SaaS has had a good run. The attraction of the model’s 80-90% gross margins drove the creation of countless businesses. And the profits those levels of gross margin can generate minted the bulk of the unicorns out of the 1290 startups currently granted that status.
Those margins come from the structure of the product. Your standard SaaS is a stack of software running in the cloud, accessed via a browser and sold by the seat via subscriptions.
Once you have the service running the cost of adding a new user is negligible – the zero marginal cost in this article’s title.
Take that same model – a software service accessed via a browser – and add one small tweak – core functionality is provided by AI – everything changes and gross margins drop precipitously.
How AI changes SaaS costs
Once you add AI to your product you suddenly have a new per-user expense: inference. And the problem with inference is if your calls to AI are small and fast it means you’re not extracting much value from AI. Which means you may not have much of a moat to protect you from competitors or from your users just talking directly to ChatGPT themselves.
This is all to say that if you are using AI, you’re probably using a lot of AI.
The more users you add, and the more they use your product, the more inference grows. As a cost it can quickly surpass your cloud infrastructure costs. Updating a few database tables, running a bit of code and sending a bunch of data to a React front end costs nothing compared to a 400B+ parameter model chewing through an 8k word user request and responding with 1k of JSON.
Your AI-powered SaaS now has serious variable costs and your gross margins have been cut in half and are in the 40-50% range.
Also – flat rate subscriptions are dead. It’s too much like an all-you-can-eat buffet in an obesity epidemic. Anyone watching the AI coding assistant space would have seen the big names – Anthropic, Cursor – learning this the hard way.
How AI-powered SaaS are pricing their services
Pricing for an AI-powered SaaS is all about finding the most palatable way to pass on the cost of inference to your customers. This requires a mix of value and clarity. That clarity will depend on the sophistication of your users.
Usage based pricing
For example, SourceGraph’s Amp coding assistant sells credits – ensuring that all inference a customer uses is paid for, capping SourceGraph’s risk – and for individual subscribers all AI costs are passed through via these credits without markup. But their enterprise plans, which come with all the SSO, management, etc features enterprise requires, have a 50% mark up on those same costs.
This structure works great for getting customers in – there’s no steep initial fee or lengthy commitment – but you need customers who can learn quickly to use your product (or already familiar with the category) and can clearly judge the value they are getting from your product.
Subscription plus Usage
The next step up from this is to have a base subscription that provides a basic feature set or limited access, and give the customer an option to buy credits for AI powered features.
OpenAI uses this for some features of ChatGPT, and AI image generators also work similarly, where a subscription gets you X images/seconds of video each period and you can purchase credits if you want to do more.
Outcome based pricing
The two previous pricing strategies have one issue – unpredictable bills. Some customers will not accept that. And with AI, which is by its nature non-deterministic, two actions that look on their surface similar, e.g. debugging two different database queries returning wrong results, might result in radically different costs for each solution. Software developers might be able to live with this, but businesses with tighter margins and detailed cost targets might not
Which leads to outcome based pricing.
This is one of the more interesting models. It’s been in use by anti-fraud services for years, but companies like Intercom and Zendesk and Salesforce have started using this pricing model. Intercom offers a fixed price per ticket resolution, Zendesk for each successful autonomous resolution, and SalesForce offers fixed prices per conversation and also per “action” (with conversations being charged at $2 and actions at $0.10).
This model makes sense. Pre-AI, SaaS was all about software supporting work – streamlining workflows, simplifying data and process management. AI-powered SaaS now has software doing the work. Outcome-based pricing makes sense under this new paradigm.
And for established industries with clear cost targets, this model makes it easy to communicate value from the start.
It’s not the simplest though. You do take on some risk. Your AI stack – the prompts, the models, the data handling – needs to work consistently within narrow parameters to make it profitable.
However, if it’s hard then it’s a moat. And if it’s a moat then it’s worth digging.
Where does AI fit in your business?
Are you at the MVP stage of an AI-powered SaaS? Or are you an established SaaS looking to incorporate AI into your offerings?
If it’s all in the backend to automate some decision making or triage or bucketing, then your concerns are more on optimising your AI stack – what’s the smallest, cheapest model that can be served the fastest to do what you need.
But if you’re exposing AI-powered features to your customers, we hope this article has given you the basics of pricing that you need to pick the right path forward.
