Library

1. Ask if you need real data at all. Most WFM questions do not. The shape of the problem is enough.
2. Describe, do not paste. "A table with date, queue, calls offered and handle time" works as well as the table.
3. Swap real for fictional. Rename queues and sites, shift dates, and scale volumes by a random factor you keep to yourself.
4. Strip the extras. File names, tab names, chart titles, screenshot headers and document properties often carry the client or employer name.
5. Know your tool. Check whether it stores or learns from what you type, and whether your organisation has approved it.
6. Keep a person in the loop. Recompute numbers, read the output as your employer would, and say when something is AI-assisted.

The Check before you post tool catches the obvious slips. It will never catch everything.

An intraday forecast splits the day into short blocks (intervals), usually 15 or 30 minutes, with expected contacts and handle time for each.

1. Turn volume into workload: calls × handle time ÷ interval length. The answer is in Erlangs, roughly "agents kept busy".
2. Convert workload into the people needed to hit your service target, then compare with who is scheduled.
3. Read the shape, not just the total. A day can balance on paper and still fail at 10:00 and 14:00.
4. Look hard at the first and last intervals. Openings and closings hide the biggest surprises.
5. Mark known events (promotions, outages, public holidays) so you can explain a miss later.

Try it in the staffing calculator.

Erlang C answers one question: how many agents do I need so that a given share of calls is answered within a given time?

• Inputs: calls in the interval, average handle time, service target (for example 80% in 20 seconds).
• Output: the smallest team that meets the target, usually with a cap on occupancy so agents are not flat out.
• Diminishing returns: the first extra agents help a lot, later ones barely move service level. This is why small teams struggle.
• Limits: it assumes random arrivals, no hang-ups, one queue and agents who can take any call. Real centres differ, so use it as a starting point.

Shrinkage is paid time when people are not available to handle contacts. It comes in two kinds: planned (leave, training, coaching, meetings) and unplanned (sickness, lateness, system problems).

1. List every kind of time away in hours, per week, from your schedule data.
2. Divide by total paid hours to get a percentage for each line, then add them up.
3. Gross up: people to schedule = people needed ÷ (1 − shrinkage).

Common slips: counting breaks twice, using last year's leave pattern, and ignoring that shrinkage is not flat across the day or the year.

• MAPE is the average of each period's percentage error. Easy to explain, but quiet days with small volumes can distort it.
• WAPE divides total absolute error by total actual volume, so busy periods count more. Often a fairer headline.
• Bias is whether you miss in one direction. A forecast that is high every week looks accurate on average but quietly wastes budget. One that is low burns service and people.

Measure at the level you staff. If you schedule by half hour, check accuracy by half hour, not just by day. Use the forecast accuracy check to see the three side by side.

Five parts, in this order:

1. Role: "You are an experienced workforce management analyst."
2. Context: only what is safe to share, described in general terms.
3. Task: one clear job, not five.
4. Format: bullets, a table, under 100 words, whatever you need.
5. Check: "Say what you are unsure about" and "Do not invent numbers."

See the prompt library for examples you can copy.