How to automate variance commentary with a Claude agent

Written up from the video AI Agents for Finance: Everything You Need to Know in Under 6 Minutes.

You can use a Claude agent to draft your monthly variance commentary in your own voice, in minutes, once you have given it the right context and a clear set of instructions. The instructions tell it how to think rather than what to write: which variances matter, how to explain the why behind each one, how to flag what it cannot evidence, and never to invent a figure. You then review the draft the way you would review a junior colleague's work, because the agent does the work but you remain accountable.

What an AI agent is, in finance terms

An AI agent is Claude set up with a clearly defined goal and the autonomy to work out how to reach it. Think of handing a task to a junior colleague: "prepare this month's variance commentary". You do not tell them every step. They pull the numbers, work out what matters, write it up and bring it back for you to check. You gave them the goal; they worked out the steps.

In a normal chat you guide Claude turn by turn. With an agent you hand over the whole goal and it carries the task from start to finish, then brings it back to you.

It is also not the same as automation. Automation has been around for decades: a task set to run on a timer, with no ability to think or make a decision. An agent can reason and make decisions to reach the goal, decisions you can then check. Whether it is triggered automatically or you press the button yourself is a separate question. It is an agent either way.

Some agents are built by developers and baked into products, such as fraud checks running across millions of bank transactions. The kind that matters most to finance professionals is built inside Claude in plain English. No code. You describe the job the way you would brief a colleague.

Step 1: create a project with a role

Create a project and name it for the job, for example Variance Commentary Expert. The name matters because you are creating a role, not a chat. I build mine in Claude Cowork; a lighter version works in a normal Claude project.

Step 2: give it the context

Give it what you would hand a new analyst on day one:

• The management pack, so it knows this month's numbers.
• The budget, so it knows what it is measuring against.
• Last month's commentary, so it knows the standard and the voice.
• A short note on the business, so it understands the context behind the numbers.

Step 3: write the instructions

This is the step that turns a one-off chat into an agent you can rely on every month. My instruction set covers eight things.

Role and reader

The agent is a senior FP&A analyst writing the variance section of the monthly pack for the founder and the board. It explains what happened, why, and what it means for the run rate. It is not a bookkeeper: it does not adjust the ledger or propose journals unless asked, and it says clearly when the data does not let it answer.

Source of truth

A fixed order of sources: the general ledger first, then the P&L and trial balance, then the approved budget, then last month's commentary, then the business context. If two sources disagree, the ledger wins and the difference is reported, not dropped. Budget means the approved budget, not the latest forecast. If a file is missing or the month is not closed, it stops.

Materiality

A dual test: a line is only discussed if the variance passes both a percentage threshold and an absolute threshold, with a separate test for margin movements in percentage points. As a starting point, 5% for a stable P&L or 10% for a small or noisy one; an absolute floor of roughly 0.5% to 1% of monthly revenue; and 1 to 2 percentage points on margins. Revenue, gross margin, EBITDA and net profit are always covered. Everything that fails the test collapses into one sentence: all other lines were within materiality. A line that is small this month but has moved the same way for three months running is flagged as a trend.

Explain the why, not just the number

For every material variance, six steps: quantify it; decompose it into the two or three transactions behind it; attribute a driver, such as volume, rate, mix, timing, headcount or a one-off; classify it as timing, one-off, mix or run rate, and say whether it reverses or repeats; test whether two variances describe the same event; and show the clean position where a one-off distorts a line. A sentence that only restates the arithmetic, such as "overheads were 11,500 over budget", is a failure, not a draft.

Output format

A fixed structure: a three-to-five sentence headline; an EBITDA bridge from budget to actual that must reconcile, with any unexplained residual shown as its own line; one short section per material line, each heading stating the conclusion; a brief note below EBITDA; a basis and reconciliation note; and numbered open questions for the founder.

Your voice

Plain professional English, conclusion first, short sentences, no filler words, and the number instead of an adjective. The single most effective tip: paste two or three paragraphs of your own past commentary into the instructions and tell it to match them.

Flag uncertainty

Flagged uncertainty is part of the deliverable. Where it cannot fully explain a variance, it says what it can evidence and what it cannot, marks the point as unverified or lists it as an open question, and labels any hypothesis as a hypothesis, with the evidence that would confirm it.

Hard rules

Never invent, estimate or interpolate a figure. Never guess a cause: if the ledger does not show why, say so. Show the arithmetic behind any derived figure. Never reuse last month's explanation without re-testing it against this month's ledger. Do not soften a bad result or oversell a good one.

Step 4: put it to work each month

Each month the prompt is one line: prepare this month's actual-versus-budget variance commentary, from these files, following the project instructions. In the video the agent produced a first draft, in my voice and focused on what mattered, in under a few minutes.

Where your judgement comes in

Do not send the draft straight to the board or your manager. Check it the way you would check a junior colleague's work, because the agent does the work but you are still accountable. You also bring what the ledger cannot: whether an explanation is the real reason or just a plausible one, what the budget owners told you, and what the board actually needs to hear. The open questions the agent raises are yours to answer.

What to build next

Once the idea lands, you start seeing agents everywhere: one that prepares you for the questions the board is likely to ask at the next meeting, or one that watches the cash position and warns you of a tight week ahead.