We put ChatGPT Enterprise in front of 640 people in January 2027 across engineering, legal, customer support and a sales team that mostly wanted help with cold emails. Six months later, the renewal is on my desk. So this is less a feature tour and more a verdict from someone who has to justify the invoice.
The product shipping in 2027 is not the one that launched in 2023. Agents, connectors, workspace-level memory and a genuinely usable admin console have changed what you are buying. Most reviews online still describe the old thing.
What the seat price actually buys
List price sits at roughly $60 per user per month on an annual commitment, and volume negotiation got us to about $52 at 640 seats. That is not cheap. Whether it is good value depends entirely on how much of the admin layer you end up using, because that is where the difference between Enterprise and a stack of Plus accounts lives.
The console is the product
SSO through SAML, SCIM provisioning from Okta, role-based permissions that map cleanly onto departments, and usage analytics broken out by team rather than by individual. The last one matters politically. We could show the finance director that support was running 1,900 conversations a week while legal ran 60, and budget the renewal accordingly instead of guessing.
Retention controls, eDiscovery export and a compliance API that our security team could actually query without filing a ticket rounded things out. Domain verification and a contractual guarantee that business data is not used for training were table stakes by 2027, but they were the reason procurement signed at all.
Connectors and agents are where it gets real
Wiring the workspace to SharePoint, Google Drive, GitHub and Slack took our platform team about three weeks, mostly spent arguing about permission mapping rather than writing code. Once live, the useful shift was that people stopped pasting context into a chat box. They asked a question and the assistant pulled the relevant document itself.
The agents build is more interesting and less mature. Our support team built one that triages inbound tickets and drafts first responses. It saves roughly 40 minutes per agent per day. It also confidently mislabels billing questions as technical ones about 8% of the time, which nobody noticed until a customer complained.
The full cost is not the seat count
Our annual line item is around $400,000 for licences. The real spend was closer to $490,000 once you add the two engineers who maintain connectors and permissions, the internal training sessions, and the 18% of seats that went dormant within four months. Dormancy is the quiet killer in any enterprise AI rollout. Someone gets a licence, uses it hard for three weeks, then returns to old habits.
Budget for seat reclamation from day one. We reclaimed 74 licences at the four-month mark after a usage review, which paid for the entire training programme twice over.
Where it earns its keep
- Drafting and rewriting. Unglamorous but universal. Our legal team cut first-draft contract summaries from 90 minutes to about 25.
- Code review support. Not writing features, but catching the same five mistakes our senior engineers were tired of flagging.
- Internal knowledge lookup. The connector layer replaced a lot of Slack archaeology. “What did we decide about the pricing tiers in March” now returns a real answer.
- Meeting follow-through. Action items extracted from transcripts, assigned, and posted into project channels without a human retyping anything.
- Onboarding. New hires asked it questions they were embarrassed to ask their manager. Adoption among the 2027 cohort was 94% in week one.
Where it still trips
Latency is inconsistent and it correlates with time of day. Between 09:00 and 10:30 CET, median response time on long-context requests roughly doubled compared to our evening baseline. Twice we suspected the workspace had quietly routed to a lighter model during peak load, because answer quality dropped in ways that were hard to prove and easy to feel.
We logged three partial outages across six months, the longest at 47 minutes. None of them were catastrophic, and all three were handled better than the average SaaS incident. Still, if your workflows depend on this, you need a documented fallback. We wrote ours after the second one, and it is worth reading this honest review of ChatGPT Enterprise reliability under real-world pressure before you sign a multi-year deal.
The compliance conversation nobody wants to have
Our legal team’s biggest objection was never about data storage. It was about authority. A model that produces a fluent summary of a statute or a clause feels like advice, and advice from a non-lawyer is a liability. We ended up writing an internal policy that treats every output as a first draft from an enthusiastic intern, and the framing question of whether AI reasoning maps onto legal reasoning at all is explored well in this piece on the gap between law and logic.
Two practical notes. Connector permissions mirror source permissions, so a folder that was over-shared in SharePoint becomes over-shared in chat. And vendor stability is a genuine procurement risk, not a footnote. OpenAI’s internal governance wobbles have real downstream effects on roadmap confidence, and if you want a sense of how messy that can get, the boardroom episode is instructive. Ask for exit terms and data export guarantees in writing.
How it compares in 2027
Claude’s enterprise tier is stronger on long document work and noticeably more cautious, which legal preferred. Gemini’s advantage is the Workspace integration, which is hard to beat if your company already lives in Google. Microsoft 365 Copilot is the sneaky competitor because for organisations already on E5 licensing, the marginal cost looks close to zero, even though real-world adoption is patchier.
We benchmarked six tools head to head on the same twelve tasks. If you want the short version of that exercise, this roundup of AI tools worth paying for in 2027 covers the ones that survived contact with an actual budget. ChatGPT Enterprise won on breadth and on the admin console. It did not win on any single task category.
What it changed about how teams work
The unadvertised effect was on job design. Junior analysts stopped doing first-pass research and started reviewing machine output, which is a harder skill and a better one. Two managers told me they now hire for judgment rather than throughput. That shift in how humans delegate to AI systems is the subject of this analysis of TML and human-AI collaboration, and it matched what we saw almost exactly.
A rollout plan that won’t waste the first quarter
If I were starting again, here is the sequence I would run:
- Weeks 1 and 2: ship SSO and SCIM only. No connectors, no agents. Get provisioning stable before anything depends on it.
- Weeks 3 to 5: pick two departments with high-volume, low-risk text work. Support and marketing, in our case. Give them 40 seats each and one named internal champion per team.
- Week 6: connect a single document repository with clean permissions. Do not start with the shared drive everyone has been dumping files into since 2019.
- Weeks 7 to 10: run a weekly 30-minute clinic where people bring a real task and leave with a saved prompt. Attendance correlates with adoption more than any email we sent.
- Month 4: reclaim dormant seats and reallocate. Publish the usage numbers internally. Nothing drives adoption like a team seeing another team’s numbers.
- Month 6: build one agent, not five. Ship it behind a review step where a human approves every output for the first month.
The single biggest mistake we made was treating this as a software deployment rather than a change in how people spend their afternoons. The licence fee is the easy part to justify. The harder question, and the one your renewal meeting will actually hinge on, is whether 640 people are visibly better at their jobs than they were in December. For us the answer was yes, in about a third of roles, and lukewarm everywhere else. That is a fair return at $52 a seat. It would stop being a fair return at $80.