Three weeks into our Copilot rollout, a financial analyst forwarded me a screenshot of an error message that just said the request couldn’t be completed. She had asked Copilot in Excel to explain a variance across 14 cost centres. It had worked the day before. It worked again an hour later. That single inconsistency tells you more about Microsoft 365 Copilot’s reliability than any service health dashboard will.
We ran Copilot across 180 licensed seats for 90 days, covering finance, sales, legal and a 12-person IT team. Some of it was genuinely excellent. Some of it cost us real hours. Here’s the honest split.
Reliability means three different things in Copilot
Microsoft reports Copilot availability the way it reports Exchange Online availability, and that number stayed comfortably above 99.9% for us. It’s also the least useful figure in the whole review, because a request that returns a plausible-looking but wrong answer counts as uptime.
In practice, reliability broke into three separate questions, and they failed at very different rates:
- Availability: is the feature there when you click it? Ours almost always was.
- Consistency: does the same prompt produce the same quality of answer twice? This is where things got wobbly.
- Latency: how long does a dependable answer take? Frequently longer than the task it replaced.
Most of the frustration people reported to IT traced back to the second and third, not the first.
Where Copilot was boringly dependable
Outlook and Teams
Drafting replies to routine email and summarising recorded meetings were the two features that never once failed us over 90 days. Meeting recaps landed within about two minutes of the call ending, every time, including on a 71-minute call with 14 participants and three people dialling in on bad hotel Wi-Fi. The summarisation quality was consistent enough that our sales team started skipping note-taking entirely by week five.
That consistency matters more than brilliance. A tool that produces a solid 7/10 summary every single time gets adopted. A tool that produces a 9/10 summary 80% of the time and a useless one the rest of the time gets quietly abandoned.
Word and the drafting loop
Rewrite, shorten, change tone, turn these bullets into prose. Five months of heavy use showed no meaningful degradation. Response times hovered between three and six seconds for short documents, which felt natural enough that people stopped noticing.
Where it got flaky
Grounding against SharePoint got noticeably less reliable as the tenant grew
Early on, asking Copilot to pull together a summary from a specific project folder worked well. By month three, with roughly 400,000 items in the tenant, we were seeing three recurring problems: answers that cited a document from the wrong project, answers that missed an obviously relevant file sitting two folders deep, and the occasional refusal where Copilot claimed no such information existed.
We scored answer quality separately in our six-week Copilot accuracy review, and the pattern matched: the tool is strongest when the source material is small, well-named and recent. Permission trimming also caused real confusion. Copilot correctly refused to surface content people didn’t have access to, but it often said nothing rather than explaining the gap, which read as a broken feature rather than a permissions boundary.
Throttling around 9am
Between roughly 8:45 and 9:30 on weekday mornings, when our sales team fires off a burst of inbox sweeps and meeting prep, we logged noticeably more failed or timed-out requests. Nothing dramatic, maybe 4% of requests, but 4% at the moment of highest intent is the 4% people remember.
Long documents in Excel
Copilot in Excel handled clean tables of a few thousand rows without complaint. Point it at a workbook with merged cells, three hidden tabs and a decade of inherited formatting, and the failure rate climbed steeply. When it did respond, we caught it misreading a header row twice in ways that produced confidently wrong totals. We now require a clean, properly formatted table before anyone uses it for anything that reaches a board pack.
The latency tax nobody budgets for
Simple prompts returned in two to five seconds. Anything requiring grounding across multiple sources took eight to fifteen seconds, and occasionally stretched past thirty. That sounds trivial until you multiply it by 40 interactions a day per person.
We measured it crudely: six users, one week, screen recordings. Average Copilot interaction came in at 11 seconds. The manual alternative for the same tasks averaged 9 seconds for short jobs and 4 minutes for long ones. Copilot is a clear win on the long jobs and a small loss on the short ones, which means the productivity story depends entirely on what your people are actually asking it to do.
Admin controls held up better than the user experience
The compliance and governance layer was the most solid part of the product. Sensitivity label inheritance worked as documented. Audit logs captured prompts and responses cleanly enough that our legal team signed off without a fight. DLP policies applied to Copilot outputs without the loopholes we’d braced for.
What cost us time was the mundane stuff: licence assignment quirks that left two users waiting three days for Copilot to appear in their apps, and a tenant-wide setting change that took longer than expected to propagate. Nothing catastrophic. Just a reminder that this is still an IT project before it’s an AI product.
How it stacks up against the enterprise alternatives
We ran a parallel pilot on ChatGPT Enterprise, and the comparison is less about raw capability than about where each tool fails. Copilot’s weak spot is retrieval across messy internal sprawl. ChatGPT Enterprise’s weak spot is that it doesn’t live inside your documents, so the retrieval work lands on your users.
Budget-wise, the two aren’t priced on the same logic either. If you’re weighing both, our breakdown of what ChatGPT Enterprise’s custom pricing actually includes is a fair place to start, and the follow-up on 640 seats and six months of ChatGPT Enterprise covers the operational reality at scale.
What we changed after 90 days
None of these are feature requests. They’re the operational habits that turned a patchy experience into a dependable one:
- Consolidated 1,400 SharePoint sites down to 180, with owners named and stale content archived.
- Banned merged cells and hidden tabs in any workbook that touches a monthly report.
- Introduced a five-prompt library per department, because prompt consistency drove answer consistency far more than anything Microsoft shipped.
- Shifted meeting-recap dependent workflows away from the 8:45am peak.
- Added a two-source rule: anything going to a client or board gets verified against the underlying document, no exceptions.
Failure reports from users dropped by roughly two-thirds over the following month. Almost none of that came from Microsoft’s side of the fence.
The reliability question worth asking before renewal
Copilot’s infrastructure is dependable. The experience on top of it is dependable in proportion to how tidy your Microsoft 365 tenant is, and that’s an uncomfortable finding for anyone who assumed the licence fee bought them a finished product.
Before you renew, pull two numbers. First, what percentage of your SharePoint and OneDrive content has a named owner and a modified date inside the last 18 months. Second, how many of your users have changed their prompt behaviour since week one. If the first number is low and the second is near zero, Copilot’s reliability problems are sitting in your tenant, not in Microsoft’s datacentres, and no amount of waiting for the next model update will fix them.

