AI at Work: What to Delegate and What to Check

You can delegate routine work to AI: drafting plans, suggesting alternative phrasing, summarizing, and sorting information. Check anything that affects decisions or their consequences: facts, calculations, conclusions, data confidentiality, and copy intended for clients or colleagues.

The key is to give AI tasks with clear outcomes and keep human oversight wherever accuracy, context, and accountability matter. Here’s how to draw that line at work and check the results without rechecking every little thing.

• 0 unconfirmed commitments should remain in a draft sent to a client
• 3 fields to check in a summary: decisions, open questions, and next steps
• 4 types of sensitive data you should never paste thoughtlessly into a public chat: passwords, access codes, card details, and secret keys

Which work tasks can you delegate to AI?

You can delegate drafting, summarizing nonconfidential materials, explaining formulas, and grouping anonymized reviews to AI—as long as an employee checks the result against the source material and makes the decision themselves.

Drafting and editing

ChatGPT or Google Gemini can draft a reply to a client from key points or turn a long draft into a task list. Before sending, check the tone and commitments: a model may soften a refusal, add a promise, or change the meaning of the original text.

Reformatting text is generally safer than asking AI to make a decision on an employee’s behalf. For example, ask it to extract action items from a draft, but not to decide who should be assigned what.

Summaries and spreadsheets

You can ask Microsoft Copilot or another company-approved service to summarize nonconfidential notes and list decisions, open questions, and next steps separately. Don’t upload confidential materials unless company policy allows it.

• Explaining a formula: Ask Microsoft Excel to explain what a formula does, then compare the answer with the formula itself and the cell values.
• Grouping reviews: Provide anonymized reviews and ask the AI to sort them by topic, then check its conclusions against the original entries.
• Summarizing notes: Cross-check decisions and next steps against the notes so you don’t mistake an omission or model error for a fact.

How can you get a verifiable result from ChatGPT or Copilot?

Set clear parameters for the model

You can get a verifiable result from ChatGPT or Copilot by specifying the task, audience, format, and constraints, then checking the answer against the source material. For example, ask it to draft an email from key points and explicitly tell it not to add new promises. That makes it easier to check that the message doesn’t include commitments you never made.

Ask the model to separate information from the material you provided from its own assumptions, and to flag gaps explicitly. For a summary, specify four fields in advance—decisions, open questions, owners, and deadlines—and tell it not to fill in missing information by guessing.

• Facts: Claims that can be found in the original key points or notes.
• Assumptions: Conclusions not supported by the source material.
• Gaps: Missing information, such as the name of an owner or a deadline.

Compare the result with the original

Check a ChatGPT or Copilot summary against the original notes: compare every decision, open question, owner’s name, and deadline with the source. If you can’t verify an item in the material, correct it or flag it for clarification—don’t copy it into a work document as an established fact.

First, test an Excel formula on a few rows where the result is already known. Compare the calculations with the expected values, and only then apply the formula to the entire spreadsheet. This can catch errors before cells are filled in at scale.

What should you check before sending text or analysis?

Before sending text or analysis, check specific claims against the source material, recalculate the figures, compare conclusions with the notes, and remove anything you can’t verify. Check names, dates, amounts, product names, and quotes against primary documents: AI can confidently fill gaps with invented details.

Repeat calculations in the original spreadsheet or a calculator: a convincing explanation of a formula doesn’t, by itself, confirm the result. In a summary, check whether the model has turned a question into a decision or attributed a conclusion to the wrong person. Verify these points against the notes, not just the finished text.

Before sending anything externally, read the draft as its author: remove unconfirmed facts and check every remaining promise against the document or agreement it relies on. If there’s no confirmation, clarify the wording or delete it. AI’s confident tone doesn’t make a claim reliable.

What data should you never send to AI thoughtlessly?

Never send passwords, access codes, bank card details, or secret keys to a public chatbot. Personal data, contracts, and internal documents should only be uploaded to a service that company policy approves for handling those materials.

Before processing a work document, check whether your organization has approved the specific service and what data-processing settings are in effect. If approval is unclear, don’t upload the document. First, check the rules with the people responsible for corporate services or information security.

Anonymizing data before analysis

When analyzing reviews, keep only the text needed for the task and remove information that could identify a client or order. Check the material against these specific categories:

• Names and contact details: Remove names, phone numbers, and email addresses.
• Order details: Remove order numbers, as well as any passwords, access codes, or secret keys that may have ended up in the text.
• Work materials: Don’t share contracts or internal documents unless the company has approved their use with the service you’ve chosen.

When does AI make mistakes and delegation fail?

AI makes mistakes when you expect it to check primary sources, resolve contradictions, or make independent decisions with consequences. Text from ChatGPT or another model is only a draft: check factual claims, legal conclusions, and calculations against documents and source data. A person responsible for the decision must confirm prices, deadlines, and commitments to the client.

What to check before using an answer

• Facts and calculations. Compare every significant claim with a primary source, and recheck calculations using the original numbers. A model’s persuasive tone doesn’t prove it’s accurate.
• Conflicting notes. If materials list different deadlines or terms, a summary may smooth over the discrepancy but can’t determine which version is correct. Check the records against a document or the person responsible.
• Terms for the client. Don’t ask a model to confirm a price, completion date, or commitment. Review the draft before sending and confirm each term with someone authorized to approve it.
• Confidential information. Don’t enter it into a public chatbot unless the company has explicitly approved using that tool for such data. If there’s no approval, use a company-approved tool or don’t share the information with the model.

How can you integrate AI into your workflow without losing control?

Start by using AI for one recurring task, check every result, and leave the decision to a person. For example, ask ChatGPT to draft an email or summarize anonymized notes while keeping the original text. That way, you can compare what changed and check whether any important details were lost.

For the chosen task, decide in advance who checks the result and who sends it to the recipient. AI suggests an option, but it isn’t responsible for its accuracy and doesn’t own the decision. Don’t send an email without checking it.

Services and fact-checking

ChatGPT, Google Gemini, and Microsoft Copilot aren’t interchangeable when it comes to data-handling rules. Use only a service approved for the specific type of data. If internal notes can’t be shared with the tool you’ve chosen, anonymize them first or don’t upload them at all.

For tasks where accuracy matters, keep the source next to each fact and remove anything from the answer you can’t verify. This is especially important for summaries: polished text may contain a claim that wasn’t in the original notes. Before sending, check the result against primary sources—not just for consistency with the requested style.

Frequently asked questions

Key takeaways

• Delegate a draft to ChatGPT or Copilot, not the act of sending an email to a client.
• Check summaries against the original notes, especially decisions and deadlines.
• Remove names, phone numbers, and order numbers before analyzing reviews.
• Test Excel formulas on rows with results you already know.

Written byValentina Soloveva

Валентина занимается новостями технологий и гаджетов, уделяя особое внимание последним трендам и инновациям. Она стремится делиться своими наблюдениями и анализом с читателями, чтобы помочь им оставаться в курсе быстро меняющегося мира технологий.