The cloud bill nobody talks about: "free" AI features in your SaaS
Every SaaS vendor shipped AI features last year. Most of them are billed as included. They are not free, they are baked into the seat price, and the seat price is moving.
The interesting line item on the next twelve months of SaaS renewals isn’t going to be the headline price increase. It’s going to be the AI features that landed quietly across the catalog through 2024 and that are now showing up in renewal conversations as either “included in the new tier” or “available as an add-on” or “you’ll need to step up to Premium to keep using.” None of those phrasings means free. They mean the cost has moved somewhere harder to track than a line item with the word “AI” in it.
Worth being explicit about where the cost actually shows up, and what to do about it before the renewal cycle locks it in.
Where the cost is hiding
Five places where AI cost lands on a SaaS bill, drawn from the public reporting and my own SaaS audits, the kind of cost the procurement conversation usually doesn’t see coming:
The seat price moved. The most common pattern. A vendor adds AI features to the existing tier, the tier price goes up at next renewal, and the sales conversation frames it as “the platform got more capable.” Sometimes it did. The pricing math is roughly: vendor pays inference costs at scale, splits the difference with the customer, embeds the customer’s portion in the seat. The customer doesn’t see the AI as a line item because there isn’t one.
A new SKU appeared. Microsoft 365 Copilot at $30/user/month on top of the base license. Salesforce Einstein 1 Studio metered separately. Zoom AI Companion as a standalone add-on for the larger plans. These are easier to spot because they’re discrete charges, but they’re also the ones most likely to get a “do we actually need this” conversation in procurement. The seat-price-moved version often doesn’t.
The storage tier you were on no longer fits. AI features in document-shaped products (Notion, Confluence, Google Workspace, Microsoft 365) index your content and produce embeddings, summaries, and search indexes. That generated content lives in your storage budget. Workspaces that fit comfortably in the included storage in 2023 now hit limits sooner because there’s effectively a 2–3× shadow content layer the AI features produce. The renewal conversation surfaces it as “you’ve outgrown the team plan.”
The usage cap got smaller in real terms. “Unlimited” usage caps have a way of turning into “fair use” caps once AI inference is in the mix. The fine print starts mentioning “AI requests are subject to throttling,” and the throttle behavior gets more aggressive on the lower tiers. You don’t see the cost, you see a feature that worked yesterday and doesn’t work as well today.
The vendor renegotiated the contract mid-cycle. A pattern that’s been visible across the larger SaaS deals: vendors come back mid-contract to “discuss roadmap” and the conversation lands on a re-papering with a new pricing structure that reflects the AI investments. Whether the customer has leverage to push back depends on switching cost. For most SaaS-in-the-business-process categories, switching cost is high.
What the vendors are actually paying for these features
The honest answer most vendors won’t say out loud: the inference cost for the AI features is real, it’s mostly going to whichever frontier-model vendor they’ve integrated (OpenAI, Anthropic, increasingly Google), and the margin on those features is currently negative or low-margin for the SaaS vendor at typical usage levels. The same dynamic that drove the Stargate-style infrastructure announcements, somebody has to pay for the inference, and the per-call cost is small but the call volume is enormous, is showing up one tier removed in the SaaS-bundling layer.
The strategic bet a SaaS vendor is making is that AI features become competitive table stakes and that being ahead on bundling them locks in customers through the renewal cycle. Margin compresses now in exchange for retention later. The vendor expects three things to happen over the next 24 months: inference costs come down (which is happening, and DeepSeek’s pricing accelerated it), they get better at routing cheap requests to cheap models, and customers get sticky enough that the vendor can re-price the bundle once the moat hardens.
You’re paying for the third part now, in a renewal cycle, in the form of bundled features whose underlying inference costs the vendor is partially eating. Whether you’d rather pay less now and not have those features is a real question the renewal conversation rarely surfaces explicitly.
What to do about it before renewal
A few practical moves for a renewal cycle in 2025:
Get a usage report on the AI features specifically. Most vendors can produce one. The interesting metric is “what fraction of seats are actually using the AI features and how often.” If the answer is “a small fraction, occasionally,” you have leverage to negotiate either a different tier or a downgrade that strips the AI features. If the answer is “everyone, every day,” the vendor knows it and the renewal will reflect it.
Audit the storage growth specifically tied to AI-generated content. This shows up in document, knowledge-base, and CRM products especially. A 30% storage growth year-over-year that correlates with feature enablement is the AI cost showing up in a category nobody’s tracking against AI.
Treat add-on SKUs as the cleanest line of conversation. When the AI cost is a discrete add-on, you can negotiate it as a separate procurement decision. When it’s bundled, you can’t. If a vendor offers both, the discrete-SKU version is usually the better long-term position even if it costs slightly more on the current invoice.
Don’t sign multi-year on AI bundling. The pricing on this category is moving fast in both directions, features expanding, underlying inference getting cheaper, competitive pressure landing. A three-year contract on the current bundle is locking in a price structure that’s likely to look generous to the vendor by year two. Annual renewal preserves your ability to re-negotiate as the market resets.
Map the AI features to whether you’d build them yourself. For a meaningful number of cases, the AI feature your SaaS vendor charges for is a thin wrapper around a capability you could implement directly against an LLM API for less. That’s not always the right move, vendor integration matters, support matters, the data already lives in their system, but it’s worth knowing where the bundled-feature premium is and how big it is.
The thing the vendor presentations elide
The framing every SaaS vendor lands on for AI features is “we’re investing in your productivity.” The honest framing is “we’re investing in our retention by giving you features the next vendor would have to match to win the deal.” Both of those are true at the same time. The bill you pay reflects the second one more than the first.
Like most cloud-bill conversations, the work is in seeing where the cost actually lives. AI features are now part of every SaaS conversation. They are not, despite the framing, free. They are paid for in the seat price, the new SKU, the storage tier, the usage cap, or the re-negotiated contract, and the version of the bill that shows up in your finance tooling almost never separates them out. The audit work to make them visible is the work worth doing before the next renewal cycle, not after.