Five checks for AI-assisted client work

Use five short checks on every piece of AI-assisted client work before you send it: source, instruction, fit, risk, and record. The checks take about five to ten minutes for a normal task. They catch the errors that clients notice first.

This matters because AI output often looks finished when it is not. A clean paragraph can hold a wrong date, a made-up number, or a tone that does not suit the client. The client sees your name on the work, not the tool’s name.

What the research says (and does not say)

Microsoft’s 2026 Work Trend Index (published May 5, 2026) surveyed 20,000 knowledge workers who use AI. Half of them (50%) named quality control of AI output as a top human skill. Also, 86% said they treat AI output as a starting point, not a final answer. These are self-reported answers from AI users in ten markets. The Philippines is not one of them. Microsoft also sells the tools it studies.

People are poor judges of their own speed with AI. A randomized trial by METR (July 2025) tested 16 experienced software developers on 246 real tasks. The developers expected AI to make them 24% faster. After the study, they believed it had made them 20% faster. The measured result was that tasks took 19% longer. The sample is small, the tools are from early 2025, and the work was coding. Do not apply the numbers to virtual assistant work. Use the lesson: feeling fast is not the same as being fast, or right.

A March 2026 paper by Huang, Xiao, and Vishnoi (revised June 2026) gives a theoretical reason to care. Their model suggests that small differences in how reliably a worker verifies AI output can lead to large differences in work quality. It is a mathematical model, not a field study. It supports a simple idea: the person who checks well gains the most from AI.

The five checks

Source check. Open the original document. Compare every name, date, number, and quote in the AI output with the original. If you cannot find a claim in a source, remove it.
Instruction check. Read the client’s request again. Tick each requirement. Look for missing items, wrong format, and wrong length.
Fit check. Read the output aloud. Does it sound like the client’s brand? Does it use the right spelling (US or UK) and the right level of formality?
Risk check. Ask: if this is wrong, who is harmed? Money, legal text, health information, and messages to the client’s customers need a slower check. Ask the client before you proceed if you are not sure.
Record check. Write one line in your task log: what the AI did, and what you changed. This helps you if a client asks later.

A hypothetical example

This example is made up to show the method. A virtual assistant uses AI to summarize a 40-minute client call. The summary says, “Client approved the vendor contract by Friday.” The assistant runs the source check. In the transcript, the client said, “I will decide by Friday.” The summary changed a plan into an approval. The assistant fixes the line and flags the open decision. Without the check, the client’s team could have acted on a decision that was never made.

Start here

Pick one repeat task, such as meeting summaries. Use only the source check and the instruction check for one week. Add the other three checks when the first two feel normal.

Advanced extension

Keep a simple error log. For each task, write the type of error you found: wrong fact, missing item, wrong tone, or other. After four weeks, count the types. You will see where each tool is weak for your work. Then you can tell a client, with evidence, which tasks you use AI for and which you do by hand.

Tradeoffs and limits

Checks cost time. For low-risk tasks, such as a first draft that you will rewrite, a shorter check is fine.
The checks do not find every error. They reduce errors. They do not remove them.
Some clients do not allow AI use on their files. Ask first. Follow the client’s rule.
The research above comes from other countries and other jobs. It shows a pattern. It does not measure your results.

Do this week

Choose one task you will send to a client this week. Run all five checks. Time yourself. Write down how long the checks took and what they found.

Sources

2026 Work Trend Index Annual Report, Microsoft WorkLab, May 5, 2026.
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, Becker, Rush, Barnes, and Rein (METR), arXiv:2507.09089, July 2025.
Delegation and Verification Under AI, Huang, Xiao, and Vishnoi, arXiv:2603.02961, March 2026 (v2 June 25, 2026).