AI demand planning for a small business without an analyst

6 days ago
3 min read

Updated: 4 days ago

Introduction

Demand planning sounds like a function that requires a department. In practice, the version that transforms a small business is modest: knowing roughly what next month looks like, in units that matter, early enough to do something about it. Most firms operate without this and absorb the consequences as normal — the panic order, the overtime week, the discount to fill a gap, the stock that did not sell.

The useful thing about the problem is how much of the gain sits in the simplest methods. Seasonal patterns from your own history, plus the leading indicators you already collect, will get most small businesses most of the available benefit. The sophisticated methods matter at a scale most firms never reach.

1. AI demand planning for a small business starts with your own history

Two or three years is enough.

Sales, orders or bookings by week, by product group. The seasonal shape emerges quickly and is usually stable, and that shape alone answers a large share of the questions you have.

2. Plan in the unit that constrains you

Not in money.

If capacity is the constraint, plan in hours or slots. If stock is, plan in units. If people are, plan in shifts. A revenue forecast is not actionable because you cannot order, hire or schedule in pounds.

3. Aggregate to a level you can actually forecast

Individual items are noise.

Demand for a single line is erratic; demand for a product family is predictable. Forecast at the family level, then allocate within it using recent share. This is the single most common technical mistake and the easiest to fix.

4. Add leading indicators you already have

They convert history into a forecast.

Enquiries, quotes outstanding, bookings on the diary, subscriptions renewing, orders on standing arrangements. These respond to what is happening now rather than to what happened last year.

5. Record and exclude the one-off events

Otherwise they contaminate everything.

A single unusually large order, a lost month, a competitor closing, a promotion. Left in the history, each one distorts next year's forecast for that period. A note against the record is enough to handle it.

6. Forecast a range and state your confidence

More useful than a point estimate.

A likely case with a high and low, and a note on which items you are least certain about. The decisions that follow — how much to order, whether to hire — are risk decisions, and they need the range.

7. Use the forecast to make one decision at a time

The value is in the action.

This month it might be the reorder quantity. Next month it might be whether to take on a temporary member of staff. A forecast produced and filed changes nothing; a forecast attached to a specific recurring decision earns its cost.

8. Track forecast error and look for bias

The improvement loop.

Forecast against actual, monthly, expressed as a percentage. A consistent direction of error is correctable arithmetic and the fastest improvement available. Random error tells you the achievable precision.

9. Keep it simple enough to be maintained

Complexity is the usual cause of abandonment.

A method one person can run in an hour a month will survive holidays and busy periods. An elaborate model that only its author understands stops being updated within a year, and a stale forecast is worse than none.

Set expectations about accuracy. Small businesses with few large customers face irreducible variance, because one customer's decision moves the month. In that case the plan should be about response speed rather than prediction.

Conclusion

Start with your own history at family level, because that is where most of the achievable accuracy sits.

Plan in the unit that constrains you rather than in revenue, aggregate above the level of individual items, add enquiries and bookings as leading indicators, annotate and exclude one-off events so they do not distort next year, publish a range with a note on your confidence, attach the forecast to one specific recurring decision, track error monthly and correct any consistent bias, and keep the method simple enough that it survives a busy month.