How to price AI features in your SaaS product
How to price AI features in your SaaS product comes down to one question most founders skip: does this feature cost you money every time someone uses it? Traditional SaaS features don't. AI features do, and if you are charging both the same flat per-seat fee, your heaviest users are already costing you more than they pay you.
The fix is not choosing usage-based or outcome-based pricing off a menu. It is figuring out which parts of your product still behave like software with near-zero marginal cost, and which parts now behave like a metered service, then pricing each one on its own terms.
Why per-seat pricing breaks the moment you add AI
Per-seat pricing works when serving one more customer costs you almost nothing. That assumption is probably baked into your whole business model: 80 to 90% gross margins, because traditional software does not get more expensive to run as more people use it.
AI features break that assumption. Every query, every inference, and every agent action consumes real compute, and that cost scales with usage in a way the rest of your product never did. SaaS companies that bolt AI onto an existing product commonly see gross margins fall from that 80 to 90% baseline to 50 to 60%, sometimes lower, purely because of the AI layer.
The damage shows up fastest in your heaviest users. GitHub Copilot reportedly lost around $80 per user per month in its early days, because power users ran far more inference than a flat subscription covered. Replit has disclosed operating at negative gross margins on some usage, meaning it loses money on the interaction itself, not just on overhead. If your AI feature sits inside a flat-rate plan today, you probably have a version of this problem already. You just have not measured it yet.
The three pricing models on the table
Three charge metrics cover almost every AI feature being priced today, and each trades cost predictability against value alignment differently.
- Consumption-based pricing charges per token, API call, or inference. It mirrors your actual infrastructure cost, so margins are predictable, but non-technical buyers do not think in tokens, and it creates real bill-shock risk. One support-tool customer reported their monthly bill jumping from $4,000 to $9,000 after a usage-pricing migration.
- Outcome-based pricing charges when the AI completes a defined job, the way Intercom's Fin agent charges $0.99 per resolved support conversation, with published real-world resolution rates between 42% and 50%, the number to model your own forecast on if you consider this path. It only works when the outcome is unambiguous and you can absorb the cost of the interactions that do not resolve.
- Hybrid pricing combines a base subscription with a usage or outcome layer on top. About 92% of AI software companies now use some form of hybrid pricing, according to Bessemer's research, and it is the default most early-stage founders should reach for, because it caps your downside while still capturing upside as usage grows.
The mistake that's quietly hiding your real costs
If your AI feature runs on promotional cloud credits from AWS, Google Cloud, or Azure, your margins look fine right now for the wrong reason. Startups routinely burn six to twelve months of credits on one provider, then move to the next, and the feature's true cost stays invisible the whole time.
The reckoning arrives when the credits run out. At that point you are choosing between a painful price increase, cutting the feature, or accepting a permanently compressed margin, and none of those conversations get easier with existing customers than they would have been with new ones from day one. Price the feature against its real cost now, while you can still design the pricing instead of retrofitting it later.
A practical framework for pricing AI features
Pricing an AI feature well takes four steps: meter the actual cost, pick a charge metric buyers already understand, set a hybrid base-plus-usage fee, and give customers a real-time usage dashboard.
- Meter it before you price it. Track cost per active user for your AI feature specifically, separate from the rest of your product, for 30 days. You cannot price what you have not measured.
- Pick a charge metric your buyer already understands. "Per resolved ticket" or "per report generated" sells. "Per token" does not, unless you are selling to developers.
- Set your base fee at roughly twice your calculated delivery cost. Layer usage or outcome credits on top of that base. It is the bridge Bessemer's portfolio companies use to move from cost-plus pricing toward value-based pricing without underpricing or shocking a buyer on day one.
- Build the usage dashboard before you need it. Founders who avoid billing disputes give customers visibility into consumption in real time, with alerts before a bill balloons, not after.
Not every AI feature needs new pricing
If your AI feature is cheap to run and mostly a retention or differentiation play rather than a core value driver, bundling it into your existing plan may be the right call for now. Some vendors are already reverting to a version of per-seat pricing for AI, charging a premium "AI seat" instead of introducing usage complexity, betting that falling inference costs make the simpler model viable again.
The point is not that usage-based or outcome-based pricing is always correct. It is that you choose deliberately, based on your actual cost per user, instead of defaulting to your existing per-seat plan because that is what you already had. If you are still deciding whether the AI feature belongs in your product at all, that is a separate question worth answering first.
What to do in the next 30 days
Before you touch your pricing page, spend the next 30 days instrumenting cost tracking for your AI feature alone: total inference cost divided by active users of that specific feature. That single number tells you whether you are already subsidizing your heaviest users, and every other decision in this piece depends on it. Once you have it, test a hybrid price, a base fee plus usage or outcome credits, with five existing customers before rolling it out broadly. If your broader pricing model has not been revisited since before you had an AI feature at all, that is worth fixing at the same time.
Frequently asked questions
Should I price my AI feature separately from the rest of my plan? Bundle it if usage is low and predictable and the feature mostly drives retention rather than core value. Price it separately once the inference cost per active user becomes a meaningful share of what that customer already pays you.
What is the difference between usage-based and outcome-based pricing for AI features? Usage-based pricing charges for consumption, tokens, API calls, or inference. Outcome-based pricing charges for a completed, verifiable result, like a resolved support ticket. Outcome pricing sells better to non-technical buyers but requires you to absorb the cost of the attempts that do not succeed.
How much do AI features typically cost to run per user? There is no universal number. Gross margins for AI-augmented SaaS products commonly land between 50% and 60%, against 80% to 90% for traditional SaaS, but your actual per-user cost depends on your model choice and usage volume. The only way to know your number is to meter it.
Is it too early to price an AI feature I just launched? No. Pricing is far easier to fix across your first 100 customers than after you have sold flat-rate access to thousands of them and have to walk back a promise.
Will per-seat pricing disappear entirely as SaaS adds more AI? Unlikely. Some vendors are reintroducing seat-based pricing for AI features as inference costs fall, betting that simplicity wins with buyers even while usage-based and outcome-based options exist for products where variable cost is higher.
Every founder bolting an AI feature onto an existing SaaS product is running the same experiment right now, and most are pricing it exactly like the rest of the product out of habit, not analysis. Meter the real cost first. The pricing model follows from that number, not the other way around. If you want a second opinion on the model before you ship it, that is worth getting right early.