Aaron Fleishman, Partner
Back in the days of early SaaS, pricing was pretty simple: Customers used workflow software and were charged a license per seat, monthly or annually. Annual recurring revenue (ARR) was the core measure of software companies and per seat subscriptions were the backbone of ARR.
Then came the AI and LLM wave that started a couple of years ago. At first, incumbent software companies decided to add a “copilot” to their existing products, an addition to a workflow system which is effectively a little chat-based agent that workers use in their workflow to get things done more quickly; a little helper, that is charged as a new SKU and added to the license per seat cost.
At $10 – $30 a month per seat though, many of us saw that this wasn’t going to be an effective model for the long term, especially when reasoning capabilities entered the LLM era. We started going from copilot to pilot (aka agent). What started as an LLM agent working on tasks for or with me is now moving toward the agentic model where the LLM agent is working autonomously on my behalf. An autonomous agent is different than a copilot agent in that it does not scale linearly with users, but in fact, does the opposite: the autonomous agent actually reduces users. It’s a virtual employee or teammate.
Measuring value
An example of this reduction in the security space is Crogl (a Tola portfolio company), which acts as a security analyst that runs investigations around breaches and threats for enterprises. With the labor shortage for cybersecurity talent, such an agentic solution gives Chief Information Security Officers (CISOs) the capacity to triage the expansive volume of threats and dynamically scale as they need to keep pace with an ever-changing threat landscape.
An LLM agent can run an entire investigation in less than a minute whereas a human will take a couple of hours, and in some instances, the agent is doing work that no one wants or is unable to do.
Key Questions: What is the value of this LLM agent? Is it the equivalent of what you would pay a human? Probably not today, but how do you charge for this?
In the SaaS days, if someone wasn’t using the software, we’d re-assign the license or focus on fewer seats in the software renewal. But now, we’re focused on trying to measure value of AI software working 24/7, across functions and entire companies.
Pricing models
While we’re still in the Wild West for how to charge for AI functionality, a number of models are arising that are moving away from per-seat pricing.
Consumption-based pricing: Charging by execution or usage
While consumption-based pricing is not new, up until now, it was largely used in the cloud infrastructure space (Snowflake, Databricks). Salesforce Agentforce (priced at $2 per conversation) is leading the way for enterprises by taking this cloud pricing concept and normalizing it in the AI realm. Other examples of software companies using consumption-based pricing that relate to charging per output include CaseMark, which charges per deposition or case summary; or the Microsoft Copilot Chat pay-as-you-go plan, that charges a per message rate with an option for a discount on a bulk package.
While this model is flexible, the issue is that while the vendor is charging say $2 per conversation, the actual cost of that conversation varies substantially depending on the specific workflow, what models are being used, how many tokens are required, and so on. A single customer conversation could in theory cost Salesforce two cents or five dollars. The former would yield a 99% gross margin while the latter would put Salesforce in the red. Such consumption style pricing could lead to unfortunate unit economics for businesses who might lack the visibility or telemetry to know their underlying cost structure.
Value-based pricing: Focusing on outcomes and business impact
Intercom and Zendesk now have products that charge customers on a per AI resolution basis, promising that their customers will be able to measure success rates of AI tasks and quantify the value of their investments.
Questions come into play when paying into a value-based based model. What is a successful outcome? Rules for a successful customer service resolution may be much easier to measure than say, a CRM outcome. What is the cost to deliver an outcome?
This suffers from similar issues as the consumption model in that the vendor may not be able to proactively predict the cost of any given execution or outcome. Time will tell if Intercom Fin AI Agent can continue to charge $.99 per resolution or if that price point is unsustainable.
For customers, value-based pricing may be the simplest to understand and measure, but it will be challenging for AI providers across functions and domains to deliver effectively.
Cost-plus pricing: The actual cost of delivering AI services, with a markup for profit
Another potential pricing model is cost-plus which ensures that businesses cover operational costs while generating revenue. This model is great for the company offering services – who wouldn’t want a guaranteed profit on cost? – but maybe not so great for customers who may have less insight into what they might be spending than with consumption-based pricing. Nobody wants a surprise bill at the end of the month. I don’t see this model lasting long, but it is one that is currently in the mix.
Hybrid models: Combining subscriptions with usage-based add-ons for AI services.
The model that seems most interesting is a hybrid– a combination of consumption and value models – and I think these make the most sense for customers. Replit, the AI app builder, has a core service which includes access to their unlimited public and private apps along with the ability to host and deploy, with a number of fixed credits for additional features like agent checkpoints, all for a single monthly rate. Pay-as-you-go for additional features ensures that customers can scale as needed while also understanding their costs in the short and long term.
Hybrid models align well with customer value, while still mapping to some extent with the underlying cost of serving these solutions. That said, tooling and telemetry will be required to do this well.
AI cost modeling solutions
At Tola, we believe that assessing the value of AI is a cross-functional challenge , which is why we were excited to be a seed investor in Pay-i, a company that is helping enterprises understand the cost to service their AI products. AI costs are variable compared to traditional SaaS costs because of diverse AI model providers (e.g. closed vs. open source), modalities (e.g. text, image, audio) and billing structures.
While Pay-i is not solving the pricing problem itself, they are surfacing the data and inputs required to make better pricing decisions for their GenAI products.
The future of AI pricing
When we’re out of the Wild West, some things around AI and pricing will be clear. Better tracking systems for usage and value and standardized metrics to improve transparency are needed for both AI products and those they serve. The potential of AI FinOps platforms like Pay-I to help businesses manage cost allocation is a start towards a culture of shared accountability among Product, Finance, and Engineering teams. And most importantly, more predictable pricing models with clearer value assessments will allow customers to better determine how to move forward with their AI strategies.