How AI Agents Are Eating SaaS: The End of Seat-Based Software?

Your bills get higher every time you add someone to the team. Another CRM seat, another project-tool seat, another helpdesk seat, and the monthly bill climbs by hundreds. Much of that pays for software those people open twice a week. For two decades that was the cost of doing business. Building your own tools was too slow, too expensive and too risky. That maths no longer holds. It raises a real question: are AI agents replacing SaaS, or just making it cheaper to walk away?

The short answer is some of it, yes. Agents now do work that used to need a logged-in human. Software has also become cheap to build. Once building gets cheap, paying per seat for a generic screen over your own data starts to look odd.

What is actually breaking, and why

Seat pricing rests on two assumptions. First, that people are the unit of value. Second, that building your own software is too expensive to bother. AI broke both at once. Agents now do the work of people who used to be licensed users. A team can build, in days, an internal tool that once took months and $100K. Remove both assumptions and the per-seat logic stops making sense for commodity software. That is the mechanism behind AI agents replacing SaaS: agents do the work, and the tool underneath is cheap to rebuild.

This is not another “SaaS is dead” headline. Plenty of software will not change at all. What is changing is the pricing model, and beneath it, the build-versus-buy decision for one slice of software. The market is already pricing this in. Deloitte expects the share of software revenue from per-seat licences to keep falling through the rest of the decade, while usage, agent and outcome models take up the slack. Gartner points the same way on enterprise software spend by 2030. Treat these as projections, not facts. The direction, though, is no longer seriously debated.

Seat compression is not SaaS replacement

When people talk about AI agents replacing SaaS, they usually mean one of two different things. Seat compression means you cut licences because a smaller team, helped by AI, does the same work. SaaS replacement means you stop paying for the tool entirely, because you built your own. Compression dents the vendor’s revenue. Replacement takes it to zero. Most real cases today sit in the middle: smaller user bases, propped up by agents, rather than a clean replacement.

The Salesforce example

Jason Lemkin of SaaStr is the most cited case. By his own account, he went from more than ten human Salesforce seats to two human seats plus one API seat. His spend still rose, from about $12k a year to $22k, up 83%. The reason is usage: 20-plus agents now drive roughly 100x the load on the same data. The seats compressed, but the bill grew. Value moved from humans to the system doing the work.

Hold onto that, because it kills a naive assumption: that AI cuts costs on day one. Sometimes it does. Often you simply pay for usage instead of seats, and the total goes up. The bottom line may not move at all, even when the system gets easier and people are happier.

At enterprise scale, replacement also takes far longer than the headlines suggest. You have to meet regulatory requirements, keep existing integrations alive and grind through procurement. Moving to a new system can easily span several years.

Which SaaS are AI agents replacing first?

AI agents are replacing SaaS fastest where the tool does the most obvious thing: create, read, update, delete. A CRM that is really just a contact book. Project apps that are really just statuses and comments. Basic help desks and ticketing. Outreach and lead-scraping pipelines. Internal reporting dashboards. Their whole pitch was a tidy front end on data you already own. That makes them the cheapest category of software to rebuild.

One question sorts most tools. Is the value in something specific the tool does, or just in the screen it puts over your own data? If it is mostly the screen, the tool is exposed. That is the pattern of AI agents replacing SaaS in practice.

This is not hypothetical. Retool’s 2026 Build vs Buy report is blunt. 35% of organisations had already replaced at least one major SaaS tool with an in-house build last year. Another 78% expect to build more during 2026. Netlify, as Business Insider reported in June 2025, swapped several SaaS tools for custom internal apps: an employee feedback survey tool, a pricing calculator, an interview-training app. None of these are moonshots. They are the boring, expensive, generic tools a team decided it no longer wanted to rent.

The pricing model is changing under everyone

Even vendors admit the old model is getting harder to defend. That is why so many are quietly changing how they charge. AI agents replacing SaaS is only half the story; the other half is how the survivors re-price. The new shapes look like this.

• Usage-based models: charge per request, task, ticket, or document processed.
• Outcome-based models: you pay per ticket resolved, qualified lead, or completed workflow, rather than per user.
• Agent credits: you buy a pool of work and draw it down by how hard each task is.
• Pay-per-resolution and pay-per-workflow: common in support and operations, where the unit of value is clear.
• Hybrid: a base platform fee plus a usage or success layer. This is where most of the market is landing.

The living examples are easy to find. Salesforce launched Agentforce at roughly $2 per conversation in late 2024. By 2025 it had moved to 10 cents per action on a Flex Credits model, with hybrid plans now sitting alongside seat-based subscriptions. Zendesk charges $1.50 to $2.00 per automated resolution. Intercom’s Fin agent costs $0.99 per resolved conversation. These are not obscure startups. They are established players changing how they charge, because per-seat no longer matches the value they deliver.

Buy the vendor’s agent, or build your own

Sooner or later you choose. Do you buy the agent your vendor is selling, or build your own? The honest answer depends on one thing: is the workflow generic, or is it really yours? This is where the trend of AI agents replacing SaaS gets specific to your business.

When buying the vendor’s agent makes sense

If you are already deep inside a platform, buying its agent makes sense. You get AI on top of the same system, with no re-architecting. The vendor supports it from day one. You know where governance sits and how the agent connects. Just understand why the vendor’s agent exists. Often it does not solve a new problem. It props up a per-seat price that is getting harder to defend, and it keeps you on the platform.

When building your own wins

Building wins when the process is genuinely yours. A platform cannot fit itself exactly to how you work, because it serves everyone else too. Salesforce users know this already. Salesforce has to serve a million-plus companies with different needs, so it stays general by design. You do not need a general system. You need one specialised tool for the one process you run every day. Build for that single use case and it can be cheaper to host, faster to ship and easier to manage than a platform built for everyone.

The caveats are real, and worth saying out loud. The hard part was never the prompting. It is deciding what to build in the first place. Nobody wants a pile of custom tools that break at 3am because a model or API changed underneath them. So treat building an agent as an engineering decision, not a weekend project. A well-defined custom AI agent, scoped and owned by your team, is the difference between real software and a cron job running a few prompts.

What to do with your stack right now

Do not rip anything out yet. Start with an audit. For each tool, work out the cost per seat, how many people really use it, which features get touched, what workflows depend on it, where the data lives and which integrations matter. Then sort the stack into three groups: keep, add an orchestration layer, or replace. An audit tells you where AI agents replacing SaaS actually pays off, and where it does not.

Here are the three groups:

• Keep: anything vital, intricate, compliance-bound, or deeply woven into your infrastructure. Your system of record usually lives here, so leave it untouched.
• Add an orchestration layer: tools you still cannot replace, but that agents can move work to and from, instead of a person copy-pasting.
• Replace: the simple tools that are really just a database, a board or a dashboard with a login in front.

The split matters. In practice, most of a stack belongs in the first two groups. The interesting part is usually one or two tools in the third: expensive, and delivering little. Start there, with one high-value workflow. You may end up building a custom internal application that does nothing but that one job. That is what strong engineering teams do anyway. Sometimes the honest answer is “do nothing”, or “a simpler tool will do”, or “do not use AI for this step at all”. Keep all three on the table from the start.

The real costs, ROI and failure modes

The savings and the costs

The economics have two sides: savings and costs. On savings, the direct line is lower licence cost, fewer users and fewer seats. The indirect savings are usually bigger: less manual work, less tool-switching, lower governance risk, and a tool you can reconfigure the day your process changes.

Costs are real too. Quick to build the first time does not mean cheap. You carry development and maintenance, plus hosting, integrations, security and human oversight on sensitive calls. A simple tool that replaces a basic CRUD app can take a few days. A production agent, with scoped permissions, guardrails and monitoring, is more like six to ten weeks for a good team. Anyone promising a couple of days is selling you a demo, not a production agent.

The failure rate nobody quotes

Then there is the failure rate, which the hype skips. Gartner expects companies to scrap more than 40% of agentic AI projects by the end of 2027, on cost, weak value or governance. Klarna is the cautionary tale. It consolidated data, cut tool usage and announced an AI assistant that replaced 700 support agents. The win did not last. A year later the company walked it back and rehired humans for the cases the assistant handled badly. Technology was not the problem. The mistake was treating AI as a moat rather than an assistant. Hand judgement-based work to a full replacement and you tend to break the thing you set out to improve.

What reliable actually means

So what does reliable mean? Only about 1 in 10 teams run a true end-to-end agent with zero human oversight. The maths is unforgiving. At 95% reliability per step, 20 steps give you a 36% chance of clean success. That is why we scope agents to 5 to 10 steps with human checkpoints, instead of shipping autonomous swarms. It is also why an agent that genuinely moves the needle lands in the £5,000 to £10,000 range, not a weekend hack. In production, cheap and brittle is the most expensive option the moment it fails.

Where this actually goes

This does not end as SaaS versus agents. The endgame is not AI agents replacing SaaS wholesale. It is a mix. Some software stays because it is genuinely hard to replace. Some becomes custom internal tooling. Some gets woven together under an orchestration layer, so work no longer needs hand-stitching by people. Some workflows go partly or fully agentic. Over time you are really assembling a company OS: your data, workflows, agents and people in one tunable layer. You shape it around how you work today, and adjust as you grow. A per-seat platform built for everyone cannot sell you that fit. That is why the generic, commoditised corner of software feels the squeeze first. The defensible products, the ones with proprietary data, network effects, compliance or deep integration, are not going anywhere.

The honest first move

The renewal question is not “should we cancel all our SaaS”. It is sharper than that. Where does build-versus-buy actually pivot for us, and which one or two workflows are worth owning? The honest first instinct is not to build. It is to audit. A good audit ends with you keeping almost everything, except the part whose worth you cannot defend. That is the work we do, and we run it on our own stack too. We have AgentWise, our own AI agent, in production, and we have built more than 50 AI features for real businesses. If the per-seat charge for tools you barely open keeps climbing, that is the conversation to have. Sometimes the answer is to stand pat. Often it is to own the one system your whole company runs on.

Frequently asked questions

In practice, how many seats can one AI agent replace, once you have excluded the exceptions?

It is not really about seats, since that varies in every case. Asking for one agent that does everything is like asking for a car that can fly and sail. Think in tasks, not seats. Which steps can you automate, move to a cheaper internal layer, or drop entirely? In high-volume, low-variation work, a well-scoped agent can take a real bite out of seat count. Think first-level support or data entry. In judgement-heavy work full of exceptions, it mostly assists, and the human stays in the loop.

We are a scaleup of around 40 people. Should we build custom agents, or buy from a vendor and layer on top?

Usually both, in sequence. Keep the platforms that are hard to replace, and layer AI on top of them first. Then find one task where you pay a lot and get little, and build a custom app for that. Rebuilding the whole stack at once is how scaleups burn quarters they cannot get back.

What does a reliable 24/7 agent actually cost per month versus our current SaaS stack?

It depends on the workflow, the volume and how often you hit expensive models. The token and cloud bill is only part of it. Most of the cost is building the agent properly, then maintaining, monitoring and securing it. For a simple internal app that replaces 10 to 20 seats, it is often cheaper to run than the licences. In complex cases, do the real calculation rather than trust an estimate.

How do we stop an agent burning a huge volume of tokens?

Put guardrails in before you switch it on. Enforce budgets. Set per-workflow caps. Alert on spikes. Add an emergency stop that pages a human. Meter and bill from the same layer that runs the request, so there is one source of truth. The horror stories are rarely about usage pricing. They happen when an agent gets an open scope and no guardrails.

If we build, who maintains it when models or APIs change?

You do, or whoever built it under a support agreement. Models change, APIs shift and prompts drift. So a proper build ships with monitoring, tests against known cases, and a fast way to roll back. This is the part the marketing skips. It also decides whether your tool is an asset or a liability a year from now. If nobody owns maintenance, keep the agent away from anything mission-critical.

Can we use the arrival of AI to reduce seats or get credits from our current vendors?

Often yes, and it is worth asking. As seat compression shows up, some vendors will discuss credits, hybrid contracts, or a move from pure per-seat to usage-based pricing. Renewal is the moment. Do not frame it as a discount request. Explain that your usage model is changing and you want a contract that reflects it. Bring your own audit numbers.

For UK and EU companies, how do you keep agents that interact with customer data GDPR-compliant?

If an agent reads emails, calls or recordings, it touches personal data. You are almost certainly in scope for UK GDPR. Define exactly what data the agent can reach. Enforce permissions by role and workflow. Log every action and keep a full audit trail. Decide where data goes and which models see it. A smaller, owned tool is often easier to govern than a sprawling one, not harder.

Which should you build first: a high-ROI workflow or a broad, general multi-agent system?

Start with a single workflow. Pick one that is frequent, repetitive, well understood, low risk and expensive in licence spend or human time. Ship it. Watch how it behaves in production, then expand. Going broad too early is where projects stall, and a big reason so many agentic projects get cancelled. Not sure which workflow to pick? That is exactly what an audit is for.