AI email assistants: what we found in Gmail and Outlook

On this page

We went looking for the case for a separate email agent. Following an unanswered enquiry, preparing for a meeting and finding a time that works are recognisable jobs. They are also the jobs that personal-agent companies keep promising to take off our hands.

The first finding complicated that story: Google's and Microsoft's native assistants already reach beyond writing replies. Their documentation describes mailbox changes, calendar actions and, in some cases, standing instructions that act later. A new assistant is entering an inbox where the incumbent can already do part of the work.

We followed those claims into feature eligibility, scheduling instructions and approval rules, then compared Spark, Cowork and Lindy's recurring workflows. Our conclusion is that follow-through is the more interesting contest. We think Lindy's unanswered-message routine is a clearer reason to consider another assistant than another email composer. Spark has a credible case in coordinating Google apps, while Cowork makes Microsoft's own execution offer harder to dismiss.

This was a primary-source investigation on 9 October 2026. We checked what the vendors document, including contradictions between their pages. We didn't connect an inbox, send a message or measure time saved.

Native Gmail and Outlook changed the comparison

Google's Gemini in Gmail guide describes search, summaries, drafts and inbox organisation, along with creating, rescheduling and deleting calendar events. Creation and event information have primary-calendar limits. Archive, delete, label and read-status changes require confirmation.

Microsoft's Copilot Chat in Outlook guide documents work-account inbox and calendar actions, including limited capabilities without a Copilot add-on. Broader enterprise context requires more entitlement. Standard Draft with Copilot still ends with the person reviewing and sending the reply.

Those distinctions matter to our assessment of the market. “AI for email” covers an increasingly capable bundle inside an existing subscription. A standalone vendor has to explain the additional responsibility its assistant takes on. We find the strongest explanation in work that starts before the next prompt arrives: noticing that a reply is overdue, assembling a recurring brief or carrying a decision between applications.

We organised the documented offers around that change in responsibility:

Product or experience	What caught our attention	What limits the conclusion
Gemini in Gmail	Inbox organisation and calendar actions alongside search and drafting	Feature eligibility varies; primary-calendar limits apply
Copilot in Outlook	Work-account actions and approved standing calendar rules	Consumer and employer accounts have different capabilities
Copilot Cowork	Multi-step execution and recurring work across Microsoft apps	Work access requires enablement and usage billing; personal access is in Preview
Gemini Spark	Tasks, reusable instructions and schedules across Google apps	Personal-account, subscription, activity and regional conditions apply
Lindy	Unanswered-message follow-up and connected administrative routines	Sending and approval behaviour depend on the routine and channel

This isn't a measured performance ranking. It records the propositions we found worth examining, then tested against the vendors' more specific instructions.

Google's scheduling documents didn't give us one eligibility answer

Google's plan comparison includes Gmail summarising and drafting across Business Starter, Standard and Plus. Help me schedule begins at Standard in that range. The comparison also lists personal Google AI subscriptions, with different entitlements.

The scheduling instructions explain a useful flow: Gemini proposes times during email composition, the sender reviews and sends them, and the recipient selects a slot to create the event. This removes some coordination without claiming access to an external client's unseen calendar.

We cross-checked the general availability table, which describes Help me schedule as US-only for eligible Workspace accounts. Its account wording doesn't align cleanly with the task instructions and plan comparison. The table also contains conflicting wording about free conversation summaries.

We therefore couldn't validate a universal Australian entitlement from those documents. That is a specific unanswered question, rather than a reason to call native Gmail incapable. The documented actions strengthened Gmail's place in our comparison. The conflicting rollout wording reduced our confidence in assigning those capabilities to an Australian account.

Microsoft's approval rules revealed several kinds of delegation

One detail in Copilot's triage instructions stood out. Text commands affecting five or fewer emails don't need confirmation. Larger actions present a confirmation card. An instruction can therefore be enough to change a small set of messages.

That is a different arrangement from reviewing an email draft before sending it. It also doesn't establish a universal five-email rule for other actions. We found that “Copilot asks first” was too broad to describe what the documentation actually says.

Calendar Instructions introduce another arrangement: approve a rule once, then let matching invitations be accepted or declined in the cloud. We find this more consequential than a better-written reply. The person is delegating future decisions within a standing instruction. The page doesn't settle every consumer-plan entitlement, so we haven't assigned that feature to all home accounts.

We also checked account differences. Consumer Family and Premium plans give their AI benefit to the subscription owner; supported Outlook apps can extend certain Copilot assistance to connected Gmail, Yahoo and iCloud accounts, while classic Outlook for Windows is excluded from that sharing. Work-account capabilities have separate organisational requirements, including Exchange Online for mailbox-grounded experiences. Consumer plans, account sharing, work requirements.

Cowork takes Microsoft's proposition further. Its setup guide separates generally available work access from personal Preview. Work requires the relevant licence, an enabled environment and usage billing. Its action controls describe approvals or handoffs for consequential actions, with scheduled tasks able to obtain approval during setup.

We think that makes Cowork a serious incumbent in this category. A scheduled assistant with advance permission to act is already competing for responsibility, not just helping inside a compose window. Our unresolved question is how reliably that responsibility survives exceptions; the instructions alone can't answer it.

Spark's strength is the connected Google job

Spark's product description brings Tasks, Skills and Schedules together. Its weekly inbox example combines priorities with focus blocks. We see a coherent proposition here: the assistant can draw on Gmail, Calendar and Drive within a recurring assignment.

The attraction is the combination. Finding a relevant message is one operation. Connecting it to an upcoming meeting and preparing the relevant material is the larger job. Spark has a documented route through those applications, which gives us a more specific reason to take it seriously than the general promise of autonomy.

Its personal setup requirements name adults with personal Google Accounts, Google AI Pro or Ultra, and Keep Activity enabled. They exclude the EEA, Nigeria, Switzerland and UK. Australia isn't excluded by that rule, although we haven't confirmed access in an Australian account. Selected business users are mentioned separately; a Workspace subscription alone isn't proof of entitlement.

Our assessment also depends on the execution route. Remote-browser work can continue with the device closed; local Chrome work requires it awake. Sign-ins and consequential actions can still interrupt either workflow. The interesting proposition is a recurring, connected job. Reliable unattended completion remains something to establish through use.

Lindy gave us the clearest example of following the work through

Lindy's follow-up bumps watch sent messages and draft a follow-up when no reply arrives within a chosen window. The feature supports Gmail and Outlook, requires Assistant features and can be disabled by an administrator. Its email drafting also leaves replies in the drafts folder.

We like the specificity. The problem is a conversation that has stalled; the work is recognising the delay and preparing the next move. A person retains the sending decision in those built-in flows. That is a useful division of responsibility even before autonomous sending enters the picture.

The routine documentation describes scheduled or event-triggered work across connected tools. Its dry run suppresses writes while showing intended results. We treat that as a way to inspect a proposed routine, rather than proof that delivery succeeds. The scheduling page describes booking after a recipient selects a slot, while the separate CC-based scheduling route remains labelled Coming Soon.

Our favourable view of the follow-up proposition has a qualification. The pricing FAQ broadly says outside-impact actions wait for approval. The integration guide limits its guardrails to writes in shared Slack threads, excluding web chat, Slack DMs, iMessage and SMS.

We give the specific operational rule more weight. The built-in draft routine has a clear sending boundary; the broader assistant does not acquire that same boundary in every channel. We think Lindy's strongest story is the concrete routine it documents. The general approval promise needs a more careful explanation.

What we now think the email-agent market is selling

The research left us with a stronger view of the native products and a narrower, more convincing case for the specialists. Google and Microsoft already cover substantial inbox work. Spark and Cowork extend their ecosystems into continuing assignments. Lindy's follow-up routine makes the value of a separate assistant tangible.

The economic question follows that division. Lindy describes seats and pooled credits; work-account Cowork adds usage billing; Spark's consumption depends on request complexity. We haven't measured any of them against a representative week of inbox work. A recurring brief that needs extensive correction could be poor value despite an attractive subscription.

Our wider personal-agent investigation found the same pattern in other jobs: the product's advantage depends on the work it can reach and carry through. Here, we would judge the next stage by whether the assistant maintains a useful obligation over time: the promised follow-up, the prepared meeting, the rescheduled appointment.

That is why we think email and calendars are worth watching. They contain commitments, other people's time and work that becomes overdue quietly. An agent that reliably takes responsibility for those details would change more than the quality of our drafts. The documents show the offers taking shape. The evidence we still want is sustained follow-through.

Primary sources checked 9 October 2026. No connected-inbox trial, outbound message, calendar change or vendor interview was conducted. Our judgements concern the documented offers. About Agentic Economy.