For the outcome, not per head: how we changed the way enterprise AI is paid for
Enterprise AI is paid for in three ways today – per head, per consumption, or per outcome. Each drives different behaviour inside a company. Why we chose the third path for our AI Assistant and what it changes for whoever is rolling AI out.
Most debates about enterprise AI revolve around what it can do. We want to shift the attention elsewhere for a moment – to a question almost nobody asks, even though it decides whether a company will actually get anything out of AI: what are you paying for?
It sounds like an accounting detail. It isn’t. The way AI is billed directly determines how many people in a company will end up with it – and whether it ends up as a tool for a chosen few or as a given for everyone. We’ve just changed the pricing model for our AI Assistant (Private Secure AI), and this is exactly the thinking behind it. But before we get to ours, let’s honestly go through all three ways that exist on the market today.
Three ways enterprise AI is paid for today
1. Per head (per-seat)
The most widespread model. You buy a licence per user and pay for each one regardless of how much – or how little – they use AI. Microsoft 365 Copilot in the full licence costs around 650 CZK per person per month, even for someone who opens it once a month.
The problem is obvious: a director works out “price times headcount”, takes fright at the result, and buys AI for just a few people – or none. The per-head model puts the company in front of an “everyone or no one” choice, and the answer is usually “no one, then”. AI that five people out of a hundred have won’t change a company.
2. Per consumption (pay-as-you-go)
A newer and in many ways smarter model – and yes, Microsoft has it too. Copilot now offers consumption-based billing: so-called Copilot Credits for advanced actions and agents, agents built in SharePoint billed by “messages” through Azure. You don’t pay per head, you pay for how much the AI consumes.
It’s a step in the right direction, but it has two catches. The first is predictability: a credit is counted by the model used, the amount of context retrieved, the number of tools called and how long the task runs. A lighter task costs a few crowns, a heavier one twenty times that – and how many there’ll be in a month, you can’t estimate up front. The second catch is what you’re actually paying for: consumption grows with the AI’s “effort”, so also with intermediate steps and failed attempts. You pay for the AI having tried, not necessarily for it having delivered you something.
3. Per outcome
And here is the path we chose. You don’t pay per head or per consumption, but per completed action – for the finished work the AI does for a person. Not for how many users you have, not for how many tokens it consumed, but for how many times it actually helped.
The difference from consumption is subtle but fundamental. A “credit” is a technical unit a customer doesn’t understand and can’t estimate in advance. A “finished answer” is something everyone grasps immediately – and something you can count and plan.
What an “action” is (and what it isn’t)
For the per-outcome model to work, it has to be absolutely clear what counts. For us it’s simple:
An action = one finished assistant answer to a person’s instruction. A chat reply, an email draft, meeting minutes, a filled-in company template. Simply: did the AI do a piece of work for a person? One action.
And what never counts:
- clicks and individual “prompts” within a single answer,
- the internal intermediate steps the AI uses to prepare an answer,
- failed attempts,
- preparing and indexing documents in the background.
In other words: you pay for delivered work, not for effort. When the AI takes three runs before it answers you, it’s still one action. When it reads a thousand documents in the background to find the right one, it’s still one action. Meanwhile the company’s administrator sees the counts day by day right in the app – the very numbers the price is calculated from.
Why per outcome specifically
Four reasons this model makes more sense to us than the two before it:
- Clarity. “You’ll pay for finished answers” is grasped by any director instantly. Credits and tokens aren’t.
- Predictability. The price is based on a band you know at the start of the month – it works like a flat fee, only fairer. No surprises on the invoice.
- Fairness. When the AI doesn’t deliver something, you don’t pay for it. When usage falls, the price falls too – the band moves both ways based on a three-month average.
- Availability for everyone. And this is the main change. When you don’t pay per head, you can give AI to literally everyone in the company – reception, the warehouse, a temp – and pay only when it actually helps them. AI stops being a privilege and becomes a given.
How it works in practice
So it’s not just theory, three things about how the model runs at a customer:
- The start is free. First it just measures. The customer sees in black and white how much their people actually get done with AI before they pay a crown.
- Invoice in advance, no surprises. You know the amount at the start of the month. The band is calculated from the average of the last three months, so one busy week won’t blow your price.
- You see everything. The app has an “AI usage” tab – counts of completed actions by day, month and year, a breakdown by feature, a daily chart, an estimate of your own Azure costs and your current licensing band. Exactly the numbers the invoice is based on. And only counts and token consumption are measured, never message content or who wrote it.
Roughly: a company of up to ten people fits, with typical adoption, around 1,900 CZK a month; a mid-sized company in the higher thousands. The exact band is always determined by the measurement at the start – which is why it’s free. Concrete figures and a calculator with a comparison are on the application page.
One more thing tied to price: the AI model itself runs in the customer’s own Microsoft Azure, in an EU region – a matter of pennies per action, which the customer pays Microsoft directly and sees day by day. And the data never leaves their environment. That’s the principle our AI has been built on from the start, and we wrote about it separately.
What you’d need for the same with Copilot
So far we’ve talked mainly about price. But a comparison only makes sense when it’s clear what is being compared for that price. AI Chat is an assistant over your company data: it chats securely over it – documents in SharePoint, mail, calendar, Teams – and straight from it writes email drafts, meeting minutes, fills in company templates, compares documents or draws charts from numbers in your material. Always only over what the signed-in user has rights to, and with data that never leaves the tenant.
What would you need for the same with Copilot? Here is the distinction the whole price difference rests on:
- Copilot Chat, free as part of Microsoft 365, isn’t enough for this. It can chat over the web – research, brainstorming, general questions – but not over your company data. It can’t see your documents, mail or calendar.
- Chat over company data (so-called work grounding) is unlocked only by the full Microsoft 365 Copilot licence – around 650 CZK per person per month. Only that gives Copilot access to your SharePoint, mail and Teams via the semantic index, and only that can write drafts in Word and Outlook over your own data.
- Part of it is possible more cheaply, via pay-as-you-go agents – an agent built over a specific SharePoint library works even without the full licence. But it’s an agent over defined content, not a full assistant across documents, mail and calendar.
And why isn’t some of those cheaper tiers enough for a customer? For three reasons, and they’re worth knowing:
- Free Copilot Chat is built over the web. It architecturally can’t see your company data – it’s not a matter of configuration, but a product boundary. It’ll help with a general question, not with your contract.
- A pay-as-you-go agent reaches company data but stays an agent over defined content – typically one library. It’s not a personal assistant that answers across documents, mail, calendar and Teams and writes a draft straight from it.
- Drafts and work directly in Office – an email draft in Outlook, writing in Word over your data – have been tied exclusively to the paid full Copilot licence since April 2026.
Put simply: the lower tiers cover a piece of what AI Chat does – but not the whole, and above all not for every user. For everyone in the company to do with Copilot what AI Chat allows straight away, they need the full licence. And this is where the pricing models diverge the most.
A worked example: a company of 50 people
Say you want to give secure chat over company data to all fifty. Roughly (with typical adoption, where about a third of people use AI regularly):
| Solution | Per month | Who gets AI |
|---|---|---|
| Copilot full licence for all 50 | ~32,500 CZK | all 50 |
| Copilot full licence for 10 chosen ones | ~6,500 CZK | only 10, the rest with no AI |
| AI Chat for the whole company | ~6,900 CZK | all 50 |
For the amount that equips ten people with Copilot, AI Chat gives secure company chat to the whole fifty. That’s the practical shape of the difference between “per head” and “per outcome”. (The figures are indicative – the exact AI Chat band is determined by the measurement at the start, and take the Copilot price from Microsoft’s current price list.)
Where the line is: we’re not against Copilot
To be clear – this isn’t an article against Microsoft Copilot. Copilot is an excellent tool and it doesn’t rule out our solution; at many customers they run side by side. Copilot also has its own pay-as-you-go variant, so it too can avoid paying per head. And for the full licence it gives things our AI Chat deliberately doesn’t do – an assistant right inside Word, Excel and PowerPoint, advanced modes like Researcher and Analyst, image generation, working with the web. AI Chat is deliberately focused on one thing: secure chat over company data and drafts from it, for everyone in the company. It’s not a replacement for Copilot in everything – it’s a different answer to the question of how to give AI to everyone.
The difference this whole text is about isn’t “us versus them”. It’s the difference in what you pay for – and what follows from it for how available AI is in a company. When you pay per head, a few people get AI. When you pay per consumption, more people get it, but at a price nobody can estimate in advance. When you pay per outcome, you can give it to everyone and pay only for what it actually delivers. Which model suits a company depends on how, and by whom, it wants AI used.
What to take away
Three sentences to close:
- With enterprise AI, it’s not only what it can do that matters, but what you pay for it. The pricing model decides how many people end up with it.
- Per outcome is clearer than per consumption and more available than per head. You pay for finished work – not for tokens, not for licences.
- The best way to find out is to measure it. That’s why the start costs nothing with us: first you see how much AI actually gets done in your company, and only then do you deal with price.
Want to work out which band you’d land in, or just compare it with Copilot? Open the calculator on the application page, or drop us a line – we’ll go through it over your specific situation. The start is always just about measuring, free.