What an AI agent does for you (and what it still can’t)

An AI agent does far more than answer questions, and far less than the demo videos suggest. It reads your data, combines sources, decides the next step within rules and takes actions in real tools. What it still can’t do well is sustain long chains of decisions without someone checking along the way.

Picture an ordinary Monday. The manager of an online store opens the laptop and finds three messages from the agent: Meta Ads CPA climbed over the weekend, two Google campaigns spent above their daily cap, and a best-selling product has four days of stock left. Nobody had to ask for any of it. Now picture the same manager telling the agent: “grow sales this month.” That’s where things fall apart.

The gap between those two scenes is what this post is about: what an AI agent actually does today, where it stumbles, and how to choose what to delegate without regretting it.

What does an AI agent do, in short?

An AI agent is a system that takes a goal, reads data from several sources, plans the steps, acts in tools (spreadsheets, CRM, ad platforms) and checks the result before moving on. It handles frequent, verifiable, rule-based tasks well. Long, vague or irreversible tasks still need human oversight.

If you want the more conceptual definition, with the history of the term, it’s in AI agents: what they are and what changes for your business. Here the focus is different: real capability versus promise.

How does an AI agent work under the hood?

Every agent, from the simplest to the most sophisticated, runs the same four-move loop. It perceives (reads data), reasons (decides the next step), acts (calls a tool) and verifies (checks whether the action produced the expected result). Then it repeats until it meets the goal or hits a stop rule.

What separates an agent from a chatbot lives in the third and fourth moves. A chatbot delivers text and stops. An agent executes and comes back to check. We covered that difference in AI agent vs chatbot.

What an AI agent already does well today

The list below isn’t speculation. These are tasks agents run in production, at companies of every size, with accuracy high enough to skip line-by-line review.

Task	Why it works	Marketing and sales example
Monitor and alert	Clear rule, easy-to-check outcome	Flag when CPA runs 20% over target
Combine data sources	Mechanical work people do badly	Join Meta Ads spend with CRM sales
Answer questions about data	Short query, verifiable answer	“Which channel brought the highest-ticket customer in October?”
Draft and summarize	A person reviews it in seconds	Weekly client report for an agency
Take small, reversible actions	Mistakes are cheap to undo	Pause an ad with frequency above 6

Notice the pattern: short, frequent tasks with an objective definition of right and wrong. That’s where an agent pays for itself. More cases by department are in AI agents: practical examples by business area.

A concrete mini case

An online cosmetics store in Recife, Brazil, set a target CPA of R$ 45 (about US$ 8) on Meta Ads. One Tuesday, the agent noticed the retargeting ad set was running at R$ 62, 37.8% over target, for three days straight. It cross-checked GA4, saw the product page conversion rate had dropped at the same time and concluded the ad wasn’t the problem: the page was, slowed down by a recent update.

The agent didn’t fix the site. It sent the team an alert with the hypothesis and the numbers. That’s the right design: it did the tedious investigation, and the decision stayed with the people who know the context.

What an AI agent still doesn’t do well

This is where most projects get frustrating. There are four limits, and the first one is math.

1. Long chains of steps

Say an agent gets each individual step right 95% of the time. Sounds great. But on a task with 10 chained steps, the odds of getting all of them right are 0.95 to the 10th power, roughly 60%. At 20 steps, it drops to 36%. This is an illustrative example calculated here, but the logic holds for any system where an early error contaminates everything after it.

The good news is that this limit is moving fast. A study by the research organization METR, published in March 2025, found that the length of tasks frontier models complete with 50% success had been doubling roughly every seven months. At the time, the ceiling sat at tasks of about 50 minutes of human work. Useful, but a long way from “run the whole operation.”

2. Context that isn’t written down anywhere

The agent doesn’t know December’s campaign has a thinner margin because the supplier raised prices over the phone. Or that a big client asked not to be called on Fridays. If the information isn’t in data it can read, it doesn’t exist for the agent.

3. A vague goal

“Grow sales” isn’t a task, it’s a wish. An agent with a vague goal and broad permissions tends to optimize the wrong thing: cutting the spend that brings in new customers to improve this month’s ROAS, for instance. In practice, this rarely works because nobody defined what counts as success.

4. Bad data

A fast agent on top of wrong data just makes mistakes faster. Messy UTMs, duplicated conversions and an outdated CRM turn into bad decisions made with confidence. Before delegating, take a look at the quality of the data that will feed the agent.

Why do so many AI agent projects never leave the pilot stage?

The 2025 numbers tell the same story. McKinsey’s annual AI survey, released in November 2025, found that 62% of organizations are already experimenting with agents, but only 23% are scaling them, usually in one or two functions.

Gartner was blunter. In a forecast published in June 2025, it estimated that over 40% of agentic AI projects will be canceled by the end of 2027, due to high costs, unclear business value or inadequate risk controls. The same analysis estimated that, among thousands of vendors calling themselves “agentic,” only about 130 offer real agents. The rest are chatbots or old automation with a new label.

The most common mistake we see is starting with the most ambitious process. A company wants an agent that “manages paid media” and discovers, three months later, that nobody trusts it enough to let it touch the budget. Starting with monitoring and diagnosis builds that trust. The full roadmap is in how to implement AI agents in your company.

How to decide what to delegate to an AI agent

Ask three questions. If a task passes all three, delegate it with autonomy. If it passes two, delegate it with human approval. If it passes one or none, keep it with the team and use the agent only to prepare the analysis.

1. Is it reversible? Pausing an ad is undone in one click. Emailing 40,000 contacts isn’t.
2. Is it verifiable? Is there a number that says whether it worked? “CPA is back on target” is verifiable. “Improve the brand” isn’t.
3. Is it frequent? A daily task justifies the effort of setting up rules. A yearly one almost never does.

This rule breaks at one point: tasks with regulatory risk, such as handling personal data or talking to customers in a regulated industry, call for human approval even when they pass all three questions.

Frequently asked questions

What can an AI agent do for me?

Monitor metrics and flag anything unusual, combine data from ads, website and CRM, answer questions about that data and take small actions, like pausing ads or updating records. The biggest gain is usually the time you stop spending on building reports.

Is there an AI agent that does everything?

Not reliably. The longer and vaguer the task, the higher the accumulated error. Agents that work have a narrow scope, clear rules and a person approving the actions that are hard to undo.

Does an AI agent replace an employee?

It replaces tasks, not whole roles. It takes data collection and cross-checking off the analyst’s plate and gives back time to interpret and decide. Companies that cut the person and keep only the agent usually lose exactly the context that isn’t in the data.

Is there an AI agent that makes phone calls?

Yes, voice agents already make qualification calls, appointment confirmations and simple collections. They work well with predictable scripts and poorly in negotiations with many variables. Let customers know they’re talking to an AI.

The right agent starts with the right question

An AI agent isn’t a digital employee or magic. It’s a way to take the reading, combining and watching of data off your team’s plate, so people spend their energy where judgment matters. Teams that start there see results in weeks; teams that start with “handle everything” tend to end up in Gartner’s statistic.

If your first step is to stop building reports by hand, Sherlok connects Meta Ads, Google Ads, GA4, your CRM and spreadsheets, lets you ask your data questions in plain language and get the analysis ready, with agents that alert you when something goes off track.