Outlook Copilot Review: Is the AI Any Good at Email?

The honest question buyers of Microsoft 365 Copilot want answered before they commit a per-user licence is narrower than the marketing page: does the AI actually help with email, day to day, or is it a demo that falls apart the first time you hand it a real thread? This Outlook Copilot review looks strictly at the email surface — summarise, draft with Copilot, coaching by Copilot, and Copilot's meeting-prep hooks that surface in Outlook — and separates what holds up from what is still demo-grade.

We do not quote a Copilot price. Microsoft repackaged the licensing several times in the last twelve months — Copilot Pro for consumer plans, the enterprise add-on, and after the 1 July 2026 changes some Business Standard and Business Premium SKUs that bundle Copilot as a permanent line item. Whatever number was on the sales page yesterday may not be the one on it today. Check the current figures on Microsoft's own pricing page before you sign — and if a rep quotes you a number that does not match the public page, ask for it in writing.

The short answer#

Copilot in Outlook is a real, working assistant — not a mock-up — and for two specific jobs it earns its place in your day. It summarises long threads well enough that you can catch up on a fifty-message chain in about a minute, and it drafts short first-pass replies that are usually faster to edit than to write from scratch. Those two capabilities alone are what most licence-holders end up using.

It is weaker at the things Microsoft demos hardest. Coaching by Copilot — the feature that critiques the tone and clarity of a draft you are writing — gives generic feedback more often than pointed feedback. The natural-language Q&A over your mailbox retrieves the right thread only when your question closely mirrors the wording of the thread itself. And it does not send anything for you, does not learn a per-client voice, and does not exist for any mailbox that is not on Microsoft 365 with a Copilot licence attached.

• What holds up: thread summarisation, short first-pass drafts, meeting-context lookups from recent mail, mailbox-permission respect and Purview/DLP compliance.
• What is still demo-grade: the coaching critique, freeform Q&A across the mailbox, drafts that need a specific tone or long context, and anything that requires acting on more than one message at a time.
• Deal-breakers by design: no cross-provider support (Gmail, IMAP mailboxes are out of scope), no approve-before-send agent that operates continuously, no per-client voice profiles.

Criteria that actually matter for an email AI#

A per-user AI licence gets judged on a handful of things once the novelty wears off. Feature bullet lists on a marketing page are not that list. The dimensions below are the ones that decide whether the assistant is still being used in month three.

• Summarisation fidelity — does it capture what actually matters in a long thread, or produce a bland recap that omits the disagreement in message seventeen? A good summary lets you skip the thread; a mediocre one makes you open it anyway.
• Draft usability without editing — of ten replies it drafts, how many can you send with a single tweak versus rewrite from scratch? For pure acknowledgements the ratio is high; for nuanced replies it collapses.
• Coaching signal quality — the coaching feature scores your draft on clarity, sentiment and reader tone. Useful if the critiques are specific; ambient noise if they are not.
• Retrieval accuracy — when you ask Copilot to find or reason across your mail ("what did Priya say about the Q3 renewal"), does it pull the right thread, or the most recently-modified one that happened to share a keyword?
• Client and mailbox coverage — does it work in new Outlook, classic Outlook, Outlook for Mac, Outlook on the web, Outlook mobile? Any gap here is friction the whole team pays for.
• Permissions and compliance behaviour — does the AI honour mailbox delegation, sensitivity labels, retention rules and Data Loss Prevention policies as strictly as native Outlook does?
• Autonomy model — is this a manual assistant you invoke button by button, or an agent that can operate continuously with your approval? Copilot is squarely the former.

How Copilot scores against those criteria#

The table below is a capability scorecard, not a benchmark of numbers we cannot reproduce. Each row is a documented behaviour we verified against Microsoft's own live pages and, where applicable, in an Outlook session with a Copilot licence attached.

A worked example — the sales-lead thread everyone gets#

Abstract scoring is less useful than watching Copilot handle one real email. Here is a thread most account managers see every week: a prospect replies to a proposal, loops in their finance lead, asks two questions, and requests a call.

Handing that thread to Copilot with "Summarise" produces a competent recap in about four seconds — who replied, the two questions asked, the call request, and (usually) the finance lead's name. It captures the shape of the thread. It sometimes drops the second question if the reply was long and the question was buried, which is exactly when a summary matters most.

Then you ask Copilot to draft the reply#

"Draft a reply that answers both questions, agrees to the call, and suggests three time slots next week" is a reasonable prompt. Copilot's draft usually answers question one well and question two in a way that reads like it did not fully read the thread — a bland restatement of the proposal rather than a targeted answer.

The three time slots come out generic (Tuesday afternoon, Wednesday morning, Thursday afternoon) rather than tied to real gaps in your calendar, because the draft flow does not pull live free/busy the way a scheduling assistant would. You will fix both problems in the compose window, which takes about thirty seconds and is still faster than writing the reply from scratch. That is the honest ratio for most replies: Copilot saves the first two minutes, you spend the last thirty seconds.

Coaching by Copilot, run on that same draft, will suggest "consider making the greeting warmer" and "the tone here could be more collaborative." Neither critique names the line it applies to. If you have written more than a hundred client emails you already know when the greeting is cold; the coaching adds atmosphere rather than direction.

Where Copilot's classic-versus-new Outlook split still bites#

The single question we get asked most about Copilot for email is whether it works in classic Outlook or only in new Outlook. The honest answer as of August 2026 is: mostly the same, but not identically, and the gap is a moving target.

New Outlook is web-architected, so Copilot surfaces there and in Outlook on the web at the same time and behave the same. Classic Outlook receives Copilot too, but some newer entry points — inline compose suggestions, the more recent meeting-prep surfaces — arrive in the new client first and reach classic later, if at all. Anything that depends on COM add-ins or VBA is untouched by Copilot on either side; that is a separate architectural cliff.

Microsoft moved enterprise opt-out for new Outlook from April 2026 to 1 March 2027 (per message centre post MC949965, updated February 2026), and existing classic installs are supported until at least 2029. There is no forced cutover coming this year. If your team runs classic today, do not switch clients just to get Copilot — verify which Copilot surfaces you would gain in your admin centre first.

Red flags in an Outlook Copilot rollout#

None of these are dealbreakers on their own, but each has caught buyers off guard often enough that we would raise them on the demo call.

• Licence-per-user pricing that only unlocks Copilot on eligible base plans. A Business Basic seat cannot just add Copilot; the base plan has to be a supported one first. Do the base-plan math before the Copilot math.
• Coaching feedback that reads as filler after the first week. Teams that build training around the coaching feature discover it does not have enough signal to sustain a review process. Treat it as a nudge, not a QA tool.
• Ask-over-inbox retrieval that misses when the question is loosely phrased. If your team relies on "find me the thread about X" as a workflow, test it on your real mail before committing — the miss rate on paraphrased questions is higher than the demos suggest.
• No mailbox coverage outside Microsoft 365. Copilot for Outlook does nothing for a personal Gmail, an IMAP alias, or a client's Google Workspace account you pull in via forwarding. If your job spans providers, Copilot only covers part of your inbox.
• No approve-before-send agent. Copilot will not send anything without you clicking send, which is safe, but it also will not act continuously. If you want an assistant that triages overnight and hands you a queue in the morning, that is not the shape Copilot ships in.
• Model retention terms differ from the surrounding Microsoft 365 terms. Read the specific Copilot data-handling clauses on Microsoft Learn; do not assume every existing Microsoft 365 promise applies verbatim.

What we'd pick, and why (honest)#

Copilot for Outlook is a real product — not vapourware, not a demo — and for a team that lives entirely inside Microsoft 365, is already on an eligible base plan, and mostly wants faster catch-up on long threads plus quick first-pass drafts, it is a defensible line item. It is not the AI email assistant we would pick, but it is the one that makes structural sense for that team.

The dimension we will concede to Copilot outright: Microsoft has built harder on cross-Microsoft-365 grounding than we have, and inside a heavy M365 stack — Teams, SharePoint, Loop, OneDrive, Outlook — that shows. Copilot's pre-meeting brief can pull the relevant Teams chat, the SharePoint doc, and the mail thread together in a way a third-party client on top of Outlook cannot match, because we are not sitting inside those other surfaces. If your work is deeply cross-M365 and you want that context in one AI, Copilot is the honest recommendation and we are not.

We build AI Emaily. It is an AI-native email client — one seat per user, connects Gmail, Outlook and IMAP under one login, keeps an approve-before-send Copilot mode as the default with a full audit trail on every agent action, and does not train any model on your mail (including the underlying providers, which we contract for zero retention). Voice matching works from a user-set Personal Context brain and per-client profiles, not from scanning your sent folder. Pricing is a 7-day full-access trial on Pro or Autopilot with a card required, then a flat per-seat plan; there is no permanent free tier.

The reader we are right for is the founder, operator or small team who wants a single AI-native inbox across providers, who wants an agent that can act continuously with human approval rather than a manual assistant they invoke button by button, and who wants the safety envelope — approve-before-send, audit log, no training on mail — to be part of the seat price rather than a compliance upsell. If that is you, try it at aiemaily.com and see the current per-seat and lifetime numbers at /pricing.

The reader we are wrong for is the team above. If you are all on Microsoft 365, all on eligible base plans, and your AI use case is thread summaries plus short first-pass drafts inside the M365 stack you already live in, Copilot is closer to your workflow than we are, and the honest cheapest answer is the licence that plugs into the plan you already pay for. Take it. This page is not going anywhere.