The week before a quarterly business review usually disappears into the same chore: hunting down numbers. You export revenue from billing, copy pipeline out of the CRM, pull product usage from analytics, chase support metrics, then paste it all into a deck and hope the date ranges match. By the time the slides exist, you are out of time to think about what they mean. An AI agent flips that. It can pull the quarter's metrics from your tools, assemble the deck or summary, flag what changed versus last quarter, and draft talking points, so you spend the meeting prep on decisions, not data wrangling.

This guide walks the QBR as a practical workflow, the same way you would brief a sharp analyst. If you are still deciding whether an agent fits the job at all, start with what an AI agent is; if you need to make the case internally, the framing in the executive business case for AI agents helps.

What a QBR agent actually does

A QBR agent is an agent pointed at one recurring outcome: turn a quarter of scattered metrics into a reviewable business-review draft. It reads the numbers from wherever they live, lines them up against the prior quarter and your targets, and writes the first version of the story. According to Anthropic's Building Effective Agents (2024), the agents that work in production are the ones built around a clear, narrow task, which is exactly what a quarterly review is.

The distinction worth drawing early is internal versus customer QBR. An internal review answers "how did the business or team do this quarter" for leadership: revenue, pipeline, KPIs, headcount. A customer QBR answers "what value did this account get, and what is next" for a specific client: their usage, their outcomes, their renewal risk. The workflow below is the same shape for both. Only the sources and the audience change. The agent does not replace your judgement; it removes the hours between you and the point where judgement starts.

1. Define the QBR outcome

Before any data moves, write down what a finished review looks like in two or three sentences. For an internal QBR: "A deck covering Q2 revenue, pipeline, product KPIs, and support health, each compared to Q1 and to target, with a one-line narrative per metric and a risks slide." That sentence is the contract for the whole run. A vague outcome is one of the most common reasons a review draft comes back wrong.

Why the outcome comes first

Naming the outcome first fixes the things the agent cannot guess: which period, which sections, what "good" looks like, and the format you actually present in. It also gives you the final check. If you cannot describe how you would know the review is complete, the agent cannot either. This is the same discipline behind running any agent well; the broader pattern is in how to set up your first AI agent. Decide the scope here, in plain words, and every later step inherits it.

2. Gather the inputs

With the outcome fixed, the agent collects each metric from its real source rather than a stale spreadsheet. Revenue and billing come from the finance tool, pipeline and bookings from the CRM, product KPIs from analytics, and customer health from the support platform. The principle here is blunt: a review is only as current as its weakest source, so the agent reads live wherever it can rather than trusting a copy someone exported last week.

Map each section to a source

The practical move is a short map: this section pulls from that system, with this date range. Q2 revenue comes from billing filtered to April through June. Net new pipeline comes from the CRM, same window. Support health comes from ticket volume and satisfaction scores over the quarter. Writing the map out surfaces gaps early, like a KPI no system actually tracks, before the agent runs. Scope every connection to read-only, on just those systems. A QBR agent never needs write access, and it never needs the systems the review does not touch.

3. Compile and compare

Gathering raw numbers is the easy half; the value is in the comparison. The agent normalizes everything to one period, then sets each metric beside its prior-quarter figure and its target, computing the deltas. A number on its own is noise. For example, something like "ARR up a few points quarter over quarter, ahead of target" is a finding. This compare step is the difference between a data dump and an actual review.

Quarter over quarter, and against target

Two comparisons carry most of the meaning. Quarter over quarter shows direction: is the line bending up, flat, or down. Against target shows whether direction is good enough. A metric can rise and still miss plan, and the agent should say so plainly. The agent groups results into the review's sections, attaches each delta to the right metric, and marks the items that moved most, because those are what the meeting will actually discuss. Estimating what a recurring run like this costs before you commit is covered in how to estimate agent cost before deploying.

4. Draft the narrative and risks

Numbers with deltas are still not a review. The agent writes the connective tissue: a one or two line narrative per metric explaining what changed and the plausible why, then a consolidated risks section pulling together the items trending the wrong way. This drafting step is where an agent tends to save the most time, because turning a grid of deltas into readable talking points is the part people dread most.

Internal versus customer framing

The narrative voice shifts with the audience. An internal QBR narrative is direct and self-critical: "Support backlog grew because we paused hiring; here is the plan." A customer QBR narrative is value-focused: "Your team ran 40% more workflows this quarter, and here is what we would do next." Same underlying deltas, different framing. The agent should draft to the audience you named in step one. The structured, multi-stage way an agent moves from data to a finished document mirrors the approach in an AI agent for contract review.

5. Hand a reviewable draft to a human

The agent's job ends at a draft, not a sent deck. It hands you a review you can scan in minutes: every metric sourced, every delta shown, the narrative written, the risks gathered, and crucially, every figure traceable back to where it came from. The human review gate is non-negotiable for anything that goes in front of leadership or a customer. The agent compresses the prep; you keep the call.

Make the draft easy to verify

A good draft is built to be checked, not just read. Each number should link or cite its source, the date range should be stated on the page, and anything the agent inferred rather than measured should be flagged as such. That way your review is a five-minute confirmation, not a re-build. If something looks off, you fix the source or the scope and rerun, rather than editing slides by hand. Understanding why a human stays in the loop, even for a capable agent, is the theme of AI agent versus chatbot versus assistant.

What to check before you trust it

An agent's review is only as good as what it reads, so two checks matter most before you rely on it. First, source accuracy: confirm each connected system is current and that the date ranges line up across them, since a billing export and a CRM pull on different windows produce a quietly wrong story. Most "the agent got it wrong" cases trace back to a stale or mismatched source, not the agent.

Scope the access tightly

The second check is access scoping. A QBR agent should read only the systems the review needs, and only in read-only mode. There is no reason for a review agent to hold write permissions or to reach systems outside the report. Grant the narrowest access that lets the job run, review the connections each quarter, and revoke anything no longer used. Tight scoping is what makes a recurring, automated review safe to leave running, and it is the same principle you would apply to any agent touching real business data.

How this runs on Gravity

On Gravity you do not configure the steps above by hand. You describe the outcome, the period, the sources, the sections, the format, and an expert-built agent that already encodes the gather, compare, and draft workflow does the work, typically handing back a draft in about 60 seconds once your sources are connected. Builders build and maintain these agents for Gravity, Gravity runs them and carries the cost, and you pay only when one runs, at $1 for 1,000 credits. You connect the read-only sources once, then trigger the same review each quarter.

Frequently asked questions

Can an AI agent prepare a quarterly business review?

Yes. An AI agent can pull the quarter's metrics from your tools, compile them into a deck or summary, compare each number to last quarter and to target, and draft talking points and risks. It hands you a reviewable draft, so your time goes to decisions instead of data wrangling.

What data does a QBR agent need access to?

A QBR agent needs read access to the systems that hold your quarter's numbers: revenue and billing, the CRM for pipeline, the product or analytics tool for usage, and support for ticket and satisfaction metrics. Scope it to read-only on just those sources, nothing wider than the review actually requires.

How does an AI agent build a QBR deck?

The agent gathers each metric from its source, normalizes the numbers to one period, then compares them against the prior quarter and your targets. It groups results into the review's sections, writes a short narrative per slide, flags what changed, and assembles the deck or summary for your review.

Is a QBR agent accurate enough to trust?

It is accurate when its sources are accurate and a human reviews the draft. The agent only reflects the numbers it reads, so check that each source is current and the date ranges line up. Treat the output as a strong first draft to verify, not a final report to forward unchecked.

How do I set up an agent for quarterly reviews?

On a platform like Gravity you describe the outcome: the period, the sources, the sections, and the format you want back. The expert who built the agent has already designed the gather, compare, and draft steps. You connect the read-only sources once and the agent runs each quarter from there.

Three takeaways before you close this tab

Sources