The SaaS contract used to be simple: we give you the tool, you supply the labour. That contract is breaking. The vendors winning the next decade will sell the tool and the labour as one product, and they'll be paid for output, not for seats.
For twenty years, the per-seat model worked because seats were a reasonable proxy for value. More people on the CRM meant more deals tracked. More agents on the support platform meant more tickets handled. Software value and human usage scaled together, so vendors priced for access and customers staffed the team that turned access into output.
That coupling has come apart. One person directing an agentic layer can now do the work that used to require a team. When that happens, seats stop measuring anything useful, and the question for vendors becomes whether they want to keep selling access to a tool or start selling the work the tool was always there to produce.
This is not yet true of every category. But it's true of enough of them that if you're building software now, it's worth taking seriously.
Selling outcomes instead of tools isn't new. Agencies, consultancies, and managed-service providers have done it for decades. What's new is who can afford to deliver it.
Until recently, packaging output required a services layer sitting on top of the software. People, billable hours, account managers, margin pressure. Software vendors stayed in their lane because crossing into services destroyed their gross margins. AI collapses that cost structure. A vendor can now deliver the work the software was designed to enable, instead of leaving that margin to the layer above.
The shift is measurable. Gartner is projecting that 40% of enterprise SaaS spend will move toward usage-, agent-, or outcome-based models by 2030, and the pure per-seat share of the SaaS market is in steady decline. Hybrid pricing (base subscription plus consumption) is now the dominant new model for AI products.
Three things made this possible. LLMs got cheap enough to operate as workers rather than features. Agentic tooling matured enough to chain real workflows end-to-end. And buyers, after three years of integrating AI into their operations, became comfortable enough to accept agent-executed work as legitimate output.
A modern software product built for this shift has three layers stacked together.
The first is the traditional software. The system of record, the workflows, the data model, the integrations. This part hasn't changed. It's still where the work lives.
The second is an LLM-native interface. Chat, natural language, document-in and document-out. Every credible SaaS vendor is shipping this layer. It's table stakes, not differentiation.
The third layer is where most vendors are still thinking too small. It's the agent layer: AI workers that operate the software on the customer's behalf, under human direction. Once it exists, the vendor isn't selling access to a tool anymore. They're selling the output the tool was always there to produce.
Two examples show the shape of this.
Intercom's Fin charges $0.99 per resolved support conversation. HubSpot dropped its Customer Agent to $0.50 per resolved conversation earlier this year. The buyer pays nothing for failed attempts or escalations. This is the cleanest version of the model. The vendor gets paid only when the work is done, and the unit of value is the resolved ticket, not the seat. The customer's support team is smaller, led by a human who handles escalations and trains the agent. The software has stopped being a workspace and become a workforce.
The second example sits closer to home. A 4D Pipeline customer in apparel has replaced traditional photography and rendering for product imagery with an AI rendering pipeline. The interesting part isn't the rendering itself. It's what sits around it. Agentic AI bots QA the output, checking that the garment fits correctly on the model, that the product is represented accurately, that the logo is placed and rendered as it should be. A human still leads the work. But producing thousands of on-brand product images, which used to mean studios, photographers, retouchers, and weeks of turnaround, now happens inside a software product that ships with its own workforce attached.
These two examples sit in very different categories (customer support and visual commerce) but the thread is the same. The customer was never buying the software for its own sake. They were buying the output the software produced. The vendors recognising this are now selling the output directly.
Three things change if you build for this.
The buyer changes, sometimes. This is more nuanced than the usual framing suggests. In some cases the buyer genuinely shifts. You're no longer selling to the head of the support team; you're selling to the operations leader who used to fund that team's headcount. Different budget line, different price point, usually a much larger one. But in other cases the buyer stays the same and what changes is their leverage. The CAD designer still buys the CAD tool, but instead of doing the bulk of the work themselves, they direct an agentic layer that produces several times the output. The marketing lead still buys the marketing platform, but the team beneath them is smaller because the platform scales their output directly.
The strategic question isn't "who's the new buyer." It's "what does my customer's leverage look like when my product ships with a workforce attached?" Sometimes you sell to someone new. Sometimes you sell to the same person but at a higher price, because their job has become more valuable.
The pricing model breaks. Per-seat assumed one seat correlated with one unit of value delivered. When one seat backed by agents produces ten units, that correlation collapses. What works instead is usage-based, outcome-based, or hybrid pricing: a base subscription plus consumption. None of these forecast as cleanly as ARR built on seat counts, which is why vendors raised on predictable per-seat revenue resist the change. The ones moving first are accepting messier forecasting in exchange for capturing the budget that used to fund headcount.
Liability shifts at the margins. More measured than the hype suggests. When the customer is still leading (setting direction, approving the work, owning the decisions), most of the liability stays where it always has. What changes is the vendor's exposure when the agent layer fails: bad outputs, missed cases, errors at scale. Contracts get teeth. SLAs start including outcome guarantees. The vendor takes on more risk than they did when they were selling a tool, but not all of it, because the customer is still in the chair.
Worth understanding before you ship, not after. Vendors who structure outcome guarantees they can't deliver will find the limits of the model the hard way.
Not every category fits this shift, and pretending otherwise weakens the argument.
The output model fits best where the customer's team is doing work they tolerate rather than work they enjoy. Compliance, reporting, procurement, tier-one support, data entry, basic analysis, high-volume content production. Work that exists because it has to, not because anyone wakes up wanting to do it.
It fits less well, and may not fit at all, where the doing is the point. Craft tools survive. Figma, Linear, Notion, the IDEs developers actually like, the design environments where the user's hands on the work is the value. In those categories, an agent layer is a feature, not a business model. The user doesn't want a workforce. They want a better instrument.
The honest question isn't "should I add agents." It's "is my customer's team doing work they enjoy, or work they tolerate?" The answer to that, category by category, tells you whether you're looking at a feature upgrade or a business model shift.
Three questions worth sitting with if you're building software now.
If you delivered the output your customers buy your software to produce, not the tool, what would that look like as a product? Not as a pivot from your current offering, but as a second SKU alongside it.
Who would the buyer be then, and is that a budget line you can credibly compete for? The budget that used to fund the team doing the work is almost always larger than the budget that used to fund the tool the team used.
What's the smallest version of this you could ship in a quarter? Not the visionary version. The smallest one that lets you start charging for output instead of access, on a narrow enough slice of your product that the risk is bounded.
The claim here isn't that every SaaS company becomes an agent company. It's that the vendors who don't seriously examine the output model are leaving a structural shift unexamined, and the ones who move first on the categories that fit will set the price floor everyone else has to work against.
The interesting question isn't whether the shift is happening. It's which side of it you'll be on.
#productmanagement #ai #SaaS #aiagents #positioning