The short version
- Buy seats when a person is present for every run: writing, research, drafting, summarising, analysis someone reads.
- Build when the work must happen without anyone opening a chat window, or has to write to a system of record with an audit trail behind it.
- The prices. ChatGPT Business is reported at $25 per seat monthly, $20 annually, two-seat minimum. Microsoft publishes 365 Copilot at $18 per user per month paid yearly, on top of a qualifying Microsoft 365 licence.
- The arithmetic. 25 seats at the annual price is $6,000 a year, forever. My published band for a focused custom build is €4,950–€19,950, once.
- Most companies want both, and that is not a fudge — it is the correct architecture.
This question arrives about once a week, usually phrased as a challenge: we already pay for ChatGPT, so why would we pay you? It is a completely fair question and it deserves the version where I talk myself out of the sale, because for a decent share of the companies who ask it the honest answer is that seats are all they need.
What follows is the line I use to tell the two apart. This is the ChatGPT-specific half of a broader comparison; if your question is really about custom software versus buying any packaged tool, the general version is custom AI vs off-the-shelf software, and the four-way including Copilot and the agent builders is over here.
What you actually get for a seat
Worth being precise, because the pitch and the product have drifted apart in both directions. A business subscription gives every employee a very good general assistant in a browser tab, with your company’s data kept out of training, an admin console, shared workspaces, connectors to common tools, and the ability to save a prompt-plus-instructions bundle as a reusable assistant. That last part is genuinely useful and under-used by most of the companies paying for it.
The pricing, as of August 2026:
| Plan | Per user / month | Conditions |
|---|---|---|
| ChatGPT Business (reported) | $25 monthly / $20 annual | Two-seat minimum. Price cut in April 2026. |
| ChatGPT Business premium seat (reported) | $125 monthly / $100 annual | Higher usage ceilings, mixable with standard seats. |
| Microsoft 365 Copilot (published) | $18 paid yearly | Discounted from $21. Requires a qualifying Microsoft 365 plan underneath. |
Copilot figures from Microsoft’s own pricing page, read 18 August 2026. OpenAI’s pricing page blocks automated readers, so the ChatGPT figures are marked as reported rather than verified at source — check the live page before you build a budget on them.
Note the Copilot asterisk, because it is the one that catches finance teams out. The $18 is an add-on. The all-in number is your existing Microsoft licence plus the add-on, which is why comparisons that put $18 next to $25 are not comparing the same thing.
Five situations where seats are the right answer
I say this to prospects regularly and it costs me nothing, because the ones who need a build find out soon enough.
- The work is writing and thinking. Proposals, first drafts, marketing copy, research summaries, tidying a messy spreadsheet, explaining a contract clause. A person is there, judging the output, every time. Nothing custom improves on this.
- You have not identified the process yet. If you cannot name the specific workflow, in one sentence, with the systems it touches, you are not ready to commission anything. Six months of seats will tell you what to build far more cheaply than a discovery phase will.
- The volume is low. Twelve of something a week does not justify engineering. It justifies a good prompt saved somewhere everyone can find it.
- The process changes constantly. Software encodes decisions. If the decisions are still moving weekly, encoding them is premature and you will pay to change them.
- You are entirely inside one vendor’s stack. If everything lives in Microsoft 365 and the job is summarising your own documents and meetings, Copilot is already sitting on the data and no integration work is required to get it there.
Underneath all five is one test: is a human present for every single run? If yes, buy seats. The subscription model is priced for exactly that shape of work and it is very good value at it.
Estimate your custom AI project in 30 seconds
Three questions, an instant cost range and timeline based on real shipped projects. After 30 minutes on a discovery call you have a written fixed-price quote.
Or build your own AI system piece by piece and send the design in for a written quote →
Five places the subscription stops
These are the failure lines I actually see, roughly in the order companies hit them.
1. Nobody is there to press the button. Every subscription assistant is reactive. Someone opens it, pastes something in, reads what comes back. The moment the requirement is “when an invoice arrives, do this” or “every night, reconcile these two systems”, there is no chat window in the loop and no seat licence that solves it. This one line accounts for most of the builds I am asked for.
2. It has to write, not just read. Connectors have got much better at reading your files and mail. Writing back into a system of record — posting a journal entry, updating an order, creating a ticket with the right owner and status — is a different problem, because it needs permissions, validation, idempotency and an audit trail of what changed and why. That is engineering, not configuration.
3. You cannot prove it is still working. A saved assistant that behaved well in March may behave differently in September, because the underlying model changed, someone edited the instructions, or the inputs drifted. A custom build gets an evaluation harness: a fixed set of representative cases with expected outputs, run on every change. It is unglamorous and it is the single biggest reason a system is still trusted in year two. There is no equivalent inside a seat licence.
4. The cost stops looking small. A per-seat model charges for headcount, not for value delivered, and it charges forever. Do the arithmetic in the table below before you decide the subscription is the cheap option.
5. It is not yours. Prompts, instructions and workflow live in someone else’s product. Prices change, plans get restructured, features get folded into a higher tier. With a build you hold the repository. That is not paranoia, it is the same reason you would not run your accounting on a tool you had no export from.
The arithmetic, plainly
Using the reported $20 annual seat price, and my own published band of €4,950–€19,950 for a focused build.
| Seats | Per year | Over three years | Against a focused build |
|---|---|---|---|
| 10 | $2,400 | $7,200 | Seats win comfortably |
| 25 | $6,000 | $18,000 | Roughly level with a build at the bottom of the band |
| 60 | $14,400 | $43,200 | A build is cheaper across three years |
| 150 | $36,000 | $108,000 | A platform build is cheaper, and does different work |
Now the honest caveat, which most comparison articles leave out because it spoils the punchline: this is not an either/or, so the crossover is not a decision point. Cancelling 60 seats because you commissioned an invoice automation would be daft — the seats do writing and thinking, the build does invoices. What the table is genuinely for is puncturing the idea that a subscription is automatically the frugal choice. Past about 25 people it is a five-figure annual line item, and it should be compared against other five-figure line items rather than waved through as software.
Add the running costs on the build side to keep it fair: model API usage on a live SMB system runs €50–€500 a month in my experience, and published sources put annual maintenance at 15–30% of the original build. I put every one of those figures, mine and other people’s, side by side in the 2026 pricing benchmarks.
What about custom GPTs and agent builders?
Fair challenge, and the gap is narrower than it was a year ago. You can assemble something with real tool access inside these products now, without writing code, and for a single-purpose internal job it is often genuinely the right call. Start there. If it holds, you have saved yourself a project.
Where they run out is not the demo, it is the second year. Version control and a review process for changes. Testing against fixed cases so a tweak on Tuesday does not silently break Thursday. Permissions that match the ones in your own systems rather than the ones the builder offers. Error handling for the afternoon an API is down. Portability, so the thing outlives both the vendor’s roadmap and the employee who built it. Every one of those is invisible in a demo and decisive in production. That is the whole distinction, and it applies equally to a no-code agent platform — I sketched the same ceiling in what a custom AI agent costs.
The decision, in one table
| Question | Seats | Custom build |
|---|---|---|
| Is a person present for every run? | Yes | No — it runs on an event or a schedule |
| Does it write to a system of record? | No, or via copy and paste | Yes, with permissions and an audit trail |
| Do you need to prove it still works? | The user checks each output | Evaluation harness on every change |
| How does the cost scale? | Per head, forever | Once, plus running costs |
| Who owns it? | The vendor | You — repository, prompts, integrations |
| Time to value | Same afternoon | 3–12 weeks depending on scope |
What I would actually do
- Buy the seats first. Whatever else is true, everyone who writes for a living should have one. It is the cheapest capability increase available to a business right now.
- Wait a quarter and watch. The processes people quietly stop doing by hand are your candidates. So are the ones they try to do in the chat window and abandon because it is too fiddly. Those two lists are worth more than any consultant’s discovery phase.
- Build the two or three that keep coming back. High volume, clearly defined, touching systems, and nobody wants to be the person doing it manually.
- Keep the seats. They were never the thing you were replacing.
If you already know which process it is, the estimator gives you a range in thirty seconds, and the builder lets you assemble the design yourself and send it in for a written fixed-price quote. And if you send it in and my answer is that seats already cover it, that is the answer you will get.
Common questions
Is ChatGPT Business enough for my company?
If the work is people writing, summarising, researching and drafting with the AI in front of them, yes, and you should not spend anything else. A subscription is the right answer whenever a human is in the loop for every single run and no system of record has to be written to. It stops being the right answer when the work has to happen without anyone opening a chat window.
How much does ChatGPT Business cost per user in 2026?
Reported at $25 per user per month billed monthly and $20 billed annually, with a two-seat minimum, after OpenAI cut the price in April 2026. Premium seats are reported at $125 monthly or $100 annually. OpenAI’s pricing page blocks automated readers, so treat those as reported figures and check the current number before you budget. Microsoft publishes its comparable price directly: Microsoft 365 Copilot is $18 per user per month paid yearly, discounted from $21, and it requires a qualifying Microsoft 365 licence underneath it.
At what point is custom AI cheaper than ChatGPT seats?
Straight arithmetic on the reported $20 annual seat price: 25 seats is $6,000 a year, 60 seats is $14,400 a year, 150 seats is $36,000 a year. A focused custom build in my published bands is €4,950–€19,950 once, plus running costs. So on pure cost a mid-size deployment overtakes a small build inside two years. But cost is the wrong reason to switch, because they do different jobs — the right reason is that the work needs to happen without a person driving it.
Can I use ChatGPT and custom AI together?
That is what most of my clients end up doing, and it is the correct answer more often than either extreme. Seats for everyone who writes and researches; a custom system for the two or three processes that run whether or not anyone is at their desk. Buying seats does not commit you to anything, which is exactly why it is the right first move.
What can a custom build do that a custom GPT or an agent builder cannot?
Write to your systems of record with permissions and an audit trail, run on a schedule or an event rather than a prompt, be tested against a fixed set of cases so you know it did not silently get worse, and survive the person who built it changing jobs. Agent builders have closed a lot of the gap on the demo and almost none of it on the second year of operation.