AI Agents Are Breaking SaaS Pricing. The Real Question Is What We Should Pay For.

For years, software companies had a wonderfully convenient business model: count the people using the product, charge for each seat, and call it a day.

Then AI agents showed up. And suddenly, one “user” might do the work of an entire department. So what exactly are we charging for now?

That was the question I brought to the stage at Web Summit Lisbon 2026, where I joined Joshua Wöhle, co-founder and CEO of Mindstone, and Edith Harbaugh, co-founder and CEO of LaunchDarkly, for a conversation about SaaS pricing in the age of AI agents. Web Summit published its recap on November 20, and the debate goes well beyond software invoices.

The seat is no longer the unit of value

Traditional per-seat SaaS pricing assumes that more human users roughly means more value. That assumption gets pretty wobbly when an AI agent can execute thousands of tasks without ever logging in like a person.

Harbaugh described how LaunchDarkly moved away from developer-seat pricing long ago, instead tying its model to service connections and the activity of customers’ end users. That matters because it aligns the bill more closely with the product’s role in the business, regardless of whether a human or an agent initiates the interaction.

Wöhle pushed the conversation toward value capture: if the software creates substantially more value than it costs, the model can work even when the customer has fewer traditional users.

Sounds reasonable. But there’s a catch.

Usage is easy to count. Outcomes are harder.

AI vendors can measure tokens, API calls, requests and completed steps. What they can’t always measure is whether the customer actually got what they wanted.

I raised the problem of paying for an AI system that burns resources without finishing the job. It’s a bit like a contractor billing you for every swing of the hammer, including the ones that miss the nail.

That distinction matters enormously. An AI agent that generates ten drafts but never produces an acceptable deliverable has created activity, not necessarily value. And when businesses start paying for agents at scale, the difference between the two becomes a CFO-level issue.

Wöhle made a useful counterpoint: early experimentation needs room to breathe, so usage can be a practical starting metric before businesses have enough evidence to measure ROI. I agree there’s a progression here. But the destination should be clearer business outcomes, not simply more machine activity.

The overlooked product is trust

The panel also explored an increasingly popular idea: if AI makes software easier to build, will software itself become nearly worthless?

I’m skeptical of the extreme version. Writing code and operating dependable software are two very different things. Harbaugh highlighted the ongoing importance of security, reliability and controlling AI agents that drift or fail without announcing it.

And that is where I think the next generation of software pricing gets interesting. Businesses may pay less for access to a tool and more for the assurance that the tool actually works, stays within guardrails and produces an outcome worth paying for.

In other words, the product isn’t just the software. It’s the accountability around the software.

What SaaS leaders should rethink now

Instead of asking how many seats an AI-powered product can sell, vendors should ask three tougher questions: What measurable result does the customer receive? Who absorbs the cost when an agent fails or retries? And what controls give the buyer confidence to let the system act?

The next SaaS pricing model probably won’t be one-size-fits-all. Expect hybrids that combine predictable access fees, usage allowances and outcome-based components. But the guiding principle should be simple: charge for value customers recognize, not activity machines generate.

Because when an AI agent can do the work of ten people, charging for ten imaginary seats is not a business model. It’s nostalgia with an invoice.

Read Web Summit’s recap of our discussion. The organizer notes that its recap was AI-generated from a transcript, so this article reflects my interpretation of the discussion rather than a verbatim transcript.