Vendors' AI Costs Fell 90%. Your Renewal Bill Rose 20–37% Anyway.

The tax has a name now

By mid-2026, procurement teams had settled on a name for it: the AI tax. Renewals that used to move a few percentage points a year were arriving 20 to 37 percent higher, and "AI" was the line item vendors pointed to when asked why. Procurement-data firm Tropic tracked the pattern across a large sample of anonymized enterprise contract renewals and found it holding across categories: ITSM, CRM, customer support, HR software. It wasn't one aggressive vendor. It was close to the median ask.

The number that made the increase land badly wasn't the 20 to 37 percent on its own. It was what sat next to it. Vendors' own AI infrastructure costs fell 40 to 70 percent over the same stretch, and the wholesale price of a token from a frontier model dropped by something close to 90 percent between 2024 and 2026 — a rate that went from around three cents per thousand tokens to roughly two-tenths of a cent. The input cost behind an "AI-powered" feature got dramatically cheaper. The number on the renewal invoice did not move the same direction.

The gap shows up in the aggregate numbers too. Software subscription and licensing prices have been rising at an annualized rate of roughly 12.2 percent, nearly five times the 2.5 percent general inflation rate across G7 economies over the same period. Per-employee SaaS spend climbed to about $9,100 a year in 2025, up from roughly $7,900 in 2023. None of that is explained by a comparable rise in what it costs a vendor to run the software. It's explained by what buyers were willing to keep paying while budgets were still catching up to how central these tools had become — SaaS now accounts for something close to one dollar in every eight of total IT spend, and around 70 percent of the typical IT budget goes to keeping the systems already in place running, which leaves finance teams with less room to simply walk away from a renewal than the negotiating advice usually assumes.

Where the increase actually lives

The AI tax rarely shows up as a single restated price. It moves through three mechanisms, and most vendors of any size now use at least two of them at once.

The first is the forced SKU migration: the tier a customer is already using gets AI features bolted on, the old tier is deprecated, and the only path forward runs through a new tier priced well above the old one. The second is the metered agent credit: a per-action or per-credit charge layered on top of the existing seat price, billed separately and hard to forecast. The third is outcome-based pricing: a fee tied to a completed task rather than a seat or a login, which sounds fair until volume scales faster than the budget that was approved for it.

The second and third mechanisms share a specific failure mode: nobody signs off on a metered line item the same way they sign off on a seat count. A recent survey of IT leaders found that 78 percent had been hit with an unexpected charge traceable to a consumption-based or AI-tied pricing model — not a vendor error, just an invoice that reflected usage nobody had modeled in advance. A support tool billed per agent action or a document workflow billed per completed task can turn a predictable per-seat line into something that reads, and arrives, like a utility bill.

Once AI capability sits inside the tier a team already depends on for its integrations and workflows, declining the AI portion stops being a realistic option in most renewal conversations. That's what makes the SKU migration the most effective version of the tax: it doesn't ask the customer to buy something new, it just removes the option to keep buying the old thing at the old price.

The quieter version: shrinkflation

Not every vendor raises the visible number. A meaningful share hold the sticker price steady and reduce what it buys instead: a slower support SLA, a smaller included-seat count, a feature that used to be bundled now sitting behind a separate paywall. Roughly six in ten software vendors used at least one value-reduction tactic in a recent renewal cycle, and just over a quarter of late-2025 renewals included some form of downgrade at an unchanged headline price.

Why buyers have more room than the invoice suggests

The negotiation data is the part vendors don't put in the renewal email. Buyers who push back through a structured negotiation process reduce the initial AI-driven ask by roughly 55 percent on average, according to Tropic's anonymized procurement data. The final number still tends to land about 12 percent above the pre-AI baseline, so the increase doesn't disappear entirely, but the bulk of the initial ask is negotiable, not fixed.

Timing changes the outcome more than most buyers expect. Organizations that start renewal conversations 90 or more days before the contract expires report average savings around 49 percent, against roughly 19 percent for teams that start within 30 to 90 days of expiry. A same-week signature under deadline pressure is close to the worst negotiating position a buyer can be in.

That's also why the single most reported effective move is a plain one: ask for an AI-free SKU, in writing, even for teams that intend to keep the AI tier. The ask forces the vendor to separate the AI premium into its own line item and justify it on its own terms, instead of letting it hide inside one bundled number. Most vendors would rather requote than risk losing the renewal outright.

A playbook for the next 90 days

1. Start the renewal clock at 90 days out, not 30. The savings gap between early and late negotiators is one of the largest levers in this entire process.
2. Ask for the AI-free SKU explicitly. Get the quote in writing before discussing the AI tier at all.
3. Get an independent benchmark for the exact SKU and tier in question, not a generic market average. Vendors reprice differently by segment and account size.
4. Write the price-protection clause to cover SKU migrations and new metered components, not just the list-price line. A CPI-indexed cap (typically 3–5% annually) that only covers list price does nothing to stop a forced tier change.
5. Model the build-vs-buy break-even at actual usage volume before the call. In one documented case, a custom-built alternative ran $371,000–$630,000 over 36 months against roughly $487,500 for the repriced SaaS option, reaching cost parity around month 18–24 at that team's usage level.

That last point is not a build recommendation. For most teams, building the replacement is the wrong call at almost any price. What matters is that the number exists and is credible enough to change what a buyer is willing to sign. A negotiator with a real walk-away option gets a different renewal call than one without.

The same pattern from the other side of the table

For a company selling software rather than buying it, the same data cuts the other way. The negotiation numbers above aren't a fringe case; they describe how a growing share of enterprise buyers now approach every renewal by default. A pricing strategy built on bundling AI into a mandatory tier and hoping the increase goes unquestioned is pricing against a buyer population that is actively organizing to push back, with benchmarks and a documented playbook to do it. Transparent, usage-capped pricing that doesn't need a 90-day negotiation to reach a fair number is a genuinely different pitch to make in that market, and it's one fewer renewal calls end in a discount that erases the trust the increase was supposed to fund.