AI for Quarterly Business Reviews: Data Pulls, Sources and Human Checks
Use AI to prepare quarterly review packs from company data, with source links on every figure and a named human check before leadership sees the numbers.

Quarterly business reviews usually fail on traceability, not on writing. A pack gets assembled from exports, spreadsheets and slide notes, and by the time leadership reads it, nobody can say where a number came from or which filter produced it. AI can speed up the assembly by pulling approved data, drafting the pack and linking every figure to its source, but the pack is only useful when a named person checks the sources, definitions and narrative before it reaches the room.
Key takeaways
- Give the agent read access only to approved sources, and require a link to a specific export, record or report for every figure it reports.
- Agree on metric definitions and cutoff dates before the first run, so the agent does not quietly redefine a term such as active account.
- Keep three layers separate in the pack: source facts, calculations and commentary. A reviewer can then check each layer on its own terms.
- Make human sign-off a recorded step with a named reviewer, a date and a list of changes, not a quick look at the final slides.
- Judge the process by how many errors reach leadership and how quickly they are caught, not by how fast the pack appears.
Why quarterly packs lose their trail
Most review packs are built in the final week before the meeting. Someone copies revenue from one system, pipeline from a second and headcount from a spreadsheet last edited by a colleague who has since moved teams. Each number looks plausible, yet the pack has no record of which cut, which currency rule or which exclusion produced it. When a director asks a follow-up question, the team reruns the work under pressure and sometimes gets a different answer.
An agent does not fix this by itself. If it is pointed at the same folders and asked to write a summary, it will produce a fluent pack with the same gaps, and fluent text makes unsupported numbers look more trustworthy than they are. The goal is a pack where every claim can be traced and every judgment has an owner.
Illustrative scenario: an operations director wants the pack to show on-time delivery. The warehouse system counts a shipment as complete when it leaves the dock, while the customer portal counts it on delivery. Both numbers are correct for their own purpose. Without a rule that names one of them as the review metric, the agent may pick whichever source it reads first. The pack stays internally consistent and still answers the wrong question.
Decide what the pack must answer first
Before any data is pulled, write down the questions leadership needs answered. Keep the list short. A useful review pack usually covers results against plan, the two or three movements that explain those results, risks that need a decision and the requests the team needs from leadership. Each question should have an owner who will confirm the answer before the meeting.
- For each metric, name the system of record, the report or field, the filter and the person who owns the definition.
- Set the cutoff rule once, such as closed periods only or the last business day of the quarter, and state it on the first page.
- Record which adjustments are allowed, such as currency conversion or excluding test accounts, and who approves them.
- Separate decisions from facts. A proposal to change a sales territory is a decision item for leadership, while the numbers behind it are fact items the pack must trace.
Build a source map the agent can follow
A source map is a short document that ties each metric to a specific location in your systems. The agent should retrieve from approved sources only, and its answers should carry references to the documents they came from. Google Cloud's reference architecture for retrieval-based generative AI applications describes this pattern: the answer is grounded in retrieved material, and the material stays attached so a person can inspect it.
Keep the map current. An export from six weeks ago can look like current data, especially when it sits in a folder next to newer files. Each entry should record when the source was last refreshed and who replaces it. For more on this failure mode, see how stale context misleads AI agents.
Step by step: from raw data to a reviewed pack
- Pull the approved exports and record the run time, filters and cutoff in an appendix to the pack.
- Have the agent calculate each metric from the pulled data, and store the formula next to the result so a reviewer can rerun it.
- Attach a source reference to each figure, pointing to the export, record or report it came from.
- Flag every figure with a missing source, a changed definition or an unexpected movement as an exception for a person to explain.
- Draft commentary only for flagged items and for facts already in the pack. Mark any sentence that states a cause as a hypothesis until an owner confirms it.
- Send the draft to the named reviewer together with the open questions list.
- After sign-off, freeze the pack version and archive its inputs, so the same pack can be reproduced later.
Where the human check happens
The reviewer's job is to test the pack, not to reread it for style. A practical checklist looks like this:
- Open three or four figures at random and confirm that each link opens the stated source and shows the same number.
- Rerun one calculation by hand or in a spreadsheet and confirm the result matches.
- Check that the metric definition on the page is the one the owner approved for this quarter.
- Read every causal sentence and confirm that an owner stands behind it, or remove it.
- Look for stale inputs, such as exports older than the cutoff or a source replaced partway through the quarter.
- Confirm that no text in the pack came from an instruction hidden in a document, an email or a meeting note.
The last point deserves attention. Shared documents and emails can contain text written to steer an AI system, and an agent that reads them may follow that text. Treat retrieved content as material to summarize, never as a command. The OWASP prompt injection prevention guidance describes this class of risk and the controls that reduce it. The NIST AI Risk Management Framework also treats human oversight and documented decisions as part of managing AI risk, which fits the sign-off record described here. That record should name the reviewer, the sections checked and any figure changed after the agent's draft. It is the evidence you will need next quarter to see which parts of the process caught errors and which did not.
How a managed setup keeps packs reliable
Be Human designs this kind of workflow as managed infrastructure. An engagement starts by defining one useful role for the agent, such as preparing the quarterly review pack, together with the memory and tools it needs, its approval boundaries and an observable success criterion, such as how many figures arrive at the reviewer with a working source link. Monitoring, recurring evaluations against earlier packs, permission reviews when a source system changes and controlled updates to the agent's instructions keep the workflow aligned with how your company reports. The agent prepares and cites. Named people approve.
After each review, track three things: the corrections made after leadership saw the pack, the figures that arrived without a source and the reviewer time spent per section. OpenAI's evals guide describes testing model behavior against defined expectations, and the same idea applies here. Rerun the agent on an earlier pack whose answers are already settled, so you can tell whether a change to instructions or sources made the output worse. A similar approach to recurring reports is described in the guide to building an AI weekly report.
If you are not sure which part of your review cycle is ready to hand to an agent, the free Be Human AI diagnostic can help you find the first task worth handing over.
Frequently asked questions
How much of the pack can an AI agent write without review? Assembly, calculations and first drafts can be largely automated, but no figure or causal claim should reach leadership without a named reviewer. The agent is most useful for pulling data, calculating results, linking sources and flagging gaps. Reviewers keep authority over definitions and conclusions.
Which data sources should the agent be allowed to read? Only the systems and exports that your finance, operations or sales owners have approved for reporting. Read access should be scoped to the reports and folders the pack needs, and reviewed whenever a source system changes. Anything outside that scope should be excluded rather than left for the agent to discover.
What if the agent's number disagrees with what the team remembers? Do not let the agent resolve the difference by choosing the more plausible figure. Open the source reference, check the definition and cutoff, and record the reason for the difference on the exception list. Often the disagreement reveals a definition that was never written down.
How do we know the process is improving after the first quarter? Compare the corrections found after leadership review, the figures missing sources and the reviewer time per section across packs. Test the agent against earlier packs whose answers are known. If a change makes the results worse, roll it back and record why.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


