How do you review an agent's work before it goes out?
Check an agent's run against its brief in Claude, ChatGPT or Gemini, see exactly what it changed, and decide what can go out.
Part of Check the workTime about 20 minutes, with the try-itChecked 16 Oct 2026Done saved in this browser
Three things to take away
- 1
An agent hands you finished work. It looks done. That is the trap.
- 2
Run the five checks before anything leaves your hands: the job, the facts, the steps, the changes, the sign-off.
- 3
Read what it did, step by step, not just the summary at the end. "The AI did it" is never the answer to "who checked this?"
Same checks, three tools (as of Oct 2026)
Where each tool shows you its steps, its sources and its changes.
| What you are checking | Claude | ChatGPT | Gemini |
|---|---|---|---|
| A task it did for you | Read the result and the steps it took; check progress from your phone | Follow its progress, then review the result and ask for changes in the same chat | Read the result and the steps it took |
| A research report | Open the sources it lists | Review the plan first, limit it to sites you name, then open the sources | Edit the plan before it starts, then open the sources |
| A spreadsheet | Claude for Excel: clickable cell references, and a warning before it overwrites | Add-ins for Excel and Google Sheets | Gemini in Sheets, for Workspace users |
| A document | Claude Docs in the conversation; in Google Docs, "Ask before edits" is the default | ChatGPT for Word, on every plan | Gemini in Docs, for Workspace users |
Worth knowing in each
- Claude: its numbers come with a way back to the source. In Excel, each answer cites the cells it used; in a dashboard, each live chart shows its query and refresh time.
- ChatGPT: you can review the research plan and limit it to sites you name before it starts, which makes the facts check smaller.
- Gemini: Gemini Notebook is built for answers drawn only from your own files, so you know where every answer should come from.
In your assistant
Your plan or your IT team may change what you see.
Claude
- A task in the conversation. Claude asks when the request is unclear and, by default, before taking an action. Long tasks carry on after you close the laptop; check progress from your phone, then read the result and the steps it took before you use it. Sources: Anthropic: Cowork is now Claude (16 Sep 2026); Anthropic: Dispatch
- Finished files. Results arrive as Claude Docs, Claude Slides or Claude Design inside the conversation, with one shareable link, and download as PowerPoint or PDF. The link makes sharing one click, so run the five checks before you send it. Sources: as above; Anthropic: Dashboards and motion (8 Oct 2026)
- Dashboards (beta, Pro, Max, Team and Enterprise). Each live chart built from your connected data shows its query and refresh time. Check both before you trust a number. Source: Anthropic: Dashboards and motion
- Claude for Excel (Pro and up). Its answers come with clickable cell references; it changes values while keeping formulas intact, finds the cause of errors and warns before overwriting. Click the references to check its numbers at the source. Source: Anthropic: Excel
- Claude for Google Workspace (public beta, Pro, Max, Team and Enterprise). A sidebar in Google Docs, Sheets and Slides, with "Ask before edits" as the default. Keep it that way while you review. Source: Anthropic: Claude now works in Google Docs Sheets and Slides (6 Oct 2026)
- Research. Open the sources and confirm the key claims before you share. Source: Anthropic: Using the research feature
ChatGPT
- Work. While a task runs, follow its progress, answer its questions, change direction and approve important actions. When it ends, review the result, then ask for changes in the same chat. Source: OpenAI: ChatGPT work and Codex
- Deep research. You review its plan before it starts, and you can limit it to sites you name. When the report is back, open the sources and confirm the key claims. Source: OpenAI: Deep research
- Office add-ins. ChatGPT for Word is on every plan, including Free; the add-ins for Excel, PowerPoint and Google Sheets are limited on Free and full from Plus. Save a copy before you let any add-in change an important file, and read each edit. Sources: OpenAI: Plans; ChatGPT release notes, 17 Sep 2026
- What it used. When an answer draws on ChatGPT's memory, "Sources" under the answer show what was used. Source: OpenAI: Memory FAQ
Gemini
- Task (personal accounts on Google AI Pro or Ultra). Read the result and the steps it took. Some actions need your confirmation, and it hands sign-ins back to you. Source: Google: Gemini Help
- Deep Research. Edit the plan, then choose "Start research". When the report is back, open the sources. Source: Google: Gemini Help
- Your own files. Gemini Notebook (renamed from NotebookLM on 16 Jul 2026) is built for answers drawn only from your own sources. Source: blog.google, NotebookLM is now Gemini Notebook (16 Jul 2026)
- Inside Docs, Sheets and Slides (Google Workspace editions). Gemini drafts, summarises, analyses a sheet and makes slides. Read every change before you share the file. Source: Google: Gmail Help
Which to use for review, and my pick
An honest view (as of Oct 2026):
- Claude makes the facts check easiest on numbers: cell references in Excel, and dashboards that show their query. In Google's apps it asks before edits by default.
- ChatGPT gives you the most control over a research report's sources before it starts.
- Gemini makes the check simpler when the answer must come only from your own files, through Gemini Notebook.
My recommendation: Claude first, ChatGPT second, Gemini third.
For work that will go out under your name, Claude shows you most plainly where its numbers came from. Whichever you use, the five checks are the same, and so is the rule: a named person signs it off.
Try it (10 minutes)
Use one piece of work an AI or an agent made for you this week: a summary, a draft, an edited sheet, a filed folder. You will review it in the tool where it already sits. Paste nothing new into any AI.
- The job (2 minutes). Write one line: what was "done" meant to look like? Does the result match?
- The facts (2 minutes). Pick the one fact that matters most and find it at its source, outside the chat. In Claude for Excel, click its cell reference.
- The steps (2 minutes). Open its steps, progress view or history. Find one step you would not have taken.
- The changes (2 minutes). Open whatever record of changes the tool or the file keeps: the steps it lists, the file's version history, or a copy you saved first. List everything it touched.
- The sign-off (2 minutes). Write your decision: send, fix, run again, or stop, and why.
If the tool showed you no steps and no record of changes, write that down too. It is the most useful thing you will learn today.
Check yourself
-
1An agent's summary says "All done, checks passed". What do you do first?
The answer: C When an AI tells you its work passed a check, ask what was checked; and asking "are you sure?" isn't checking. -
2Which is the clearest sign that a task's scope is creeping?
The answer: A Files or websites you didn't mention are the pattern to watch for; stopping to ask is the agent doing its job. -
3Claude for Excel changed some values in your workbook. What helps you check them (as of Oct 2026)?
The answer: A Anthropic's Excel page describes both: answers with clickable cell references, and a warning before overwriting. You still click through and check.
Read the whole lessonThe idea in full, the worked examples, common mistakes and what to check
The idea
A chatbot hands you words to copy. An agent hands you finished work: an edited spreadsheet, a new document, a filed folder, sometimes an email already drafted in your name. It looks done. That is the trap.
Small mistakes often don't stay small, because each wrong step becomes the starting point for the next one. The good ones check their work as they go, for example by rereading the calendar after booking instead of assuming it worked. But a summary that says "all done" is the agent's account of its own work, not a check of it.
And you stay responsible. Anthropic's safety page for Claude Cowork says you remain responsible for everything Claude does on your behalf, including messages sent, data changed and what scheduled tasks do (as of Oct 2026). The same holds in every tool.
Why it matters now. In a September 2025 survey of US full-time desk workers by BetterUp and the Stanford Social Media Lab, 40 percent believed they had received "workslop" in the last month: AI-made work that looks polished but is unhelpful, low-quality or off the mark. In Microsoft's 2026 Work Trend Index (a survey of 20,000 AI users in 10 countries, February to April 2026), quality control of AI output was the human skill people most often rated as more important as AI takes on more work, at 50 percent. Both are self-reported surveys, run by companies that sell workplace services.
Why it matters more in 2026. The agents now finish more on their own. Claude's tasks carry on after you close the laptop, ChatGPT's Work can keep going in the cloud after you leave, and Gemini's Task can run in a remote browser. You were not watching, so the review is where you catch up.
An illustration, not a real case: an agent is asked to update the regional sales tab and write two lines for Monday's meeting. Its summary says every region was updated and the totals were checked. Its steps show it also changed the forecast tab, which nobody mentioned, and that its "check" was that the totals column adds up, not that the figures match the source report. The summary was true and the work was still wrong.
The five checks
Run these in order, before anything leaves your hands. They are the same five checks in every Neosaipien kit.
1. The job: did it do what the brief asked? Put the brief's "done looks like" next to the result. Is it the right format, length and audience, and did it land where it should? Judge it against the brief, not against what you now wish you had asked for.
2. The facts: is every name, number, date and quote real? Open the sources yourself, outside the chat. When an AI tells you its work passed a check, ask what was checked. A pass on one part says nothing about the parts it did not test.
3. The steps: how did it get there? Read what it did, step by step, not just the summary at the end. Look for the step where it guessed, skipped something or gave up and worked around it.
4. The changes: what is different now? List every file, tab, email, folder and setting it touched, including ones you did not mention. Two questions are the right ones: is it accessing files or websites you didn't mention, and is the task's scope creeping beyond what you asked for? Use the tool's own record (below) rather than your memory of the file.
5. The sign-off: would you put your name to it? Decide one of four things: send it, fix it, run it again with a better brief, or stop. Whoever briefed the agent runs these checks, and "The AI did it" is never the answer to "who checked this?"
Scheduled runs, and tasks that carried on while you were away, need the same five checks on every run, not only the first.
Common mistakes
- Reading only the summary. The summary is the agent's own account. The steps are the evidence.
- Judging the result against what you want now, not against the brief's "done". If the brief was wrong, fix the brief and run it again.
- Accepting every change at once. Read each edit first.
- Trusting "checks passed" without asking what was checked.
- Asking "are you sure?" It tells the AI which way you're leaning; it does not check anything.
- Sharing the link before the checks. A Claude Docs link or a Work deliverable is one click from your whole team.
- Reviewing only the first scheduled run, then letting the rest go unread.
- Missing what else it touched: another tab, another file, an email it sent along the way.
- Forwarding it because it looks finished. Polished is not the same as right.
What to check before you trust it
- The result matches the brief's "done looks like".
- The facts that matter were found at their source, outside the chat.
- You read the steps, and you know where it guessed.
- You have a list of everything it changed, and nothing is there that the brief did not allow.
- Any "check passed" claim says what was checked.
- A named person signs it off, and it is not "the AI".
- For a scheduled task, someone reviews every run.
Check it. Run the five checks: the job, the facts, the steps, the changes, the sign-off.
Keep it safe. If the steps show it opened a file or a site you did not mention, stop the task and tighten its permissions before the next run. That is the lesson "What should an agent be allowed to touch?"
Checked against · read 16 Oct 2026Every claim, its source and the date read
As of: product steps as they were on the vendors' own pages on 16 Oct 2026.
Every product name, feature and step above comes from Neosaipien's platform reference, checked on these vendor pages on 16 Oct 2026.
- Anthropic, Claude Cowork and chat are now one Claude (16 Sep 2026)
- Anthropic, Run tasks in the background with Dispatch
- Anthropic, Build live dashboards and animate explainers with Claude (8 Oct 2026)
- Anthropic, Use Claude for Excel
- Anthropic, Claude now works with Google Docs, Sheets, and Slides (6 Oct 2026)
- Anthropic, Use research on Claude
- Anthropic, Use Claude Cowork safely (you remain responsible; files or websites you didn't mention; scope creep)
- OpenAI, ChatGPT Work and Codex
- OpenAI, Deep research
- OpenAI, ChatGPT plans (add-ins by plan)
- OpenAI, ChatGPT release notes (17 Sep 2026, ChatGPT for Word on all plans)
- OpenAI, Memory FAQ (Sources under answers)
- Google, Manage your tasks and workflows in Gemini Apps
- Google, Use Deep Research in Gemini Apps
- Google, NotebookLM is now Gemini Notebook (16 Jul 2026)
- Google, Get started with Google Workspace with Gemini
- Microsoft, 2026 Work Trend Index (5 May 2026; 20,000 AI users, 10 countries, fielded February to April 2026: its charts say 18 Feb to 20 Apr, its method note 18 Feb to 7 Apr)
- BetterUp, The hidden cost of AI "workslop" (29 Sep 2025; survey with the Stanford Social Media Lab, September 2025, US full-time desk workers)
Not legal advice. Where this lesson mentions rules, it gives plain-English orientation only; your organisation's own policy comes first.
Where next
How can AI help with your inbox without sending the wrong thing?
Use AI to summarise threads, find details and draft replies, and keep every send in your own hands or behind an approval.
Open the next lessonThe AI Fact-Check Kit, a five-step checking routine anyone can run on AI work: Sort, List, Trace, Hedge, Log, with review cards that include one for an agent's run.
For the whole team: Workshop in a Box: Check What AI Writes, a 60-minute session a team lead can run with their own team.
The Team workshop. Hands-on training built around your team's real work. I can run it live.
Book a 20-minute call