How to Price AI In SaaS

Start by understanding how much differentiated value your AI features are creating

By definition, these foundational models are accessible to every SaaS provider, so how should you think about pricing what is, in effect, a commodity you’ve integrated into your product? Start with first principles: how much differentiated value does this AI feature create? 

By integrating AI features into the flow of your broader platform, you are saving the user time from having to leave their flow to go to the underlying model (ChatGPT, etc). Keeping the user in context can be a powerful unlock (see our thoughts on Deep Collaboration here). 

However, be honest with yourself as to how much value your AI is actually creating. Many AI features in SaaS today are getting a flood of initial tire kicks from curious users but aren’t seeing meaningful sustained adoption. Start by understanding retention and value creation.

Then ask yourself how differentiated your AI offerings are. If the majority of the value your AI feature creates can be garnered by going directly to ChatGPT, don’t try to make a significant margin on that feature. Reselling is not a sustainable value-creation strategy (nor a differentiation strategy, though that’s a topic for another post). 

Even if you aren’t able to charge much for your AI features today, they can create meaningful value by making your current product more valuable and perhaps stickier. They can also be used to drive upsell to higher tiers, all of which can result in increased net dollar retention.

Over time, you can leverage initial features that may today just be a thin wrapper around a third-party model to build more differentiated value (more on how below). And when you get to that point, you can consider a more value-extractive pricing approach.

The state of AI pricing in SaaS today: early days

As you launch your AI features, your priorities should be to learn about how/where they add value and to drive user adoption. To that end, many SaaS companies are starting by offering their AI features for free. The leaders we’ve chatted with behind these strategies envision monetizing over time as they learn more about how much value is being created and for whom.

That said, some SaaS providers have already begun to charge for their AI features. Most start with a freemium approach typically gated on usage volume (not features) so the user gets a full taste of the feature’s power before seeing a paywall. The pricing models they use prioritize simplicity, which helps drive adoption. GitHub CoPilot and NotionAI, which have both seen strong adoption, charge on a per seat basis. 

The aforementioned SaaS providers charge an incremental fee for their AI features on top of their core product offerings. There are also SaaS providers whose entire product is focused on providing AI-enabled services. Companies like Jasper and Copy.ai leverage OpenAI to deliver copywriting as a service. They charge higher prices per seat, though likely have a higher cost of goods sold (COGS) given their bills from OpenAI.

There’s an open question around how much of these foundational model COGS SaaS providers will be able to pass along to their customers over time. It may be the case that for those that provide truly differentiated value, they’re able to pass all of these costs along (and add a healthy margin). Those that provide more commoditized value may not be able to and will thus have lower gross margin profiles than traditional SaaS providers (70-80%). These players may be able to leverage AI to improve their internal operations (e.g., automating major parts of sales and marketing) and recover some of the gross margin loss in their operating expenses. 

How SaaS companies should think about COGS for their AI features

The pricing offered by underlying model vendors is rapidly evolving; not just price level but approach (some charge on a token basis, some on compute, etc). SaaS companies should assume that, with more model providers and better technology, cost of inference will continue to decline over time (though there may be some near term stagnation given GPU supply constraints).

In our portfolio of SaaS companies, we’re seeing that most AI features being integrated today (summarization, editing, text generation, etc.) aren’t breaking the bank. Thus, most SaaS players aren’t gating usage for their paid AI features though they do reserve the right to throttle for certain extreme users (providing intentionally vague “fair use” policies).

As you consider COGS mitigation strategies over time, one important approach will be optimizing which underlying model you use for a given task, trading off cost and performance. As an example, very few portfolio companies are leveraging GPT-4 today given cost and latency, despite its tremendous power; GPT-3.5 Turbo is sufficient for their use cases. While these model-specific points will evolve quickly, the core idea of choosing the right model for the job will likely become the dominant approach SaaS vendors take to delivering AI features. As open-source models in particular continue to rapidly improve, we think this dynamic will get even more favorable for SaaS vendors.

Driving towards differentiated value creation

Ultimately, your goal for your AI features should be to have them drive differentiated value. You can start with a simple integration of a third party model, but you should be using this approach to capture proprietary data specific to your “job to be done” (JTBD). You’ll want to track and correlate which suggestions are surfaced by the model, what the user does with the suggestion, and what the business outcome is. You can then use this proprietary data to tune your model, ultimately delivering JTBD-specific value that can’t be obtained from a generic third party model. At Emergence, we call this approach a Coaching Network.

Ironclad is a good example here. Ironclad provides a digital contracting platform, on which their customers can draft, negotiate, and sign contracts. They recently integrated a feature that allows customers to get AI-enabled coaching as they redline contracts. Ironclad started with an integration into OpenAI’s models, but is now gathering valuable data on what their users do with the model’s suggestions, and critically, what the business outcomes are, e.g., which redlines drive faster contract close timelines. As they integrate this data into their AI features, they’ll be able to deliver insights that others using generic third party models won’t be able to. And they’ll be able to price accordingly.

As SaaS companies move toward this nirvana state, I suspect many will move away from per seat pricing. Ultimately, as discussed in my 2019 post, this approach may limit data gathering and ultimately cannibalize the SaaS provider as its AI features improve and make more user seats less necessary to drive the same value. SaaS companies will likely move toward more value aligned models at this point. As the industry gets there, we’ll share more thoughts on what we’re seeing.

In the interim, SaaS companies should be solving for simplicity and adoption in their AI feature pricing. This is a time for learning and iteration. 

We’re all learning quickly as the space evolves at blinding speed. Feel free to reach out (jake@emcap.com) if you’re building an AI-enabled SaaS product and want to jam on these ideas.

An earlier version of this article was featured in TechCrunch+.