copilot type seats | Per user per month | List move plus tier mix shift | 10 to 25% |
| Agent products | Per conversation or action | Unit price and minimum reset | 20 to 45% |
| Model API consumption | Per token, tiered | Tier reset and overage rate change | 15 to 40% |
| Bundled AI in platform | Platform line item | Hidden inside platform uplift | Opaque, often largest |
| AI compute commits | Annual unit commit | Recommit at new minimum, new rate | 10 to 30% |
The first AI renewal is the test the vendor is running on the market. The buyer with a cap pays the test once. The buyer without one pays it every year.
The cliff is easiest to see in absolute dollars on a representative estate. A mid sized enterprise that signed Copilot for 20,000 users at 30 dollars per user per month carries a base AI line near 7.2 million dollars per year, before any cliff.
A 20 percent cliff at the first renewal moves that line to roughly 8.6 million, an absolute increase of 1.4 million dollars at the AI layer alone. A 35 percent cliff moves it to 9.7 million, an increase of 2.5 million. Across larger estates the absolute numbers scale linearly.
The realized number after negotiation is materially lower than the opening ask. In the same 20,000 seat example, a capped 8 percent realized renewal lands at 7.8 million, an increase of 600,000 dollars. The cap clause is the difference between a six figure increase and a seven figure increase.
That ratio is consistent across the engagement file. The cap clause is roughly worth between four and seven times its own legal cost over a typical three year AI contract, before any negotiation effort is counted.
The AI cliff does not replace the platform uplift. It stacks on top. The same mid sized enterprise also faces a 3 to 8 percent platform renewal on the base license, which is a separate increase on a separate line.
Reading the cliff and the platform uplift as two distinct events is essential. Vendors prefer a blended figure because it understates the AI move. Buyers should insist on the two lines reported separately to understand what the actual repricing is.
Across the 60 to 80 AI engagements in the file, buyers with caps at signing saved 12 to 20 percent of their five year AI spend versus peers without caps. On a 30 million dollar AI commit that is a 4 to 6 million dollar swing across the term, earned at signing.
That cumulative number is what should drive the cap conversation at any new AI signing in 2026 and 2027. It is not a renewal lever. It is a signing lever. Buyers who push the cap at signing pay the cumulative reward across every renewal in the contract life.
For finance planning, the prudent assumption is a 25 percent AI line increase at the first renewal unless a cap is in place. The number can be revised down once the negotiation completes, but the budget reserve should reflect the bands the cliff is actually producing.
Underbudgeting the AI cliff is the most common failure mode on this category. The signing rate is treated as the run rate, the renewal lands as a surprise, and the AI program loses credibility inside the business because the cost moved without warning.
The 2027 cliff will be larger by volume because the broader enterprise AI rollouts came in late 2024 and through 2025. More contracts will reach their first renewal anniversary inside that window. Gartner IT spending releases have already flagged software and AI as the fastest growing categories, with price as a material share of that growth.
The 2027 cliff will also widen by vendor. Agent products and AI assistants from platform vendors are still finding their pricing model. Until that model stabilizes, every renewal is also a repricing event, not a routine uplift.
The dominant 2027 cliff driver will be promo expiries on contracts signed inside the 2025 enterprise AI surge. Many of those carried a year of promotional pricing, a discount stack, or an introductory tier. As those expire, the realized rate moves to list, and the cliff lands.
Map every AI signing inside your estate against its promo expiry date. The cliff is foreseeable to the month if the contract is read carefully. Vendors rarely volunteer that calendar.
Regulation may bend the cliff trajectory in two ways. New AI transparency rules in Europe and pending US guidance may require vendors to publish AI usage and pricing terms in more detail, which would narrow the gap between opening ask and realized rate. The effect would be small at first but compounding.
Regulation may also slow the cliff in regulated sectors specifically. Financial services and healthcare buyers facing strict AI governance requirements have more leverage than other sectors because the cost of switching vendors mid term is higher. Vendors recognize that and price more conservatively in those segments.
The defensive posture is straightforward. Put a cap on every AI line that does not have one. Pull the AI add on out of the platform renewal date. Build a multi provider option in every AI category. Start the AI renewal conversation nine months before the anniversary, not three.
None of these moves is new. All of them are routine on platform renewals. The cliff is large because the same disciplines have not yet been applied to the AI layer in most enterprises.
Through 2027 we expect vendors to shift the AI cliff away from the headline rate and into the consumption meter. Microsoft, Salesforce, and ServiceNow have all started leaning on consumption units inside their AI products, which moves the repricing onto a less visible part of the bill.
The same shift is visible at the model API layer. OpenAI and Anthropic are pricing more new capability into per token tiers rather than into headline list moves. The buyer side adjustment is to watch the meter, not just the list rate.
Bundling pressure will increase through 2027 because bundling hides the cliff. Account teams will push to roll AI into platform agreements, into security agreements, and into developer agreements. Each bundle is another place where the AI repricing can disappear into a blended line.
The buyer side response is unbundling pressure. Insist on AI as a line item, separately priced and separately capped, in every contract that contains it. The added work at signing pays back every renewal.
If the average AI cliff lands at 25 percent on the first renewal and 15 percent on the second, an uncapped AI line is roughly 44 percent more expensive after two renewals than at signing. Capped AI lines compound much more gently, typically holding near the signing rate plus the negotiated cap.
That cumulative gap is the entire value of the cap and term work. It is a one time clause negotiation that pays back every renewal for the life of the AI add on.
The AI renewal cliff is the gap between the rate a buyer signed in 2024 or 2025 and the rate the same vendor now quotes. The opening ask lands in a 20 to 45 percent band. It is the first real test of how AI prices outside a promo window.
Because most early enterprise AI add ons carried promotional pricing inside their first year, and those promo windows are now expiring on contracts signed between mid 2024 and late 2025. Twelve months later, the contracts reach renewal and the promotional rate falls off. The promo year was the runway. The renewal is the price the vendor really wants.
The opening ask on the first AI renewal lands in a 20 to 45 percent band across vendors, with outliers above. The realized uplift after a structured negotiation typically lands at 40 to 60 percent of that opening ask. The vendor mix and the clauses in place at signing drive the spread.
Salesforce Agentforce and ServiceNow Now Assist are repricing hardest in our engagement file, with opening asks in the upper end of the band. Microsoft 365 Copilot is repricing inside a narrower range but is the largest absolute exposure because of seat count. Anthropic and OpenAI enterprise contracts reprice through tier resets and overage rather than headline uplifts.
A capped uplift on the AI line, written at signing, cuts the realized cliff roughly in half in our engagement file. A separate term lets the buyer renegotiate the AI layer on its own clock. An explicit exit right prevents lock in. The cap matters most by a wide margin.
Yes, at least on a separate term and a separate cap, even inside the same master agreement. Rolling the AI add on into the platform renewal hides the AI repricing inside a blended uplift, and the platform cap does not protect the AI line. A separate AI term forces the AI conversation onto its own clock.
Anchor on realized usage from year one, because most enterprises overbought their first AI add on and actual usage sits below the contracted minimum. Bring a credible multi provider alternative before the vendor proposes the new rate. Treat the second renewal as a right sizing exercise, not a discount exercise.
Map every promo expiry date and open the renewal conversation at least nine months out. Pull the realized usage. Identify the AI lines that lack a cap. Cost a credible alternative in every AI category. Buyers who start early pay materially less than buyers who start late.
Only if a capped uplift is written into each year of the term. A two year AI commit at a fixed first year rate with no cap on the second year defers the cliff by twelve months. The protection is in the cap clause, not the term length.
It will keep repeating until AI pricing models stabilize, which we do not expect inside the next two renewal cycles. Every renewal until the model settles is also a repricing event. The defensive posture, caps and separate terms, will need to be maintained through 2027 and into 2028.
Large enterprises face the largest absolute cliff because their AI commits are concentrated. Mid market enterprises face smaller individual cliffs spread across more AI lines. Both profiles benefit from the same defensive clauses, but the leverage is different. Large estates lean on the multi provider option, mid market estates lean on calendar discipline.
You pay the test. In our engagement file, buyers who accepted the first AI renewal quote without negotiating typically paid the full opening ask. That sets the baseline for every future renewal. Future increases compound on a higher base, and the AI line becomes harder to reset later. The do nothing path is rarely the cheapest path.
The vendor by vendor cliff bands, the realized uplift model, the per category defense playbook, and the AI clause checklist that holds the gap widest.
Used across more than five hundred enterprise engagements. Independent. Buyer side. Built for procurement and finance leaders running the next AI renewal cycle.
The first AI renewal is the test the vendor is running on the market. The buyer with a cap pays the test once. The buyer without one pays it every year.
