Six months ago I sat on a procurement call where a vendor said ChatGPT Enterprise would “transform how the company works.” Our operations director asked a fair question back: what does it do that the $20 version doesn’t? The answer took eleven minutes and never quite landed.
We bought it anyway. This is the review I wish I’d read before signing: 140 employees eligible, 62 seats activated at peak, roughly 4,800 conversations in our busiest month, and a renewal decision that turned out to be far less obvious than I expected.
What ChatGPT Enterprise actually is, stripped of the sales deck
The models are the same ones you get in Plus. The difference is everything wrapped around them: SAML single sign-on, SCIM provisioning so a leaver loses access the moment HR closes the ticket, an admin console with real usage analytics, retention controls, a compliance API, and a contractual commitment that your prompts and uploaded files don’t go into training data.
You also get the practical kit: a 32,000-token context window on GPT-4o, near-unlimited usage under fair-use terms, shared GPTs published to a private workspace, Projects, and connectors into Google Drive, SharePoint and GitHub. Encryption at rest is AES-256, and enterprise accounts sit under SOC 2 Type 2 reporting. None of that is exciting. All of it is the reason a security team says yes instead of sending you back to a personal Plus subscription.
It is not a different brain. If your team wasn’t getting value out of the $20 tier, Enterprise won’t manufacture it. What it removes is friction, and excuses.
The rollout took three weeks. Adoption is still going.
Domain verification and SSO took an afternoon, mostly because our identity provider admin knew what she was doing. The remaining twenty days went on governance, which nobody warns you about: who gets a seat, which document types are acceptable to paste in, who approves a shared GPT before it goes company-wide, and what happens when someone leaves with a folder full of chats.
By day 30, 62 of 140 people had ever activated. The admin dashboard, which is genuinely good, told a blunter story. Eleven people accounted for 41% of all messages. The data team alone was 19%. Sales, the group that asked loudest for licences, sat at 22% weekly active use. That pattern held for six months with only modest drift.
Where it paid for itself
First-pass contract review
Our two-person legal team built a shared GPT loaded with our standard positions on liability caps, indemnities and data processing terms. They estimate it saves 15 to 20 hours of outside counsel time a month on NDAs and vendor paperwork. Take that as their estimate rather than a measured figure, but even discounted it clears the seat cost for the whole legal function.
Support and localisation
We answer tickets in nine languages. Drafting in English and having the model produce a first pass in German or Japanese cut median first response for non-English tickets from about four hours to 90 minutes. An agent still edits every reply. That editing step is not optional, and it’s where most of the labour moved.
Engineering and incident work
This is where the long context window earns its keep. Paste a stack trace plus four related files and ask for the five most likely causes, ranked. Postmortem drafts that used to take an hour start at 70% done. Nobody ships code the model wrote unreviewed, but nobody writes a blank page anymore either.
The quiet win: shared GPTs
An onboarding assistant, a tone-of-voice editor, an RFP boilerplate generator. Collectively they replaced about 40 pages of process documentation that people used to hunt for in Slack bookmarks. When they work, this is the most underrated part of the product.
Where it disappointed, honestly
- Stale shared GPTs. Nobody owns the knowledge base. Within six weeks our onboarding assistant was confidently citing a policy we retired in March. We now assign a named owner and a quarterly review date to every published GPT.
- Connectors are useful but noisy. Ask for last quarter’s pricing decision and you may get twenty results, three of them relevant. Better than nothing, well short of enterprise search.
- Usage plateaus. We never pushed past about 40% of licensed seats in weekly active use, and that’s with internal training sessions and an exec sponsor. You are paying for chairs nobody sits in.
- It doesn’t police shadow AI. The biggest risk we found wasn’t the enterprise tenant. It was the marketing manager pasting a spreadsheet into a personal account because the login was faster.
The pricing conversation nobody will have on the record
Custom, annual, per-seat, quoted quietly after a discovery call. Our per-seat number landed in the range most mid-size buyers report, and it is not the number on the website for a reason. What you get at that price is mostly the admin and compliance layer rather than better answers, which is worth understanding before you compare it to a $30 consumer tier. There’s a fuller breakdown of what the custom price actually gets you if you want the line-item view.
Below roughly 50 seats the maths gets ugly fast, because you’re paying for a compliance stack that small teams rarely need.
Reliability matters more than benchmarks
We had one bad afternoon in the fourth month: about 40 minutes where the web app and the API both stumbled. Because we’d routed customer support drafting through it, two agents went back to writing from scratch and a queue that should have cleared at 4pm cleared at 6pm. Not catastrophic. Educational.
The lesson is to assume the web app and the API fail at the same time, because they did, and to keep a documented fallback for the three workflows where a 40-minute gap actually costs money. Our testing of this under load is written up in this review of ChatGPT Enterprise reliability under real-world pressure, including the days it held up fine.
Enterprise versus Team, without the marketing gloss
Team runs $25 to $30 per seat with caps, lighter admin controls and no compliance API. For a 12-person agency that is almost certainly the right buy, and the limits of the $25 seat are worth reading before you assume you’re upgrading for features you’ll never use.
The honest dividing line is this: buy Enterprise when you have an auditor, a procurement team or a regulator asking questions. Buy Team when your only question is whether people will use it.
Who should actually pay for this
- Organisations of 60+ people where data governance has an owner and a policy document.
- Regulated or audited industries where “we use the free version” is not an answer you can give.
- Teams already seeing heavy voluntary use of consumer AI, where the risk is uncontrolled rather than absent.
- Companies with a real onboarding and enablement plan, not just a licence email.
Skip it if you’re under 30 people, if nobody owns AI governance, or if your honest forecast is fewer than 40% weekly active users three months in.
A 30-day test that beats any demo
Pick 15 people across three functions. Give them seats. Track three numbers weekly: weekly active users, distinct workflows that survive past week two, and how many published shared GPTs still have an owner. If weekly active sits under 40% at day 30, negotiate down the seat count rather than the price, and revisit the value for money question with your own data instead of a vendor’s slide.
The tool is good, occasionally excellent, and priced for organisations that need the paperwork as much as the answers. Our renewal came down to one number: 41% weekly active use across 62 seats, with the top eleven people driving most of the value. We kept the seats, cut twelve, and put a named owner on every shared GPT. That’s a modest win, honestly earned.