Forecast and budget conversations

Forecasts must be grounded in your history and your assumptions

Here is the rule that ought to be printed and pinned above every accountant’s desk before they open a chatbot to “help with the budget”: a forecast is only as good as the history and the assumptions it stands on, and neither of those can come from a language model.

The model has not seen your client’s sales records. It has not seen the seasonality of their Birgunj distributor’s order pattern, the way the Dashain bump arrives three weeks earlier in Province 1 than in Bagmati, the slow-pay habit of one specific government counterparty, or the fact that the managing director plans to add two regional sales executives in Magh. All of that lives in your spreadsheet, in your meeting notes, and in the partner’s head. A budget built without those inputs is fiction, however fluently presented.

What the model is allowed to do is structure the conversation. Which line items should be projected separately versus rolled up. What sensitivities are worth showing. Which scenarios — baseline, upside, downside — are conventional for the type of business and the audience. How the output should be laid out for a board pack versus a bank submission. The model knows the shape of a good budget conversation. It does not know the substance of yours.

A concrete pattern that works

The pattern that holds up across small-firm and SME work in Nepal has four steps.

1. Pull the last 12-24 months of actuals from your accounting software into a clean table — revenue, cost of goods sold, gross margin, operating expense lines, EBITDA. Twenty-four months is better than twelve because it lets the model see one full annual cycle including the Dashain-Tihar peak and the Asadh-end fiscal close. Twelve is acceptable if that is what you have.

2. List your own assumptions explicitly, in plain language: “Dashain bump assumed at +35% on Aswin revenue, in line with the prior two years. Input prices assumed to rise 8% from Magh based on the supplier letter dated 14 Poush. Two new sales executives joining in Magh, fully loaded cost NPR 1.2 lakh each per month. No major regulatory change assumed.” This is the part the model cannot supply. It is also the part that, if it is wrong, will make the budget wrong — which is why writing it out explicitly is itself useful, regardless of AI.

3. Hand both the historical table and the assumption list to the model and ask: “Project the next four quarters in a baseline scenario, an upside scenario (specify what changes), and a downside scenario (specify what changes), in a single table. Show monthly granularity for revenue and quarterly for the rest. Use only the figures and assumptions I have provided; do not introduce new assumptions without flagging them.”

4. Re-compute the output. Every projection in the table goes back to the spreadsheet. The model’s role is to lay out the structure and write the scenario narratives; the arithmetic, as always in this chapter, lives in the spreadsheet.

The “explain it to my partner” pattern

Once the scenario table exists, there is a second prompt that earns its keep on every budget cycle: “For each of the three scenarios, write a one-paragraph rationale in plain language suitable for a non-finance audience. Cover what assumption is driving the difference, what would have to happen in the business for that scenario to come true, and one risk that could push the actual outcome outside the scenario range.”

What this produces is the executive summary that normally takes an extra evening to write. The board chair can read three short paragraphs and understand the shape of the year without parsing a twelve-column table. The bank credit officer can extract the narrative for their committee paper. The managing director can use the same paragraphs in the next quarterly all-hands. One prompt, several downstream uses.

The verification rule applies here too — read the paragraphs against the table and against your assumptions. If the rationale says “downside reflects a 5% drop in volume” and your assumption list specified 10%, the model has drifted. Catch it.

Iterate explicitly, do not start over

The most common AI mistake during budget cycles is starting a new chat every time an assumption changes. The right pattern is the opposite: keep one working session, and feed changes in as explicit deltas.

“Change one assumption: salary increase moves from 8% to 12%. Hold everything else constant. Re-do the table and re-write the rationale paragraphs to reflect this change.”

The model handles incremental changes well. It will hold the rest of the structure stable and update only what should change. Over the course of a budget meeting, you can run four or five such iterations in the time it would have taken to rebuild one scenario from scratch in Excel. The partner sees the impact of each assumption shift in real time, which is the conversation the partner actually wanted to have but rarely got.

The honest limit

There is a temptation — especially after the third or fourth iteration goes smoothly — to let the model “help with the assumptions too.” Resist it. The model can list assumption categories it would expect in a budget of this type. It can flag if you have left out a line that competitors typically include. What it cannot do is tell you whether this client’s Dashain push will work, whether this managing director will actually hire the two sales executives, or whether this receivables pattern will hold up against monsoon disruption to the East-West Highway.

Those calls are the partner conversation. They were the partner conversation before AI and they will be after. What AI changes is that the spreadsheet work around the partner conversation — the laying out, the scenario modelling, the rationale writing, the re-iteration — drops from days to hours. The conversation itself stays the same length and stays exactly as hard. That is the honest limit, and it is also why the budget accountant who masters AI does not get replaced; she gets more time to be in the room when the conversation happens.

Check your understanding

Quick check

—

A junior accountant generates a three-scenario forecast for a client by giving the model only the prompt 'Build me a baseline, upside, and downside forecast for a Nepali manufacturing firm.' No historical data and no business-specific assumptions are provided. Is the resulting forecast reliable for board sharing?

1. AYes — large language models are trained on industry data and can produce reasonable forecasts for standard business types.
2. BYes, provided the model is asked to "be conservative" in its assumptions.
3. CNo — without the client's historical actuals and explicit business assumptions, the model has no grounding for the projections and is effectively inventing the business; the output may sound plausible but is not a forecast in any defensible sense.
4. DYes, as long as the accountant rounds the numbers before presenting.

Check answer→

What comes next

This chapter has been about turning numbers you already trust into prose, analysis, and projections that save the accountant time without surrendering professional judgement. The next chapter changes terrain. Tax, compliance, and Nepal-specific filings — VAT, TDS, income tax, IRD circulars, the Companies Act — are the area where the cost of an AI error is highest and the verification path is most specific. Chapter 4 takes the same discipline into that higher-stakes territory.