How an AI agent actually works: a CMO’s guide

You want a shortlist of events for next spring. Not a 40-tab research expedition. Just three plausible options, their costs, deadlines and the reason each might put you in front of buyers.

That sounds like a job for an AI agent. But what does the agent actually do between receiving that request and handing back a document? If you can picture those steps, you can decide what work to give it and what should remain yours.

An agent works more like a concierge than a search box

Imagine asking a hotel concierge to find a restaurant. A search box returns restaurants. A decent concierge checks when you want to eat, looks at availability, notices that one place is closed on Mondays and comes back with options. If you’ve allowed it, the concierge might make the booking too. Finding, checking and booking are different permissions.

An AI agent follows a similar pattern. It receives a goal and instructions, uses the tools it has been given, looks at what comes back, and decides on its next step. That cycle can repeat several times. Anthropic describes it as planning, acting, observing and adjusting. A conventional automated workflow follows a route someone has already laid out; an agent can choose its route as it goes.

The concierge comparison has a limit: the agent does not have a seasoned marketer’s judgement tucked inside it. It can only work with the information, access and boundaries available to it. OpenAI’s September 2026 description of its Agents API separates the task, tools and working environment for precisely this reason. The system around the model shapes what it can do.

Watch it build that event shortlist

Suppose you ask an agent to find three European industry events for a proposed spring campaign. You give it a target audience, a budget ceiling and permission to read your event calendar. It can search public event sites and write a briefing, but it cannot contact organisers or commit spend. This is an illustration, not a report of a real deployment.

First, it interprets the request. Does “next spring” mean the next calendar spring or the next planning cycle? What counts as an industry event? If the brief leaves those open, the agent may make a plausible assumption. A good setup tells it which ambiguities it may resolve and which ones require a question.

Next, it chooses a move. Perhaps it searches for events, then checks an organiser’s page for dates and audience details. A search result is a lead, not proof. If the page shows last year’s event, the agent needs to notice and look again. If it cannot find a current sponsorship price, “price unavailable” is a better answer than an attractive guess.

Then it compares what it has found with your calendar and criteria. One event may have the right audience but clash with a customer conference. Another might fit the dates yet offer no evidence about who attends. The agent can return to a source, try a different search or stop and flag the gap. This is the useful bit of the loop: the next step responds to what the previous one revealed, rather than following the same checklist regardless.

Finally, it writes a shortlist with source links, unresolved questions and a note on what it checked. It stops. It has not booked a stand, emailed an organiser or spent a pound. Those would be separate actions requiring separate access and, in my view, explicit approval.

Where the apparently sensible answer goes wrong

The easiest mistake to spot is a made-up price. The more interesting failure is a reasonable-looking shortlist built on the wrong premise. Perhaps the agent treats general attendance as evidence that your buyers attend. Perhaps it reads a current event page but misses that the pricing PDF belongs to a previous year. Every step looks orderly; the conclusion still falls apart.

There is also a less obvious risk when an agent reads outside material. A webpage or document can contain instructions aimed at the agent rather than information for the task. Anthropic identifies these prompt-injection attempts as a risk when agents process material they do not control. An event shortlist probably does not need a security seminar, but it does benefit from a simple boundary: event pages supply facts about events, not new instructions about what your agent should do.

This is why “give it more context” is not automatically the answer. A folder full of old plans, inconsistent audience definitions and expired price sheets can create more chances to pick the wrong thing. Anthropic’s work on agent context makes the broader point: what an agent sees during each step affects what it does next. For a CMO, that means a clear brief and a few trustworthy sources can beat unrestricted access to every folder.

I like agents most when the job involves exploring, checking and coming back with a useful answer, while the cost of a wrong turn is still small. The distinction is almost comically ordinary: I’m happy for the concierge to find the restaurant. I still want to see the menu before they order for the whole table.

This blog was created by Jan’s AI content agent. Get in touch if you want to know how it works.