How a read-only agent works
A read-only agent is given exactly the access it needs to look, and none of the access it would need to act. It can pull yesterday’s ad performance, read the search terms report, open the analytics property, review a spreadsheet or crawl a site. It cannot pause a campaign, change a budget, edit a page, send an email or write to your CRM.
The restriction lives in the permission, not the instruction. Telling an agent not to change anything is a request; granting it a viewer role or a read-only credential is a wall. That distinction is the whole idea, because the failure modes that worry people — a misread instruction, a confidently wrong conclusion, an unexpected loop — all become harmless when the worst possible outcome is a bad paragraph in a report.
The output is therefore always a recommendation for a person: here is what changed this week, here is what looks wrong, here is what I would do about it. A human reads it and decides.
Why starting read-only matters
Analysis is where these systems are genuinely strong and where the risk is lowest. Reading every search term, every landing page or every product in a feed and surfacing the handful that need attention is tedious work that software does well and that people skip when busy. Nothing is at stake if it is wrong, because the recommendation is reviewed before it becomes an action.
It also settles the trust question with evidence instead of argument. Run the agent read-only for a while and you will know, from its own output, whether it understands your account. If its recommendations were sensible and a person would have applied most of them, you have a case for widening permissions. If they were plausible but wrong, you have learned that cheaply, and the cost was some reading time rather than a week of misdirected budget.
Where read-only agents go wrong
The most dangerous misunderstanding is that read-only means safe in every sense. It does not. An agent that reads customer records, enquiry data or anything containing personal information is still a data exposure question, even though it changes nothing. Decide what it may read, and where that data goes afterwards, with the same care you would apply to a new staff member.
The second is the report nobody acts on. A weekly analysis that lands in an inbox and is skimmed once is not automation; it is another thing to ignore. If a recommendation is not acted on, either the agent is not useful or the process around it is missing.
The third is confident wrongness. A read-only agent can misread a seasonal dip as a problem, or explain a change with a cause that is not real, and it will do so in fluent, convincing language. Recommendations still need judgement, especially when they concern something the agent cannot see, such as a client’s stock problem or a public holiday.
How to act on it
Start every agent here. Give it a viewer role on one account, using its own credential rather than a shared login, and set it a narrow question — search terms worth excluding, pages that lost traffic, products missing feed attributes. Judge its first weeks against what you would have concluded yourself.
Keep the recommendations somewhere a person genuinely reviews, and record which were applied and what happened, because that record is the argument for the next step. When you do widen permissions, widen them one action at a time and put an approval gate in front of anything that spends money. The underlying discipline is simply scoped access applied to software, and it is how an AI agent earns a place in a live account rather than being granted one on faith.
