Most monthly business reviews follow the same pattern. In the days after close, analysts reconcile numbers from several systems, paste them into slides and write commentary for every page. The meeting then walks through the pack page by page, and the questions that matter most end with a promise to come back next week. Much of the effort goes into producing the pack rather than discussing it.
Generative AI can change each of those steps. Assistants such as Copilot in Power BI can already summarize a report page in plain language, and the same technology can draft variance commentary, answer follow-up questions and assemble a narrative. The benefit is real, but only if the review is redesigned around one governed version of the numbers. Otherwise AI simply produces fluent commentary on numbers nobody has agreed on.
From a stack of slides to a short narrative
The review document becomes two things. The first is a short narrative of a page or two: what happened, why it happened and what needs a decision, in the order the executive team should read it. The second is a governed report that holds the detail, with every number traceable to its source.
The narrative is read before the meeting, so the meeting can be spent on decisions rather than on reading. Slides do not disappear entirely, but they stop being the product. The product is a shared understanding of the month, reached faster.
A useful test for the narrative: an executive who reads only the first paragraph should know whether the month is on track, and one who reads the whole page should know what is being asked of them. Everything else is reference material.
What to automate
The most time-consuming part of the pack is also the most automatable: explaining movements that the numbers already explain. AI drafts these well when it works from a governed model with a clear variance bridge. It can:
- Draft the variance commentary. Rank the largest movements against plan and last year, and explain each one from the drivers in the model, such as volume, price, mix and cost lines.
- Summarize each business unit. One paragraph per unit, in the same order every month, so readers find the same information in the same place.
- Track open actions. List what was promised last month and what has changed since.
- Check the numbers in the text. Compare every figure in the draft with the certified model before anyone reads it.
Drafting works best when the bridge is complete. If a movement cannot be explained from the drivers in the model, the draft should say so and flag it for the owner rather than invent a cause.
What stays human
AI explains what the data contains. It does not know what the data leaves out: the customer negotiation that delayed an order, the price increase announced but not yet in effect, the one-time cost that will not recur. Deciding whether a variance matters, whether it is a trend or noise and what to do about it remains the work of the people accountable for the results. The same is true of the outlook. AI can extend a trend, but whether the second half will look like the first is a judgment about customers, competitors and the organization’s own plans, and it belongs to the people who will be held to it.
The practical rule is simple. AI writes the first draft; the owner of each business unit edits it and signs off, and the commentary goes out under their name. Owners learn quickly where the draft is reliable and where it needs their context, and the time saved on writing goes into the forward view.
One version of the numbers
The biggest risk in an AI-assisted review is not a badly written sentence. It is a well-written sentence about the wrong number. If the AI reads exported spreadsheets, last month’s file or a different cut of the data from the one in the report, the narrative and the pack will disagree, and the meeting will be spent reconciling them.
The safeguards are straightforward. The AI reads only the certified model that the review uses. Every number in the narrative is tied to a measure in that model and checked automatically before release. The narrative states its data cut-off and says so when the books are not yet closed. The draft is produced only after the certified model has refreshed, never from a spreadsheet someone updated by hand. The variance bridge lives in the model, not in a side spreadsheet, because executives ask why a number moved far more often than they ask for the number itself, a point that applies equally to rolling forecasts.
Questions answered live
The meeting changes too. When a question comes up, the answer can come from the same governed model, through a question typed in plain language or an analyst working in the report on screen. “We will come back to you” becomes the exception rather than the rule. This works only if the model holds what the meeting needs: plan, forecast and prior year side by side, with the drivers behind them. Preparation shifts from building slides to making sure the likely questions can be answered.
Keep a log of the questions asked in the meeting. They show what next month’s narrative should cover, and which definitions still confuse people. Over a few months, the log also shows whether the narrative is answering questions before they are asked.
How to start
Pick one review, ideally one where the numbers are already governed and the owners are willing. Run AI drafts in parallel with the current process for a few cycles, and compare each draft with the commentary the owners actually published. Agree on a house style: lead with the number, sign every variance, name the driver. Keep a handful of the best past commentaries as examples; they teach the AI the house style faster than any written instruction. Measure the hours spent before and after, and the edits owners make, so the case for change rests on evidence.
Then switch, and retire the slides that the narrative and the governed report have replaced. Tell the executive team what changed and why, so the first narrative is read as a deliberate improvement rather than a shortcut. The review keeps its monthly rhythm. It simply starts from a draft instead of a blank page.