Blog · AI and self-service analytics
Agentic analytics lets software coordinate an analysis. Set the scope, keep calculations in tools, and decide what a person must approve before acting.
Agentic analytics means delegating a sequence of analytical steps within limits. An agent can choose a permitted tool, inspect the result and decide which permitted check comes next. For a finance or commercial owner, the question is how much work to delegate before a person must review the evidence and decide what to do.
There is no single business specification behind the label. Ask a provider to demonstrate the sequence it actually performs. A useful starting point for one desk is an approved set of exports, a named question and a stopping rule. Covirage's tools compute and check every figure from the supplied data; the external AI model only explains the computed results.
The existing AI data analyst guide covers the software role and tool evaluation. AI data analysis covers the division between language and calculation. This guide owns the delegation decision: which next steps may software take, and what needs a person?
“Review September's cost movements” leaves too much open. A bounded assignment is: compare approved September actuals with budget by workstream, list movements above the agreed review threshold, retain the offsetting lines, and draft questions for the owners. Stop if inputs do not reconcile or a cause is unsupported.
Specify the source files, period, budget version, permitted comparisons and recipients. Permission to read a file does not authorize changing the source system, moving budget or distributing the result. Agree those actions separately with the responsible person.
A program finance lead supplies synthetic September 2026 workstream totals in USD. For this exercise only, the lead sets a review threshold of an absolute movement of at least $20,000. That threshold is a review choice, not an accounting standard.
Movement = actual cost − approved budget.
Review flag = absolute movement ≥ the agreed dollar threshold.
| Workstream | Budget | Actual | Movement | Threshold flag | Follow-up the agent may draft |
|---|---|---|---|---|---|
| Migration | $200,000 | $250,000 | $50,000 | Yes | Ask the owner to confirm the invoices behind the movement |
| Testing | $150,000 | $135,000 | −$15,000 | No | Retain the favorable offset in the total |
| Training | $100,000 | $120,000 | $20,000 | Yes | Ask whether timing or scope evidence exists |
| Total | $450,000 | $505,000 | $55,000 | Total reconciles | Prepare the review with all three lines visible |
Migration and Training contribute $70,000 of unfavorable movements; Testing offsets $15,000, leaving $55,000 over budget. The lead can delegate the ranking and draft questions. The result does not establish that Migration has a vendor problem or that Training's increase will recur. Those claims require evidence and an accountable owner. The full movement method remains in variance analysis.
| Work | What an agent may do within the assignment | What a person must check or approve |
|---|---|---|
| Interpret the question | Propose a measure, period and breakdown | Whether those choices answer the business question |
| Select sources | Use the files and fields authorized for this task | Correct export, completeness, access and cutoff |
| Choose calculations | Call approved tools with visible inputs | Right definition, scope and comparison version |
| Continue the investigation | Inspect permitted supporting rows and compare approved segments | Whether the next step remains within the assignment |
| Explain results | Draft commentary grounded in computed outputs | Every figure agrees; causes and assumptions have evidence |
| Recommend a response | Present options and state missing evidence | Business tradeoffs, ownership and approval to act |
| Share or change something | Prepare a proposed next step for review | Authorized audience and explicit approval for the action |
Stop the sequence if a required field is missing, totals fail to reconcile, definitions conflict or the next action exceeds permission. The agent should identify the gap. The business owner resolves it before relying on the result; see when the assistant cannot compute an answer.
Keep the question, source version, definition, tool inputs, returned figures and reviewer decision together. Repeat the same assignment on the same approved data: the figures should agree even if the explanation is worded differently. Then test a missing input and an out-of-scope request to see whether the sequence stops appropriately.
For Covirage, the calculation rule applies throughout: tools compute; the external model explains. Delegating more steps does not transfer arithmetic into generated text. It also does not transfer a manager's authority over a budget to the model.
A broad goal becomes broad permission. Define the allowed sources, comparisons and actions before work begins.
The total is right but the question is wrong. A reconciled result against an obsolete budget still needs correction.
Several plausible steps become a claimed cause. A chain of comparisons can locate a movement without establishing why it occurred. Keep association and cause separate.
A draft recommendation is treated as approval. A person must judge the evidence and authorize operational or financial changes.
For a business desk starting with exports, see self-service analytics. Related reading: prescriptive analytics and action choices and versioned metric definitions.
NIST's AI Risk Management Framework, checked October 17, 2026, provides general guidance on managing AI risks and trustworthiness. The delegation checklist above is a business review method, not a certification or a statement that NIST prescribes these exact steps.
An approach in which an AI agent coordinates a sequence of analytical steps toward a goal, rather than only answering one isolated question. The useful test is what it may do, which calculation tools it uses and where it must stop for a person.
An agent can select a follow-up tool based on an earlier result within its permitted scope. A chat interface alone does not establish that capability or guarantee numerical accuracy. Check the actual tool calls and results.
A named business owner remains responsible for the definition, the evidence and any action. An agent can draft a recommendation, but a computed difference does not establish its cause or authorize a budget change.