Planning a hybrid workforce: when part of your capacity is software

Intermediate level · ~7 minute read · AI and the workforce planner, part 2 of 5

Until recently the capacity side of a plan had one input: people. Now there’s a second workforce that handles a share of contacts, never takes a break, and sometimes fails all at once. Planning it with the tools built for the first is the most common mistake in AI-era capacity plans.

A cautionary tale about believing the headline

In February 2024 Klarna announced that its AI assistant was doing the work of 700 customer service agents and had handled around 2.3 million conversations (about three quarters of chats) in its first month. By May 2025 the company’s chief executive was telling Bloomberg that the AI-only approach had produced lower quality, that Klarna was recruiting human agents again, and that “there will always be a human if you want”. Headcount had fallen 22% over the freeze, to 3,500, mostly through attrition.

The point isn’t that AI service fails. It’s that the number a vendor or a board sees in month one (containment, volume handled) is not the number that determines your human capacity needs in month eighteen. A plan that takes the early figure at face value and strips out the headcount has no way back when quality, complaints or repeat contacts say otherwise.

Three assumptions that quietly break

Arrivals are no longer independent. Erlang models assume contacts turn up randomly and independently. When an AI agent reaches a confidence or policy limit, or the system behind it has a bad hour, escalations arrive in clusters: a burst of customers who have already spent several minutes with a bot arrives in the human queue together. The human queue’s variance now includes the bot’s failure modes. Industry commentary on hybrid operations has started making exactly this point, and it’s why a ten-minute bot problem can become a forty-minute service-level problem.

The mix is harder. Automation takes the simple contacts first, so the contacts left for agents are longer and more complex, a mechanism we worked through in when AI deflects the easy contacts, your AHT goes up. Any plan that removes volume without raising handle time has double-counted the saving.

Patience is lower. An escalated customer has already invested effort and often arrives frustrated. The abandonment curve in your Erlang A or simulation model was probably fitted on customers who came straight to the queue, which makes it optimistic for this group. See abandonment and caller patience.

Plan the AI as a queue of its own

Give the bot the same treatment as any other resource. It has a demand profile (which contact reasons it receives, by interval), a throughput, a resolution rate, and a failure rate. Containment should be a range, not a number, because it varies by contact reason, time of day and release. And it needs leading indicators that tell you the human plan is about to be wrong: escalation rate by reason, repeat contact within seven days, and the share of “contained” conversations that end with the customer simply leaving. The last of these is the one that exposes false containment, covered in the false-containment trap.

Build the demand model bottom-up by contact reason (demand decomposition is the method), tag each reason as bot-suitable or not, and model deflection at that level rather than applying a flat percentage to the total.

Budget both workforces, and keep a fallback

A hybrid capacity plan has a human headcount line and an AI line. The AI line isn’t free: platform and usage costs, the people who maintain knowledge and prompts, quality review, and supervision all belong in it. Some WFM vendors have started building long-term capacity planning that budgets AI headcount alongside human headcount, which is a sign of where the discipline is going.

Then decide what you hold in reserve. The question isn’t “how many agents do we need if the bot works?” but “what is the worst credible outage, and how long would it take to cover it?” A trained flex pool, cross-skilled overtime agreements and a pre-agreed callback or message-back mechanism (moving the work, not removing it) are the practical answers. Test them, because a fallback you haven’t exercised is a hope.

Treat the big forecasts as scenarios

You will be shown projections such as Cisco’s forecast that agentic AI will handle 68% of contact centre interactions by 2028, or Gartner’s view that routine service resolution will approach 80% by 2029 (both as cited in trade coverage). They are useful for scenario planning and nothing more. Your own contact mix, regulatory exposure and customer base decide how much of that applies to you. Run a low, central and high deflection case, and plan the human side so that it survives all three.

What changes in the human queue when a bot sits in front The old assumptions ▸ Contacts arrive independently ▸ Mix is stable and mostly simple ▸ Patience matches the fitted curve ▸ Capacity is people, and they are available ▸ Volume saved equals headcount saved What you now plan for ▸ Escalation bursts when the bot stumbles ▸ Harder, longer contacts reaching agents ▸ Lower patience after bot effort ▸ AI capacity with its own failure rate ▸ Deflected workload, not deflected volume Plan the AI queue, then the human queue behind it.

The closing thought

Capacity is no longer a single number of heads. It’s a system with two workforces and a handover between them, and the handover is where plans fail. Part three looks at the vendors selling that system, and how to read their claims.

See also

Sources and further reading

External sources used in researching this article, accessed October 2026. Vendor-authored and trade-press material is identified as such; figures are quoted as reported and have not been independently verified.