Loading...

Loading...

AI needs new pricing. Customers need a reason to accept it.

• 7 minute read

AI has made software pricing a lot more complicated. Charging per seat, the default for SaaS, is a poor fit. Unlike traditional software, AI has a real cost every time it runs, so heavy users cost far more to serve than light ones. And when AI does work a person used to do, the value no longer scales with the number of seats. Gartner expects at least 40% of enterprise SaaS spend to move to usage-, agent-, or outcome-based pricing by 2030 [1].

This is all part of the wider shift towards paying for the work software does, instead of paying for access to it (as I covered in a previous article, Monetising the agentic layer). But there isn’t yet a “right” answer, and companies are still experimenting. Many are aiming for outcome-based pricing, where the customer pays only for a result and price lines up most closely with value. Others see hybrid models, typically a platform or seat fee plus a usage charge, as the pragmatic middle ground. And it seems every other company has introduced credits for AI usage, though no two define a credit the same way.

None of this has settled yet, and more change is coming. As pricing models become more closely tied to consumption, customers are being asked to accept more uncertainty. It shows up in three questions they keep asking:

  1. What will I pay?

  2. How do I know it works, and that it’s worth the cost?

  3. Will you change your pricing?

At Hg, we keep a close eye on how software companies are answering these questions and turn what we learn into action for our portfolio. In this article, we share some of those learnings, looking at what companies are doing in practice and what is and isn’t working.

What will I pay?

Customers rarely object to paying for usage. What they object to is a bill they can’t predict. Vendors are tackling this in three ways: upfront commitments, visibility and spend controls, and a choice of how to buy.

Upfront commitments are the most common. The customer buys a bundle of usage for the year, draws it down as they go, and tops up if they need more, much like a phone contract with a data allowance. For example, HubSpot includes credits with its subscriptions and sells more in packs, all drawn from one pool shared across its AI agents [2]. A commitment turns variable usage into a predictable cost that customers can put in next year’s budget. But customers still have to guess how much they’ll need. Guess too low and they face top-ups they didn’t budget for. Guess too high and they pay for usage that goes to waste. Clay, the AI go-to-market platform, tackles the latter by letting monthly customers roll over unused data credits, up to twice their monthly allowance, so a wrong guess costs less [3].

Even with a commitment, customers can’t be sure how quickly they’ll burn through their allowance. They want to see their usage as it happens and set caps so they don’t overspend. Workday alerts administrators at 80%, 90%, and 100% of their credit limit [4]. Cognition, maker of the Devin coding agent, offers more granular controls, letting companies divide one account-wide limit into department budgets in Devin Enterprise. Department heads can then split their budget across teams and projects and cap spend for individual users [5].

Some vendors go further and allow customers to choose the price model that best suits them. Salesforce sells Agentforce four ways: per conversation, per action through credits, per user under enterprise licences and, since July, per resolved customer issue [6]. Decagon offers both per-conversation and per-resolution pricing, and says most customers pick per-conversation to avoid disputes over what counts as a resolution [7]. ServiceNow keeps both doors open. Around half of its new business is no longer seat-based, but it still sells seats to customers who want them. CEO Bill McDermott said customers value the predictability of seats, “particularly now where a lot of pricing out there has been less than predictable” [8].

How do I know it works, and that it’s worth the cost?

A predictable bill is only part of the equation. As agents take on work that customers’ own staff used to do, the next question is how well they’re doing it, and what that’s worth. Most companies are yet to put a number on it. In Q2 2026, just 11% of S&P 500 companies quantified how AI productivity affected a specific business use case, and only 2% quantified its impact on earnings [9]. One answer is to tie the price to the result, so the customer pays only when the work gets done and the vendor carries the risk that it doesn’t. This is most established in customer support. Fin (formerly Intercom) charges from $0.99 for each conversation its AI agent resolves [10], Zendesk bills per automated resolution [11], and Sierra says customers “pay for results, not seats” [12].

But outcome-based pricing is hard. It only works when vendor and customer agree on what counts as an outcome, and the customer can check that it happened. Even in customer support, there’s no agreed definition of a resolution. Fin counts a conversation as resolved if the customer simply leaves without asking for more help within 24 hours [10]. Some customers have pushed back, arguing they’re being billed for conversations Fin didn’t actually resolve. Zendesk sets a higher bar and bills only for resolutions it has verified. Even so, customers weren’t always clear what they were paying for. So in May 2026, Zendesk began reporting verified resolutions separately from “contained” ones, where the agent handled the request but the outcome wasn’t confirmed, and no charge applies [11].

Outside customer support, outcomes are even harder to define, so vendors are finding other ways to show the product works before customers commit. Microsoft’s Dynamics 365 case management agent has a shadow mode that runs on live cases, drafting replies and proposing updates without sending or saving anything, so customers can compare what the agent would have done with what their staff did [13]. Other vendors absorb the cost of proving it. HubSpot has offered free trials of its agents and accepted slower buying cycles as the trade-off. Cognition goes further with a guarantee, offering enterprise customers up to $10 million of free usage if its Devin agent delivers less engineering value than they pay for [14].

Will you change your pricing?

Almost certainly, and more often than customers are used to. AI products are still evolving, and companies are still learning what works, so their pricing will keep evolving too. Customers would prefer stability, but they can accept change if it’s handled the right way. That comes down to two things: whether the change is fair, and how clearly it’s communicated. Cursor learned this the hard way. In June 2025, changes to its Pro plan left users with large, unexpected overages that many only discovered when the bill arrived. Within three weeks, Cursor had apologised, refunded the surprise charges, made it easier to see on its dashboard when users were nearing their limits, and promised advance notice of future changes [15].

In contrast, Clay shows what doing this well looks like. The pricing model it launched in March 2026 was more complex than the one it replaced, but Clay made design choices that put customers first. Existing self-serve customers could keep their current plans with no end date, and Clay expected 85% of customers using its AI features to spend less under the new model [3]. It also went well beyond the standard email announcement. It published the internal memo behind the decision [16], its founders walked through the change on video, and it contacted customers who would save money by switching, a move it expected to cost around 10% of revenue [3].

Trust is what makes the pricing stick

Most of the debate about AI pricing centres on the model and the price point. Getting those right is only half the job. The other half is earning customers’ trust. Alerting customers before they hit a limit, verifying a resolution, letting them watch an agent work before it acts: none of this changes what a company charges. It changes what the customer can see, check, and control.

All of this comes at a cost. Usage dashboards and spend controls take engineering time, billing only for verified resolutions leaves revenue on the table, free trials slow down sales cycles, and guarantees put real money at risk if the product falls short. The right mix depends on the product and the customer. But building trust upfront is usually cheaper than winning it back later.

And it isn’t one team’s job. Product and engineering build what customers see and control, business systems make billing verifiable, go-to-market explains a change before it lands, and finance tracks the numbers that show when the pricing needs adjusting.

That’s why at Hg we treat AI monetisation as a cross-functional effort. When we support a portfolio company, our Catalyst team brings together experts in pricing, product, go-to-market, finance, and business systems, so the work needed to make the pricing land is planned alongside the model itself. Across the portfolio, we also share case studies and bring leaders together to swap lessons on what’s working. The clearest lesson so far is that choosing the model is only the start. Whether the pricing sticks depends on everything built around it.

Sources

  1. HubSpot, Product and Services Catalog (HubSpot Credits): https://legal.hubspot.com/hubspot-product-and-services-catalog

  2. Clay, Introducing Clay’s new pricing: https://www.clay.com/blog/introducing-clays-new-pricing

  3. Cognition, hierarchical budgets: https://devin.ai/blog/hierarchical-budgets

  4. Salesforce, Agentforce pricing: https://www.salesforce.com/agentforce/pricing/

  5. Decagon, Pricing AI agents: https://decagon.ai/blog/pricing-ai-agents

  6. ServiceNow, Q2 2026 earnings call transcript (Fortune): https://fortune.com/company/servicenow/earnings/q2-2026/

  7. Goldman Sachs Q2 2026 earnings analysis, via Investing.com: https://www.investing.com/news/stock-market-news/when-will-ai-start-boosting-corporate-earnings-goldman-sachs-weighs-in-4862480

  8. Microsoft 365 Message Center: https://mc.merill.net/message/MC1296468

  9. Cognition, AI Productivity Guarantee: https://cognition.ai/blog/ai-guarantee

  10. Cursor, Clarifying our pricing: https://cursor.com/blog/june-2025-pricing

  11. Clay, The thinking behind our new pricing: our internal memo: https://www.clay.com/blog/clay-pricing-memo-internal

The views expressed are Hg’s own as at the date of publication and are subject to change without notice. This article is for general information only and does not constitute research, investment advice, or an offer or solicitation to buy or sell any financial instrument, fund or product in any jurisdiction. Certain information has been obtained from third-party sources and has not been independently verified; references to third-party companies are for illustration only and do not constitute an endorsement. Hg accepts no liability for any reliance placed on this article. Hg Pooled Management Limited is authorised and regulated by the Financial Conduct Authority.

Share this article