How much cross-training is enough?
One second skill for everyone captures most of what full flexibility offers, but only if the skills join the queues into a chain. Simulated, not guessed.
Published
Everyone agrees cross-training helps. Nobody agrees how much of it to do. Train too few people and the flexibility never reaches the queue that needs it; train everyone on everything and you pay for skills that go stale and work that takes longer. This guide puts numbers on the trade, using a simulation rather than an opinion, because no formula covers it.
The centre used throughout: four queues, 75 contacts each in a half hour, five-minute handle time, 80 per cent answered in 20 seconds, 57 agents split across the queues in proportion to their work.
What perfect flexibility would be worth
Start with the ceiling. Erlang C says each of those queues needs 17 agents on its own, so 68 in total. The same contacts handled as one pooled queue need 57. Separating the work costs 11 agents, about 16 per cent of the requirement, and that is the prize cross-training is chasing.
It is a ceiling, not a forecast. It assumes anyone can take anything at the same speed, which no real centre manages. The interesting question is how much of the 11 agents a realistic skill design actually reaches.
One second skill for everyone gets most of the way
Simulating the same centre with 57 agents, nobody cross-trained, gives a service level of 70.5 per cent. Training everyone on everything reaches 80.4. That 9.9-point span is the whole prize; here is how the designs in between divide it.
| Skill design | Service level | Share of the prize |
|---|---|---|
| Nobody cross-trained | 70.5% | 0% |
| 1 per queue has a second skill | 72.3% | 19% |
| 2 per queue | 73.8% | 34% |
| 3 per queue | 74.9% | 45% |
| 8 per queue (about half) | 77.4% | 70% |
| Everyone has one second skill | 79.2% | 88% |
| Everyone knows every queue | 80.4% | 100% |
Everyone holding exactly one extra skill captures 88 per cent of what total flexibility offers. The third, fourth and fifth skills between them are worth the remaining 12. This is not a quirk of the example: the same result appears in the academic work on skills-based routing, where a centre of agents with two skills each needed 91 agents against 89 for agents who could do everything, and 108 for specialists.
The practical reading is that cross-training is a breadth exercise, not a depth one. A training budget spent giving every agent one more skill beats the same budget spent making a few people universal.
Which second skill matters as much as how many
Here is the part that gets missed. Take the same number of second skills and arrange them differently, so that queues one and two cover each other and queues three and four cover each other. Every agent still has exactly two skills. The result drops from 79.2 per cent to 76.1: the same training investment captures 57 per cent of the prize instead of 88.
The reason is that flexibility only pays if it can move work to wherever the spare capacity happens to be. Two closed pairs of queues can never send work across the middle, so the centre still behaves as two smaller centres, and small centres are exactly what pooling was supposed to cure. A chain, where each queue’s cross-trained agents cover the next queue along, joins everything into one.
So when you choose second skills, do not ask which pairing is most natural. Ask whether your skill map is connected. If you can draw a line through every queue by following shared skills, it is. If it falls into islands, you are paying for flexibility you cannot use.
The limit: cross-skilled work is slower
Everything above assumes an agent is equally fast on both queues. They usually are not. Repeat the simulation with secondary work taking 10 per cent longer and the picture changes:
| Skill design | Service level, no penalty | With a 10% penalty |
|---|---|---|
| Nobody cross-trained | 70.5% | 70.5% |
| 3 per queue | 74.9% | 73.9% |
| About half per queue | 77.4% | 74.8% |
| Everyone has one second skill | 79.2% | 73.6% |
| Everyone knows every queue | 80.4% | 72.6% |
With the penalty, training everyone on everything is the worst of the cross-trained options. The best design becomes a partly cross-trained team. Once every agent can take every contact, every contact risks landing on someone slower at it, and past a point that costs more than the flexibility earns.
This is the same effect that makes pooling a poor idea for queues with very different handle times. The published rule of thumb is that pooling helps on average while the longest handle time is at most about five times the shortest, and helps every queue only below about three.
How big is your own penalty? Nobody knows. There is no published figure for how much slower agents are on a secondary skill, which is why the simulator asks you rather than assuming. Measure it if you can: compare handle time on primary and secondary work for the agents who already hold both.
What it saves in agents
Service level is the wrong currency for a business case. Ask instead how many agents each design needs to hit 80 per cent. In the same centre: 61 agents with nobody cross-trained, 56 with everyone holding a second skill. Five agents, about 8 per cent of the team. With the 10 per cent penalty on secondary work, the saving falls to three.
Notice this is smaller than the 11 agents Erlang C promised. Two reasons. Erlang C assumes nobody ever hangs up, which makes the separate queues look worse than they are; and no real skill design reaches perfect pooling. Both gaps are real, and quoting the Erlang figure to a finance director you will later have to face is a bad trade.
Try it on your own queues
The simulator splits a headcount evenly and cross-trains the same number per queue, which is the right way to ask how much is enough. Once you know, switch the team to groups and describe the one you actually have: ten agents whose first skill is billing and second is sales, five the other way round, twenty who only do support. It runs that team against the same people with their second skills removed, so the gain is what those second skills are worth.
Simulating…
Running the first batch…
Deciding this for real, across every interval and every skill? Pebble WFM schedules multi-skilled teams against the forecast. Free month, no card needed.
A short checklist
- Work out the ceiling first with the queue pooling calculator. If separating your queues costs two agents, no skill design is worth the training budget.
- Prefer breadth over depth. One extra skill for everyone beats a few universal agents, by a lot.
- Check your skill map is connected. Draw the queues, draw a line for each shared skill, and make sure you can walk from any queue to any other. Islands waste the investment.
- Measure the slowdown on secondary work before committing to universal cross-training. It decides whether the last step is worth taking, and it is usually the step people take first.
- Route to the narrowest qualified agent. Sending work to the most specialised person who can take it keeps the flexible ones free, and in the simulation it was worth about a point of service level on its own.
- Re-check when volumes change. The value of pooling shrinks as queues grow. A saving that justified the training last year may not survive the queue doubling.
Frequently asked questions
- Is it better to train a few people on everything, or everyone on one more thing?
- Everyone on one more thing, by a wide margin, as long as the second skills join the queues into a chain. In the four-queue centre simulated here, giving every agent one extra skill captured 88 per cent of what training everyone on everything would have bought. Training a handful of people deeply captured far less: three per queue reached 45 per cent.
- Does cross-training help the queues equally?
- It helps the worst-off queue most, which is usually the point. Before cross-training, the gap between the best and worst served queue in the example was 8.0 percentage points. With one second skill each it was 0.7. Small queues are the ones drowning in their own randomness, and flexibility rescues them first.
- What if cross-skilled agents are slower on the second queue?
- Then there is a limit, and it arrives sooner than people expect. At a 10 per cent slowdown on secondary work, the best design in the example was no longer full flexibility: training everyone on everything scored 72.6 per cent, worse than a partly cross-trained team at 74.8. The flexibility is still worth having; training everyone on everything is not.
- How do I know what my own numbers are?
- Run them. The simulator on this page takes your queues, handle times and headcount and reports the gain with a margin of error attached. Anything smaller than that margin is not a real difference, however much you want it to be.
Stop doing this one interval at a time
Pebble WFM forecasts your demand, computes the staffing requirement for every interval, builds the roster and publishes it, with self-service for agents and a copilot that can do what a planner can. Explore a sample organisation on day one.