Working effectively with agents
Practical habits for getting the most out of Numen — how to ask, when to review, and how to build a routine you can trust.
You don't need to be technical to get great results from Numen. The people who get the most out of it treat an agent like a sharp junior analyst you direct and review — not a magic button you press once. This guide is the working style behind that, written for you as the person doing the finance work.
Ask a sharp question, not a vague one
The single biggest lever is the question. A vague ask gets a vague answer; a specific ask gets a specific, traceable one. Name the metric, the period, and the comparison.
| Instead of | Try |
|---|---|
| "Analyze my finances." | "Why did gross margin fall in Q2 vs. budget?" |
| "How are sales?" | "Show revenue by product for the last 6 months, and flag anything down over 10%." |
| "Make a report." | "Draft a board summary of budget vs. actual for the Manufacturing division this month." |
Set the context before you ask
An agent answers from what you point it at. Before a big question, take five seconds to set the stage — it changes the answer more than any clever wording.
- Pick the period and entity in the canvas header so the numbers cover the right scope.
- Use the right agent for the job — GL Data to load and clean data, Rules to define how things are calculated, then the analysis agents on top.
- Attach a file (Excel/CSV) when the data isn't in Numen yet, so the chat can work from it.
Review — don't rubber-stamp
The agent proposes; you dispose. It never posts an entry, changes a plan, or finalizes a report without your say-so — and you should use that checkpoint. Before you act on a number, open the reasoning trace to see how it got there and check the cited sources. In finance you are accountable for the figure, so the two-minute review is the point, not an obstacle.
Build once, reuse forever
Most finance work repeats — the monthly close, the board pack, the same variance review. Don't rebuild it each time. Save the widgets and layouts you like to your Library, and save a finished report as a template. Next month it's one click instead of an afternoon — and reusing a saved widget also spends fewer credits than generating a fresh one.
Keep the human in the loop
Agentic AI is excellent at the heavy lifting — gathering data, computing, drafting, watching for anomalies around the clock. You own the parts that need judgment: the assumptions, what counts as material, and the final call. The best results come from that split, not from handing over the whole decision.
A simple weekly rhythm
If you're not sure where to start, borrow this routine and adapt it:
- 1Refresh the data — upload or sync the latest ledger in GL Data so everything downstream is current.
- 2Ask what changed — open the chat or Anomaly Detection: "What moved most vs. budget this week, and why?"
- 3Drill into what matters — follow the one or two items worth your time; check the sources.
- 4Capture it — drop the key widgets into a report and save it, so the story is ready to share and easy to repeat next week.
Common pitfalls to avoid
- Vague prompts — "analyze this" wastes a turn (and credits). Name the metric, period, and comparison.
- Acting on a number without checking its sources — always open the trace for anything that drives a decision.
- Treating a forecast as a fact — it's a modeled estimate; pressure-test the assumptions.
- Rebuilding the same report every month — save it as a template once and reuse it.