Variance analysis commentary is where many month-end packs lose days: the numbers are final, but nobody has yet written down why they moved. Most of that time goes on gathering evidence rather than applying judgement, which is exactly the part AI agents handle well. This article explains how agents draft budget vs actual variance analysis from ledger and operational data, how the draft should be structured so a controller can test it, and how review and sign-off should work. It also covers where the approach breaks down and when it is not worth doing.
1. What is variance analysis commentary, and why does it take so long?
Variance analysis commentary is the written explanation that sits beside the numbers in a management pack: revenue is behind budget because a key distributor delayed orders, freight is over because spot rates rose in the quarter. Finance directors read the management commentary before they read the tables. Boards often read nothing else. Yet in most organisations it is produced last, under time pressure, by analysts copying last month's wording and chasing budget holders by email for reasons they half remember.
The delay is rarely the writing. It is the investigation. To explain a cost centre overspend, someone has to open the general ledger detail, find the journals that drove it, check whether an accrual was reversed, match supplier invoices to purchase orders, and then ask the department why. Each step lives in a different system. By the time the evidence is assembled, the reporting deadline has arrived, and the explanation that goes out is whatever was quickest to defend rather than whatever was true.
2. How do AI agents draft budget-versus-actual commentary?
An agent starts where an analyst would: with the variance itself. It reads the trial balance against budget and forecast at the level the pack is reported, applies the variance thresholds finance has set, and decides which lines need an explanation. Then it drills down. For a revenue line it pulls invoiced volumes, prices and customer mix from the sales ledger and runs a price volume mix analysis. For an operating expense line it lists the journals, vendors and purchase orders behind the movement, separating timing items such as late accruals from genuine overspends.
Operational data is what turns a number into a reason. Ledger detail shows that maintenance cost rose; the work order history shows it was an unplanned compressor repair at one site. Take a distribution group with entities in Dubai and Riyadh: an agent can link a margin shortfall in Riyadh to a shipment of discounted stock recorded in the warehouse system, then check the credit notes that followed. The result is specific, sourced and short, which is what automated variance commentary should be.
Each sentence in the draft carries its evidence: the journal numbers, document references and the query that produced the figure. That traceability matters more than the quality of the prose. A fluent explanation with no link to source is a confident guess, and controllers are right to distrust it. Where the agent cannot find a cause in the data, it should say so plainly and route a question to the budget holder, rather than invent a plausible story to fill the gap.
3. The data foundation the commentary depends on
Agents can only explain what they can see. If cost centres are mapped inconsistently across entities, if the budget lives in a spreadsheet that does not share the ledger's chart of accounts, or if operational systems use different product codes, the agent will produce variance analysis commentary that is precise and wrong. Most failed attempts at automated explanations trace back to this, not to the model. Fixing the mapping is unglamorous work, but it is the work that decides whether the drafts can be trusted.
The practical answer is a governed data layer on Microsoft Fabric or a similar platform, where ledger, budget and operational data sit in one model with agreed hierarchies and definitions. Budget versions need to be locked and labelled, so the agent compares against the approved budget and not a draft someone saved later. Operational sources should be added in order of explanatory value: sales and inventory first for most businesses, then procurement, then headcount and HR data for payroll variances.
4. How controllers review and sign off agent-drafted commentary
Review should be designed, not assumed. The controller's job changes from writing to testing: does the explanation reconcile to the variance, is the cause real, and is it the main cause or merely a convenient one. A good review screen shows each variance, the proposed explanation, the share of the movement it accounts for, and the linked evidence. Anything the agent could not explain, or explained with low confidence, is listed first so attention goes where judgement is needed.
Controller review and sign-off then follows clear decision rights. The controller can accept, edit or reject each comment, and edits are stored alongside the original draft so the change is visible in the audit trail. Judgements such as whether a shortfall is timing or permanent, or whether a provision is needed, stay with people. Sign-off is recorded against the period, the version and the reviewer. This is how AITHENTIC governs its finance AI agents generally: a defined role, decision rights, escalation logic and a full audit trail.
One mistake to avoid is allowing commentary to change silently after sign-off. Once approved, it is a record that may go to the board, auditors or lenders. If a late journal shifts a variance, the agent should reopen the affected comment, redraft it against the new figures and send it back through review, keeping both versions. Teams that skip this end up with packs where the words and the numbers disagree, which does more damage to credibility than a late pack.
5. Where agent-drafted variance analysis commentary falls short
Agent-drafted variance analysis commentary is reliable when the cause is recorded somewhere. It struggles when it is not. A sales shortfall caused by a competitor's promotion, a deliberate decision to delay hiring, or a negotiation still in progress will not appear in any ledger. The agent may find the symptom, such as lower volumes at certain customers, but the reason sits in someone's head. The workflow therefore needs structured questions to budget holders, with answers fed back into the draft rather than pasted in by hand.
There are also cases where this is not worth doing. A single-entity business with a short pack and a handful of material variances each month may be better served by a disciplined template and a good analyst. The return comes with scale: many entities, many cost centres, several currencies and a tight reporting deadline. Teams should also resist using agents to generate more commentary. Explaining every line below materiality adds reading time and hides the movements that matter.
6. Building commentary into the month-end rhythm
Commentary works best when drafting starts before the close is finished. Agents can produce provisional explanations on soft-close figures, so budget holders are asked about large movements while they still remember them, and the final draft becomes a refresh rather than a fresh start. This fits naturally with efforts to reduce month-end close cycle time, because the month-end reporting pack stops being the last bottleneck after the ledger is locked.
Good commentary should also feed forward. The causes it identifies, such as a supplier price increase, a delayed contract or a shift in product mix, are the same drivers a driver-based rolling forecast needs. If the agent records each cause in a structured form, tagged as timing, one-off or ongoing, FP&A can decide which belong in the next forecast instead of rediscovering them a month later. Over time this builds a history of explained variances that makes the budget itself more realistic.
Where teams start matters. A sensible first scope is one reporting pack, one region and the operating expense lines that consume the most analyst time. A related pattern, monthly spend anomaly reporting to business unit heads, is often a good companion: agents tell each owner what moved and why before the formal pack arrives, so many month-end questions are already answered. Extend to revenue and margin once the data mapping and review routine have held steady for a few closes.
Key challenges
- Evidence for a single variance sits across the ledger, budget files and operational systems, so investigation consumes the reporting window.
- Inconsistent mappings between budget, chart of accounts and operational codes produce explanations that are precise but wrong.
- Commentary copied from prior months repeats stale causes and misses new ones.
- Budget holders are chased late, when they no longer remember why a cost moved.
- Edits made after sign-off leave packs where the words and the numbers disagree.
The framework
A workable operating model has four layers. Data: one governed model holding ledger detail, locked budget versions and operational sources ranked by explanatory value. Drafting: an agent that applies agreed variance thresholds, decomposes each material movement and writes a sourced explanation, flagging gaps instead of filling them. Review: a controller queue ordered by risk, with accept, edit and reject actions recorded against the original draft. Sign-off and feedback: approval logged by period, version and reviewer, structured causes passed to FP&A for the forecast, and any change after approval sending the item back through review.
How to implement it
- Agree variance thresholds, reporting levels and a standard comment structure with the controllers who will sign off.
- Map budget, chart of accounts and operational codes into one governed data model, and lock approved budget versions.
- Pilot the drafting agent on one region's operating expense lines for a few closes, comparing its drafts with the analysts' commentary.
- Build the review queue with evidence links, confidence flags and recorded accept, edit and reject decisions.
- Add structured questions to budget holders for causes that do not appear in any system.
- Extend to revenue and margin with price volume mix decomposition once mapping and review have proved stable.
Frequently asked questions
Can AI write variance analysis commentary?
Yes. AI agents can write a first draft of variance analysis commentary, provided they can read the ledger, the approved budget and the operational systems where causes are recorded. The draft should quantify each material movement, attribute it to specific transactions or drivers, and link to the evidence. A controller still reviews, edits and signs off, because judgements about timing, permanence and provisions should stay with people.
How do you write good budget vs actual commentary?
Good budget versus actual commentary states the variance, the main cause and whether it is timing, one-off or ongoing, in one or two sentences. It should explain most of the movement rather than a convenient part of it, reference the transactions behind it, and say what happens next. Avoid repeating last month's wording and avoid explaining lines that fall below your agreed variance thresholds.
What data do AI agents need for variance analysis?
At minimum, agents need general ledger detail at journal level, the approved budget in the same chart of accounts and hierarchies, and the latest forecast. Explanations improve considerably once operational data is added: sales orders and invoices for revenue, inventory movements for margin, purchase orders for cost and headcount data for payroll. Consistent mappings across these sources matter more than the volume of data.
Is automated variance commentary accurate enough for board packs?
It can be, once a controller has reviewed and signed it off, and not before. Accuracy depends on clean mappings, locked budget versions and every statement being traceable to source. Agents should flag what they cannot explain rather than guess. Most teams run agent-drafted variance analysis commentary alongside analyst commentary for a few closes before relying on it for board or external reporting.
