What can an AI agent actually do? Real examples by team

Here is the concrete version of the question. A finance lead types, in Slack, "reconcile yesterday’s Stripe payouts against Xero and flag anything that does not match." A few minutes later a spreadsheet comes back with two mismatched transactions highlighted and the fee difference totalled. No dashboard opened, no CSV exported, no manual cross-checking. That is what "an AI agent doing the work" looks like in practice, and once you see one example the rest of this list reads as variations on the same move: plain-English request in, finished artefact out.

Most explanations of AI agents stay abstract. This one does not. Below are real request-and-output examples grouped by the team that would ask, so you can find your own job in the list and picture the exact thing that comes back.

The shape of every example

Across every team the pattern is the same. You describe an outcome the way you would brief a coworker. The agent works out the steps, uses your connected tools to actually perform them, and hands back the finished thing, a report, a draft, an updated record. For anything that changes the outside world, sending, editing, deleting, a well-built agent shows you first and waits for your go-ahead. Reading data is automatic; acting is approved.

Finance

• "Reconcile yesterday’s Stripe payouts against Xero and flag mismatches." → a spreadsheet with the deltas highlighted.
• "Chase every invoice more than 14 days overdue." → drafted reminder emails per client, sent on your approval.
• "Pull last month’s revenue and write a one-page board brief." → a finished PDF summary.
• "Match these receipt photos to the right entries in QuickBooks." → matched lines, held for sign-off before posting.
• "Tell me which subscriptions churned this month and the lost MRR." → a ranked list with the totals.

Marketing

• "Audit our ad accounts and tell me what to cut." → a prioritised PDF of losing creatives and wasted spend.
• "Write three launch email variants in our voice and put them in Notion." → a Notion page with all three drafts.
• "Turn this blog post into five LinkedIn posts and a newsletter." → the repurposed drafts, ready to schedule.
• "Pull last week’s campaign numbers and brief me on what moved." → a short performance summary with the so-what.
• "Draft replacement copy for the three worst-performing ads." → new variants to test, written from the audit.

Operations

• "Summarise this 80-message thread and list the decisions and owners." → a tidy recap with action items.
• "Build a small internal tool to track equipment checkouts." → a deployed app from a one-line request.
• "Pull this week’s numbers from across our tools into one status doc." → a single assembled report.
• "Find every open task with no owner and list them by project." → an organised, assignable list.
• "Schedule the quarterly review with everyone in #leadership." → calendar invites, sent after you confirm.

Sales

• "Research this inbound lead and tell me if they fit our ICP." → a brief with company size, fit, and a recommendation.
• "Update HubSpot from the notes in this call thread." → CRM records updated, shown before they save.
• "Draft a tailored first-touch email for this prospect." → a personalised draft, not a template.
• "Which deals have gone quiet for more than ten days?" → a stalled-deal list with last-contact dates.
• "Build a one-page recap of this account before my renewal call." → a finished prep sheet.

Engineering

• "Triage new error reports and open tickets for the top three." → three assigned tickets in Linear.
• "Summarise what shipped this sprint from our merged PRs." → a written changelog draft.
• "Find the open issues tagged urgent and who they are assigned to." → a prioritised, owned list.
• "Draft release notes for version 2.4 from the closed issues." → ready-to-edit notes.
• "Build a quick internal dashboard for our deploy status." → a deployed tool from one request.

Customer support

• "Triage the new support inbox and draft replies for the routine ones." → drafted responses, held for review.
• "Summarise this week’s tickets and the top three recurring issues." → a themed summary with counts.
• "Find every ticket mentioning the billing bug and tag them." → a grouped, tagged list.
• "Draft a help-doc answer for the question we keep getting." → a ready article draft.
• "Flag any ticket where the customer sounds about to churn." → a watchlist with the at-risk accounts.

The common thread

Notice what every example shares. None of them end with advice on how to do the task yourself; they end with the task done. None require you to pick the integration or wire up the steps; you described an outcome and the agent handled the route. And the ones that change something external paused for your approval, because reading your data and acting on it are different categories of trust.

Echo is an AI employee that does exactly this from inside Slack. It connects to over 3,000 tools, returns finished work rather than a conversation, asks before it changes anything external, remembers your context between tasks, and does not charge per seat. The fastest way to answer "what can it actually do" is to point it at one of the tasks above on your own tools. The first $50 of work is free at /signup.