Somewhere on your computer there is a note holding your best prompt. You know the one: the paragraph you paste into an AI chat at month-end, refined over months until the commentary comes out the way you like it — the format, the thresholds, the rule about never guessing. It is quietly one of the more valuable documents in your finance department, and it comes with a monthly toll: find the note, paste it in, re-attach the files, and spend twenty minutes reminding a very smart tool of things you told it thirty days ago.

If that sounds familiar, it is because this column has spent six weeks working exactly that way: bring the task to a chat window, explain it well, check the output. That workflow has one weakness, and it shows up on exactly the tasks finance does best — the ones that repeat. This week we fix it, using a feature most Microsoft 365 businesses already have sitting in the Copilot app: the ability to build your own agent, no code, in about an hour. We will build one that earns its keep in every finance department on the island — a variance commentary agent.

A prompt you write once: name and description, instructions, knowledge, and a starter prompt
— your standard, written down and put on duty. (Branded graphic by PGH Consulting, LLC)

A prompt you write once

An agent, in Microsoft’s world, is nothing exotic: a prompt you write once, saved with a name, a set of standing instructions, and a folder of reference files attached. Instead of re-explaining the job in a chat window every month, you open the agent and say “do June.” Everything you would normally paste in — the format you want, the thresholds that matter, the rules about what it must never do — is already there, applied every time, for everyone who uses it.

That last part matters more than the time saving. When the commentary standard lives in one person’s notes, it leaves when they go on holiday. When it lives in an agent, the whole team drafts to the same standard, including the analyst who joined last month.

The monthly retyping tax: find the note, paste the prompt, reattach the files, re-explain the
rules — twenty minutes before the work even starts. (Branded graphic by PGH Consulting, LLC)

The ones Microsoft already built

Before building your own, it is worth knowing what ships in the box. The Microsoft 365 Copilot app comes with prebuilt agents, and two are directly useful to finance: Researcher, which digs through your files, emails, and the web to assemble a sourced brief, and Analyst, which reasons over raw data like a junior data scientist — give it a messy sales export and it will run actual code to summarise and chart it. There is an agent store with more, and your administrator decides which ones are switched on. Spend ten minutes with Analyst on a copy of last month’s workbook; it is the fastest way to feel what these tools can do.

But prebuilt agents are generalists. The real value for a finance team is an agent that knows your format, your thresholds, and your rules — and that one you have to build yourself.

Building the variance commentary agent

In the Microsoft 365 Copilot app (on the web, in Teams, or in the Office apps), go to the agents section and choose to create a new agent. You can describe what you want in plain English and let Copilot set it up, or skip straight to the configure screen and fill in the fields yourself. Either way, four fields do all the work.

Set it up once inside Copilot: name it, write the rules, attach the knowledge, and give the team
a starter prompt — then test before you publish. (Branded graphic by PGH Consulting, LLC)

Name and description are simple: call it “Variance Commentary” and describe what it does in a sentence.

Instructions are the heart of the agent — you get about 8,000 characters, which is more than enough to be specific. Here is a starting block you can adapt:

“You draft monthly budget-versus-actual commentary for a Jamaican distribution business. Work only from the workbook the user attaches and the reference files in your knowledge. For every line that is more than 10 per cent or J$500,000 off budget, state the variance in dollars and per cent, then the driver if the data shows it. If the data does not show the driver, write ‘requires management input’ — never speculate. Follow the structure and tone of the sample commentaries in your knowledge. Spell out ‘per cent’. No adjectives about performance; the numbers carry the story.”

The agent’s first draft: variances stated in dollars and per cent, one line flagged “requires
management input” — the agent writes the first pass, finance still owns the judgement. (Branded graphic by PGH Consulting, LLC)

Knowledge is where the agent gets its memory: you can attach up to 20 sources, including SharePoint folders and files. Attach the folder holding your monthly packs, plus two or three past commentaries you were proud of — that is how the agent learns the house style without you describing it.

Starter prompts are the buttons users see when they open the agent: “Draft this month’s commentary from the attached actuals” is the obvious one.

Then test it before anyone else touches it. The builder has a try-it tab; give the agent a month you have already written up, and compare its draft against what you actually sent. The gaps you find go back into the instructions — you are editing the standard, not correcting the same mistake in chat every month. When it holds up, publish it and share it with your team the way you would share a document.

What it costs
The full experience — building agents grounded in your files, plus the prebuilt Researcher and Analyst — comes with the Microsoft 365 Copilot licence, about US$30 per user per month on an annual commitment, on top of a Microsoft 365 business plan. That is real money in Jamaican terms, roughly J$4,700 a month per seat, so start with the people who own recurring reporting rather than the whole office. If your business is on an eligible Microsoft 365 plan without the add-on, Copilot Chat still lets you build simpler agents with pay-as-you-go metering, with knowledge limited to SharePoint items and public websites — a reasonable way to trial the idea before committing to licences.

A note on security and data

Two things to check before an agent goes anywhere near the management accounts. First, an agent inherits Microsoft 365 permissions: it can only read what the person using it can read. That is reassuring, but it cuts both ways — if your SharePoint permissions are untidy, an agent will cheerfully surface a salary file that a curious user technically had access to all along. Tidy the folder permissions before you attach the folder. Second, whoever you share the agent with can see what its knowledge sources produce, so attach the specific folders the job needs, not the whole finance drive. Your administrator controls which agents can be created and shared at all; on a work tenant, that conversation comes first.

What it will not do

The agent will not know why the numbers moved. It can see that distribution costs ran 14 per cent over budget; it cannot know about the fuel surcharge or the new route to Montego Bay, and if your instructions let it guess, it will guess fluently. That is why the “requires management input” rule is the most important line in the instruction block. It will not fix a messy workbook — if the actuals tab is half-linked and the budget column is out of date, the agent inherits every error. And its first month is a supervised run, not a hands-off one: every figure in every draft gets checked against the pack before anything leaves finance, this month and every month after.

What to try this week
1. Open the Microsoft 365 Copilot app and try the prebuilt Analyst agent on a copy of last month’s management workbook — ask it what drove the biggest variance, and check its answer against what you know.
2. Before touching the agent builder, write your instruction block in a document: format, thresholds, the rules, and what the agent must do when it does not know.
3. Create the agent, paste in the instructions, and attach a folder with two or three past commentaries as knowledge — good ones, because they become the template.
4. Test it on a month you have already written up, and compare its draft line by line against what you actually sent.
5. Fix the gaps by editing the instructions, then share the agent with one colleague first — the whole team can have it once a second pair of eyes trusts it.

Work on copies, keep salary and other restricted files out of the agent’s knowledge, and verify every figure in every draft against the pack before the commentary goes anywhere.

Peta-Gaye Hardy is the founder of PGH Consulting, LLC, where she helps finance and operations teams adopt AI in practical, low-risk ways. She writes the weekly AI in Finance & Business column and is based between Jamaica and the United States. Learn more at www.pghconsultinggroup.com. Follow on Instagram @pghconsultinggroup.

Disclosures: This article is informational and does not constitute investment, tax, legal, or accounting advice. AI tools can produce errors and figures should always be verified against source data. Some features described require paid subscriptions — building fully featured agents and using the prebuilt Researcher and Analyst agents requires a Microsoft 365 Copilot licence (about US$30 per user/month on an annual commitment) on top of a Microsoft 365 business plan; lighter agent features are available on eligible plans with pay-as-you-go metering. The author has no commercial relationship with Microsoft or any product mentioned and was not compensated by them. Readers should consult a qualified professional before acting.