What is agentic analytics, and when is it useful?

A growing order backlog raises several questions for an analyst. Which groups account for the increase? Are the periods comparable? Did a late refresh distort the numbers? The report shows the movement; answering those questions establishes what it means for today's work.

An agent can take on parts of that investigation. Its value depends on whether the explanation helps someone make a better decision, and how much work that person must do to check it.

The work an agent takes on

An agent works through a question by inspecting data and choosing further checks from what it finds. The tools available to it perform the calculations and queries. The model decides which of those tools to use next, subject to the instructions and limits set for the investigation.

In a fixed workflow, the next step is already specified: refresh the table, check the threshold, send the alert. An agent has room to choose a different follow-up based on the answer it receives. Anthropic describes these two approaches as predefined workflows and agents directing their tool use in Building effective agents.

The flexibility is useful when the next sensible question depends on the answer to the last one. It also makes the investigation less predictable. The agent needs boundaries for what it may read, how long it may continue and what evidence it must return.

When a fixed workflow fits

Keep the scheduled query when the question already has a dependable sequence of checks. Repeated calculations, threshold alerts and regular reports often fit that pattern. The expected output can be tested directly, and a failed step is easier to locate when the sequence is known.

Choosing the next step during an investigation brings a trade-off. The agent may run more queries, take longer or produce an explanation that needs correcting. Limit its queries and execution time. Then compare the additional findings with the time and attention required to review them.

Leave the action with someone authorized to take it. An agent's finding can help a manager decide about staffing or operations while the agent remains read-only. Before automating an action, define its conditions and permissions and assign responsibility. For some work, delivering the evidence to a person is enough.

Choosing a useful investigation

Start where people repeatedly spend time finding out what changed before they can act. Identify who needs the answer and which decision it will inform. Otherwise, an automatically generated explanation can become another item for someone to read without changing the work.

Look at the current investigation before choosing the technology. How often does it occur? Which steps require judgment? Where does the analyst spend time, and what does the business owner do with the answer? A recurring question with several possible causes is a stronger candidate than a request for a summary of every measure in a report.

Keep the first scope small enough to evaluate. One operational review, a defined set of sources and an agreed comparison give the team a way to judge usefulness. Expanding the agent's access does not establish that it understands more of the business.

Deciding who acts on the finding

Define what the recipient will do differently when the explanation arrives. Work backwards from that decision to the measures, comparisons and evidence it requires. The process owner and the people who currently investigate the question should help set those instructions.

They need to settle the meaning of a change as well as the calculation. A late record, a changed status and a deterioration in performance can produce similar movements in a chart. Specify the relevant dates, excluded records and conditions that call for review. These definitions establish the business question the agent should investigate.

Decide who receives the finding and who may act on it. Permission to inspect an order does not authorize changing its priority. If the investigation is read-only, make the handoff to the responsible person explicit. Also define what happens when the data is stale, a query fails or the agent cannot support an explanation. An incomplete run should reach someone who can resolve it.

Building a reviewable explanation

Use completed investigations to test the agent. Your team knows what happened in those cases and can compare the explanation with its own review. Include normal periods and exceptions, especially cases where the first plausible explanation turned out to be wrong.

Read the query as well as the answer. Were the population, dates, joins and statuses right? Does the result support the explanation? Finding many delayed orders in one group is a reason to look there. It does not, on its own, explain why the orders were delayed.

The reviewer needs a way to follow the evidence without starting the investigation again. Include the comparison, relevant records or totals, the executed query and any open question. Each part of the explanation should be traceable to something the agent read or calculated.

Then break the normal path. Try a missing input, an unavailable permission and an old extract. Does the agent stop, or does it continue with an answer that looks complete? Check what the run reports to its operator. The same information will be needed to diagnose a failure after rollout.

Running it in daily work

The pilot needs an owner for its move into regular use. That person must arrange for the inputs to be ready before the investigation starts and for the output to reach its intended recipient. The recipient needs to know how to review the evidence and what the explanation can support.

Save the instructions used on each run with its tool calls, queries and output. Keep versions so that changes in behavior can be investigated. Repeat the relevant evaluation cases when a definition, source structure or operating process changes.

During rollout, show recipients how to question an explanation and report a problem. Follow their use of the result: the time spent checking it, the corrections it needs and the findings they can act on. These measures reveal whether the agent reduces investigation work after accounting for review.

Ongoing support can cover the infrastructure, the investigation itself or improvements to the wider process. Agree who owns each part. A working data connection and a completed agent run establish that the system operated; the business owner still needs to judge whether the result helped.

Example: investigating a growing order backlog

This illustrative morning review investigates the change in orders awaiting dispatch so an operations manager can decide what to check next.

The agent chooses follow-up checks within an agreed review. It reports evidence for the manager's decision and leaves order priorities unchanged.

Before the first rollout

1. Choose the decisionIdentify the person who needs the finding, the decision it supports and the investigation that person currently relies on.
2. Set the boundariesRecord the definitions, sources, permitted tools and conditions for stopping. Decide who handles a failed run and who authorizes action.
3. Test known casesUse completed investigations to compare queries and explanations. Include failures in the test before adding questions or widening access.
4. Assign ongoing ownershipAgree who prepares the data, changes instructions, trains recipients and provides support. Review the investigation's usefulness with its recipients after rollout.

Sources