SaaS seats are melting

The seat was never the product. It was a proxy — a crude, convenient stand-in for the thing you actually wanted to buy, which was work getting done. For twenty years the proxy held, because work came in exactly one shape: a human, logged in, clicking.

That assumption is now dissolving under the industry's feet, and with it the pricing model that built the modern software economy. The position of this essay: per-seat SaaS is structurally mispriced for a world where agents do a growing share of the work, the vendors know it, and buyers who keep signing seat-locked multi-year deals are financing the vendors' transition out of their own pockets.

Why per-seat won in the first place

Give the model its due — per-seat pricing was a brilliant fit for its era:

• Easy to meter. Counting logins is trivial. Counting value delivered is hard. Vendors picked the countable thing.
• Aligned with growth. In the headcount economy, a growing customer meant more employees, more seats, more revenue — expansion revenue without a single sales call.
• Predictable for both sides. CFOs could budget it; vendors could forecast it. Wall Street learned to worship the metric that fell out of it.

The unstated axiom underneath: the amount of software-work a company needs scales with the number of humans it employs. Nobody wrote that down as a bet, because nobody imagined it could lose.

The melt

An agent doesn't log in. It doesn't have a seat, a session, or a mouse. It hits an API a thousand times an hour, does the work of what used to be several humans-with-licenses, and shows up in no seat count anywhere.

So watch what happens to the proxy: a company that adopts agents seriously stops adding seats while consuming more software-work than ever. The correlation that per-seat pricing metered — humans in, value out — quietly breaks. First in the workflows where the work was most mechanical, then everywhere the work is structured enough for an agent to run it (we've mapped that frontier in Stop hiring for jobs agents can run).

The vendors are not confused about this. You can read the awareness directly off the industry's pricing pages: the sudden bloom of consumption pricing, per-conversation pricing, per-resolution pricing, "agent" SKUs, credits, actions, and outcome-flavoured experiments across the major platforms. When an entire industry starts inventing new units of charge at the same time, it is telling you — in its native language — that the old unit stopped measuring the business.

Per-seat pricing is a bet that work stays human-shaped. The vendors have already started hedging that bet. Have you?

What melts first — and what survives

The melt is not uniform, and pretending it is would be lazy. A rough sorting:

Melts first: seats that measure presence. Licenses whose daily reality is a human moving data from one screen to another — ticket triage seats, data-entry-heavy CRM seats, report-assembly seats, the long tail of "operations" logins. When the work is extraction, transformation, routing and drafting, an agent does it through the API and the seat evaporates.

Melts slower: seats attached to judgment. Approval authority, negotiation, design, accountability. A seat that maps to a person who decides and answers for the decision still measures something real — though notice how many "judgment" seats, audited honestly, are 80% data movement wrapped around 20% judgment.

Survives, transformed: systems of record and infrastructure. The database of truth — the ERP ledger, the source of customer records — retains gravity because agents need somewhere authoritative to read from and write to. But its pricing migrates toward what infrastructure already uses: usage, capacity, API volume. The system of record survives; its per-human toll booth doesn't.

What buyers should do now

Not in 2028. Now — because the contracts you sign this year will still be running when the melt reaches your workflows.

1. Audit seats against work, not against people. For every license tier, ask: what fraction of this seat's actual daily activity is mechanical data movement an agent could run? You are not asking to fire anyone; you're asking what you're paying per unit of judgment, which is the part that deserves a seat.
2. Stop signing multi-year seat-locked renewals without an agent clause. If the vendor's own roadmap includes agents doing the work, why are you committing to three years of human-count pricing? Negotiate caps, exit ramps, and conversion rights to usage pricing.
3. Make API access a contractual right, not a premium add-on. An agent-ready stack is one your own systems can act through. A vendor who charges rent for machine access to your data is charging you for the privilege of your own future. This is now a standard clause we recommend in any serious automation programme.
4. Prefer usage-aligned pricing where your volumes are honest. Usage pricing is not automatically cheaper — it's automatically more truthful, and truth is what you want in a contract during a phase transition.
5. Build the agent layer you own. The deepest risk isn't overpaying for seats; it's letting the automation dividend accrue to the vendor. If their "AI seat" does the work, they capture the margin. If your agent layer — orchestrating across tools through APIs — does the work, you do. Owning it is not free — a focused automation build typically runs 4–8 weeks and €18k–45k, an agent layer with retrieval over your own knowledge 6–12 weeks and €30k–70k, with ongoing evolution in the €2k–6k a month range — but that is a one-off plus a budget that compounds, set against seat rent that buys the same twelve months over and over. That layer is precisely what we build in our AI systems practice, and the ownership argument is the same one we make in Don't buy AI, buy outcomes.

The two objections, answered

"Our vendor says their new AI seat handles all this." It might — and notice what it costs. The vendor's copilot is priced to capture the automation dividend inside their margin: you pay a premium seat for the privilege of their agent doing work on their platform, on their terms, with their lock-in deepened. Sometimes that trade is fine for a narrow tool. As a strategy, it's outsourcing the most valuable layer of the next decade to whichever vendors you happened to have contracts with in 2026. The question isn't whether their AI seat works; it's who owns the leverage when it does.

"Our people like their tools, and seats are how we pay for them." Nobody is proposing to take working tools away from humans who use them well — judgment seats earn their keep. The proposal is narrower: stop paying human prices for machine-shaped work, and stop signing contracts that assume the 2019 shape of your company is permanent. Affection for a tool is real; it is not a pricing model.

What this looks like in practice

Honest examples from our own operation, because we run on this thesis:

Our publishing hub replaced a stack of would-be seats. One brief goes in; drafted, brand-consistent content for the blog, LinkedIn and Facebook comes out through one pipeline with a human approval gate. Priced in the old world, that workflow is several content-tool subscriptions and the human hours to shuttle text between them. Priced in ours, it's compute plus one person's judgment. The seats didn't get cheaper — they stopped existing.

Our automation spine is self-hosted n8n. Automation capacity priced in compute we control, not in per-user tiers — which changes behaviour: when adding a workflow costs nothing marginal, you automate things that would never have justified another subscription. The cheapest software often turns out to be the software that doesn't know how many of you there are.

And a clearly illustrative archetype: picture a 60-person professional-services firm running the seat audit from step one. A plausible honest finding: a meaningful minority of its licensed seats — the operations coordinator's project-tool license, the junior who assembles the weekly report (three hours a week is roughly 150 hours a year of seat time), the admin whose CRM seat exists to retype emails into fields — are, functionally, data-movement seats. Each one is a candidate for the same restructuring: agent does the movement, human keeps the judgment, license count quietly stops growing. Multiply that finding across every mid-size firm in Europe, and you understand why vendor pricing pages are being rewritten this year — and what the back office looks like by decade's end (we sketched that in The 2030 back office).

The endgame

Follow the logic to its resting point: software converges toward pricing work done — outcomes, actions, resolutions — because that's the only unit that still correlates with value when the worker might be silicon. In that world, the strategic divide among buyers is brutally simple: companies that own their agent layer capture the automation dividend; companies that rent everything hand it back as vendor margin.

The seats are melting either way. The only open question is who collects the water.

If you want the honest version of that seat audit — including the seats that should absolutely stay human — talk to us. We'll bring the framework; you bring the license spreadsheet you're slightly embarrassed about.