Per-seat pricing is a measurement error.

I have signed a lot of software contracts that priced the wrong thing. Not overpriced — mispriced. The number on the page was defensible and the unit it was attached to was nonsense. Somewhere in the last two years the seat quietly stopped being where the work happens in most of the systems I run, and the invoice has not caught up.

Why is per-seat pricing a problem now?

Per-seat pricing charges for human logins. It worked for thirty years because a login was a decent proxy for how much work a system did. Once agents perform the reads, the writes, and the drafts, that proxy breaks in both directions: usage triples while headcount falls, or you hire four people who barely touch the system. The price stops tracking the value either way.

This is not a moral complaint about vendors being greedy. It is a measurement complaint. A unit of account is supposed to correlate with the thing you are buying, and for a long time seats did. If one person could handle forty support conversations a day, then twelve seats meant roughly four hundred and eighty conversations, and charging by the seat was a rough charge by the work. The correlation did the work, not the seat.

Break the correlation and you are left holding a number that means nothing. That is what has happened. The seat is now a measure of how many people have credentials, which is an access-control fact, not an economic one.

These figures come from the vendors' own pricing and documentation pages — HubSpot and Salesforce checked on 30 July 2026, Zendesk and Intercom rechecked on 5 August 2026. Treat them as indicative: list price is a starting position, every one of these companies negotiates, and the definitions move. Zendesk restructured what counts as a billable resolution on 18 May 2026, and no longer publishes a rate per resolution at all.

What are vendors charging for instead?

They moved the meter rather than removing it. Salesforce sells Flex Credits in packs of 100,000 for $500, where an Agentforce action costs 20 credits — ten cents — and a Voice action costs 30. Zendesk charges for verified automated resolutions. Intercom charges $0.99 per outcome. HubSpot still charges per seat, with tiers.

Each of these is a bet about where the value sits, expressed as arithmetic. It is worth reading them as arguments, because they disagree with each other in interesting ways:

• Per actionSalesforce. Flex Credits meter the agent's effort: 20 credits per Agentforce action, 30 for Voice, sold in packs of 100,000 for $500. Honest about what costs the vendor money — though a loop that retries twice bills you three times. Credits and the older Conversations SKU cannot coexist in one org.
• Per conversationSalesforce's previous model, at $2 each, now being retired in favour of credits. A conversation is a unit the customer recognises without being taught it, which was its main virtue.
• Per verified resolutionZendesk. Since 18 May 2026 resolutions are tiered and two of the three cost nothing: an assisted escalation, where the AI helped but a human finished, and a contained resolution that fails verification. Only a verified resolution — no customer follow-up for 72 hours, then confirmed by a separate model reading the transcript — draws down your allowance. The rate is not published: the pricing page carries no per-resolution figure and the numbers in the help centre are marked as placeholders, so it reaches you through a salesperson.
• Per outcomeIntercom, at $0.99, charged at most once per conversation however many questions Fin answers. Four outcome types bill: resolution, Procedure handoff, and disqualification at $0.99, qualification at $9.99. Attempts are never billed, and a resolution is deducted if the customer returns to the same conversation.
• Per seat, stillHubSpot. $90 per seat per month on Sales Hub Professional billed annually, $100 monthly, with cheaper core seats around $45–50 and a $1,500 one-time onboarding fee. Legible, budgetable, and measuring the wrong noun.

Is outcome-based pricing better than per-seat?

Mostly yes, and I want to give it real credit. Charging when a customer's problem actually got solved is a closer fit to value than charging $90 because someone has a login. Both of the serious implementations have gone further than they had to: Zendesk made two of its three resolution tiers free, so the case where the AI helped and a human finished costs nothing, and Intercom deducts a resolution if the customer comes back later asking for more. Those are refusals to bill, written into the product. They cost the vendor money and they didn't have to be there.

But outcome pricing has two failure modes that per-seat does not, and the industry is not saying them out loud.

The first is definitional. The vendor writes the definition of the outcome, counts the outcomes, and adjudicates the edge cases. Zendesk's verification is performed by Zendesk's own language model reading the transcript. That is better than not verifying at all, and it is still the vendor marking its own homework — and in Zendesk's case the price attached to that verdict is quoted rather than published, so you cannot check the arithmetic before the call. When Intercom prices a sales qualification at $9.99 and a disqualification at $0.99, it has encoded a ten-to-one judgment about which result matters — a judgment you now pay for, made by a party with an interest in the answer. Read the edge cases before you sign: at Intercom, a frustrated customer who reads Fin's answer and leaves without asking for a human is an assumed resolution, and it bills.

The second is that a per-outcome meter charges you for succeeding. Automate more of your support and the bill goes up, more or less linearly. That is the alignment story told as a feature, and for a while it genuinely is one. It stops being one at the exact moment automation starts working, because the thing you were promised — cost falling as software absorbs the work — arrives instead as a variable cost that scales with volume.

When is per-seat pricing still the right unit?

When humans really are the operators. If your work is high-touch sales where a named person owns each relationship, or consulting where you bill hours yourself, or a regulated process where every human needs distinct entitlements and an individual audit trail, then seats measure something real and per-seat pricing is a fair, legible way to pay for it.

There is a second argument for it that I think is underrated: per-seat pricing is predictable. Finance teams dislike variance more than they dislike paying too much. A metered bill transfers volume risk from the vendor to the customer, and a company that cannot forecast next quarter's software spend will often rationally choose the flat number that is too high over the variable one that might be lower. Anyone selling metered pricing as straightforwardly pro-customer is skipping that.

So the honest position is narrower than "per-seat is dead". Per-seat is wrong when software does most of the operating, and right when people do. Most companies are somewhere in between and drifting in one direction.

What does Munin charge, and what is the trade?

A flat €99 per month per organisation, not per seat, with metered overage above the included allowance. Cloud Free is €0 and covers 5,000 MCP calls, 250 contacts, and 100 MB — enough to run a small operation properly rather than enough to look at a demo. Self-hosting is free forever under MIT.

I should be precise about what that is and isn't. Flat-per-org is not a meter-free utopia. There is an allowance, and past it there is metering, so we have made the same kind of guess Salesforce and Zendesk made. Ours is that the fairest thing to count is the MCP call, because that is the unit that actually costs us money to serve. We would rather bill for our own cost than grade your outcomes — and our rate is on the pricing page rather than in a quote.

The trade is where the variance lands, and we put it on our side on purpose. A ten-person team driving 155 tools all day pays the same €99 as a solo founder who checks in on Tuesdays. The customer gets a number they can forecast a year out; we carry the volume risk. We already argued that the exit should be open, and the same logic constrains the price: the ceiling on what we can charge is enforced by docker compose up, not by our restraint. Every module has a matched export and import pair. If the number stops being fair, leaving is a sequence of tool calls.

What should you actually count in an agentic system?

Count the thing that consumes a resource you can name. Compute, storage, calls, messages sent — units where both sides can point at the same meter and neither party has to be trusted to define what happened.

Outcomes are seductive because they sound like value, but they require someone to adjudicate, and adjudication is where pricing quietly becomes a governance problem. Seats are seductive because they are simple, but they measure credentials. The middle path is dull and defensible: bill for consumption of a resource, publish the rate, and let the customer's own logs reconcile the invoice.

This is the same argument I have been making about the software itself, one layer down. Once the cheapest operator in the system stops being a person, the assumptions built around people stop describing the system — and pricing is one of those assumptions, not an exception to them. It just happens to be the one with a number attached, which is why it is the easiest place to see the error.

If you want to see what the flat number buys before it costs anything, Munin Cloud has a free tier, and the comparison of the open-source CRM field has the licences alongside the prices.

A seat was never a unit of work. It just used to correlate.