We have spent the past year running AI agents on real client work: research, first drafts, reporting, outreach and a long list of small jobs nobody enjoys. Some of it worked better than we expected. Some of it failed in ways that would have embarrassed a client if a person had not caught it.
Where agents earn their keep
- Research. Scanning search results, competitor sites, ads and reviews, then summarising what changed this week.
- First drafts. Posts, ad variations, email sequences and outlines, written to a clear brief and brand guide.
- Repurposing. Turning one interview, report or webinar into a month of content across formats.
- Reporting. Pulling numbers from several tools into one weekly summary with the changes highlighted.
- List building. Finding and enriching target accounts and contacts against a precise profile.
Where they still need a person
Agents are confident when they should not be. They will state a statistic that does not exist, flatten a brand voice into something generic, or miss that a perfectly reasonable sentence would offend a specific client. Judgement, taste, relationships and anything that makes a claim on the client's behalf stay with people.
The rules we work by
- Every output is reviewed by a person before it reaches a client or the public.
- Agents get narrow jobs with clear inputs, never open-ended goals.
- Every fact an agent produces must come with a source we can check.
- We measure agents on time saved and errors caught, and switch off the ones that do not pay back.
Used this way, agents let a focused team produce the volume of a much larger one without lowering the bar. Used carelessly, they produce a lot of confident noise. The difference is almost entirely in how the work is designed.
Written by Thar Labs
