To plan staffing for AI support at 100+ agents, forecast volume by channel and intent, decide which intents the AI agent will resolve end to end, then re-run Erlang C on the volume and handle time that is left for people, not on your old totals. Headcount falls by less than the automation rate, because the contacts that remain are harder and smaller queues need more agents per unit of work.

The rest is ordinary workforce management done more often: occupancy caps, shrinkage, escalation capacity for peaks, and a monthly re-forecast as the AI takes on more intents.

Key takeaways

• Never cut headcount by the automation rate. In our illustrative example below, the AI resolves 40% of calls but rostered headcount falls 27%, from 113 to 82, once the remaining calls take longer.

• Plan by intent slice. A blended "AI rate" hides the intents that still spike your human queue.

• Keep the classic guardrails. Call Centre Helper uses 85% maximum occupancy and 30% shrinkage as industry averages in its Erlang C example.

• Expect to keep people. Gartner found 95% of customer service leaders plan to retain human agents to define AI's role.

• Re-forecast monthly while automation is climbing, and weekly in your peak season.

How we chose and checked these steps

The eight steps below follow the standard planning cycle (forecast, size, schedule, review) and add the two places AI changes the maths: which volume leaves the human queue, and what the remaining contacts look like. Every external figure links to its source, which we read on 8 and 22 October 2026. The worked example uses illustrative inputs, not data from any company, and we show the arithmetic so you can repeat it with your own numbers.

Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

Staffing and volume planning for AI support: the steps at a glance

1. Baseline volume by channel, intent and interval

Start with what actually arrives: contacts per 30-minute interval, split by channel and by intent (refunds, identity checks, payment failures, complaints). Channel matters because the maths differs. Erlang C was built for one agent handling one call, and Call Centre Helper's guide to Erlang C for chat explains how to adjust it for agents handling several chats at once with a concurrency factor: in its table, an agent handling 2 chats counts as 1.7, and an agent handling 3 counts as 2.5.

Intent matters because AI will not take every intent at the same rate. Pull at least 12 months of tagged history if you have it, so seasonality shows up. If your tags are messy, fix the top 20 intents first; they usually carry most of the volume. Our workforce forecasting guide covers the forecasting methods in more depth.

2. Forecast peaks separately from averages

Average volume does not break a support floor; peaks do. Build separate curves for your peak days (tax deadlines, rent weeks, billing runs, product launches) and for the peak intervals inside them. Call Centre Helper notes that chat forecasting should use historical actual concurrency rather than the maximum an agent can handle, because concurrency drops when you are overstaffed and rises when you are short.

Peaks are also where AI changes the picture most. Flex handles 4x chat spikes during rent week, and Hnry's published story shows automation climbing from roughly 58% across the full period to around 70% in its peak week. If your AI agent absorbs a larger share at peak than on an average day, plan two residual curves, not one.

3. Size the human queue with Erlang C, or Erlang A

Erlang C turns calls per interval, average handle time (AHT) and a service level target into a raw agent count. Call Centre Helper's Erlang C worked example uses 100 calls per 30 minutes, a 3-minute AHT and 80% answered in 20 seconds: that is 10 Erlangs of traffic, and 14 agents reach 88.8% service level with a 7.8-second average speed of answer.

Know its limits. Mandelbaum and Zeltyn's Erlang-A paper calls Erlang-C the model most frequently used in workforce management of call centers, but notes it ignores abandonment, which can lead to under- or over-staffing. If many callers hang up while waiting, Erlang A or simulation is the better tool. For the method step by step, see our Erlang C staffing guide.

4. Add occupancy caps and shrinkage

Raw agents are not rostered agents. Two adjustments follow. First, cap occupancy: Call Centre Helper's example uses 85% maximum occupancy, and adds agents until the cap is met. Second, divide by one minus shrinkage. Call Centre Helper's shrinkage guide says most professionals agree shrinkage normally comes out between 30 and 35%, and cites a Dimension Data benchmark average of 35%.

So 14 raw agents at 30% shrinkage becomes 14 / 0.7 = 20 rostered. At 100+ agents the shrinkage line alone is worth a full team, which is why it deserves its own review. Our guide to adherence, shrinkage and occupancy covers how to measure each one.

5. Model the AI resolution rate by intent slice

Do not plan on a single blended automation rate. List your intents, and for each one estimate the share the AI agent will resolve end to end, from a pilot or a replay of historical tickets. Start conservative and update it from production data. Carmoola's published story is a useful shape to expect: about 30% resolved by its earlier automation, 40% end to end from day one with Lorikeet, and 60% of all inbound on WhatsApp, chat and email today.

Be clear about what "resolved" means. A contact the AI hands back to a person still costs an agent's time, often more than before, because the customer has already tried once. Count escalations as human volume in your plan.

The external direction of travel supports caution. Gartner also predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention. That is a forecast about common issues, not a planning input for your complaint queue.

6. Re-run Erlang C on the residual volume and handle time

This is the step most plans skip. When AI removes the simpler contacts, the ones left take longer. Re-run Erlang C with the residual volume and a higher AHT, then re-apply occupancy and shrinkage. The illustrative example below shows why the result is not a straight-line cut.

We computed these with the standard Erlang C formula, the same method that reproduces Call Centre Helper's 14-agent example exactly. The inputs are assumptions chosen to show the effect, not benchmarks. Two things drive the gap: longer residual handle time, and the loss of scale efficiency, since a smaller queue needs more agents per Erlang of traffic to hit the same service level.

7. Plan escalation and peak capacity for the harder mix

Your human team now handles the contacts the AI could not finish, plus the ones it should never touch. That mix needs more experienced people and a buffer for escalation spikes. Plan three things: an escalation queue with its own service level, a skills matrix for complaints, vulnerable customers and regulated actions, and a peak playbook (overtime, cross-skilled back-office staff, outsourced partners) that you trigger from the forecast, not on the day.

The evidence on keeping humans is consistent. Gartner predicted in June 2025 that by 2027, 50% of organisations that expected to significantly reduce their customer service workforce will abandon those plans, and a March 2025 poll of 163 leaders found 95% plan to retain human agents. Agents also get faster with AI help: an NBER study of 5,179 customer support agents found an AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers.

8. Redesign roles and set a review cadence

At 100+ agents, AI changes job design as much as headcount. Expect to need people who own AI workflows and knowledge, QA reviewers who score both human and AI conversations, and WFM analysts who forecast two queues instead of one. Some of these roles come from your strongest agents.

Then set the cadence. Re-forecast monthly while automation is still climbing, weekly in peak season, and after every material change to what the AI handles. Compare forecast against actual by intent, not just in total, so you can see which slice moved. Our contact center workforce management guide covers the full planning cycle.

Tools that help

Lorikeet's Workforce Manager details come from its launch post, which says Lorikeet customers can skip the roster upload because it reads ticket history directly. Lorikeet's honest limit for this use case: it is not a full enterprise WFM suite, and its launch post describes intraday management as the next step rather than a shipped feature. Large voice-heavy centres with deep skills-based routing will likely still want a dedicated WFM platform alongside it.

On cost, Lorikeet's published pricing sizes plans by monthly ticket volume: Start for under 5,000 tickets, Scale for 5,000 to 20,000, and Signature for 20,000 or more, with escalations to a person not charged. Signature is custom-priced and is the plan published for 20,000+ monthly tickets or dedicated support.

Questions to ask before you change headcount

• Which intents will the AI resolve end to end, and what share of each, based on replayed tickets rather than a demo?

• What is the AHT of the contacts that remain, measured after a pilot?

• How are escalations counted in the forecast, and do they carry their own service level?

• Does the AI share hold at peak, or fall away when volume spikes?

• Who owns AI workflows, knowledge and QA, and are those hours in the roster?

• How often will you re-forecast, and what triggers an off-cycle review?

What still needs a human

Complaints, vulnerable customers, regulated decisions and anything that needs judgement or discretion should stay with trained people, with the AI routing them fast. Headcount decisions themselves also need a human: model several scenarios, reduce through attrition and redeployment where you can, and keep a buffer until production data proves the AI share holds through a full peak season.

If you want to test the residual-volume maths on your own tickets, you can start a 30-day free trial of Lorikeet.