Frequently asked questions
The questions workforce planners and contact-centre managers ask most often. Short, opinionated answers — with a link to the deeper article on each.
No fluff, no consultancy-speak. Where the honest answer is “it depends”, we say so — and tell you what it depends on.
-
What service level should I set?
There’s no industry-standard answer — but there is an industry-standard mistake, which is to inherit 80/20 from a previous operation without checking whether it fits yours. The right SL target is the one that matches what your customers will tolerate, what your unit economics will support, and what your operation can sustainably deliver. For most retail and service operations 80% in 20 seconds is reasonable; regulated and complaint-handling queues often need higher; back-office and lower-urgency queues often justify lower.
The bigger question is whether SL is still the right primary metric at all. Some operations are moving toward outcome metrics — first-contact resolution, abandonment, segment-level service — and using SL as a secondary check rather than the headline.
Read: how to set a service level target →
Read: is service level a dead KPI? → -
What’s a normal shrinkage rate?
Including paid leave, holidays, training, coaching, breaks, meetings, and unplanned absence — a healthy operation runs at 28%–35%. Operations above 40% are usually carrying hidden time the agents themselves don’t realise they’re consuming. Operations below 25% are usually under-investing in training or coaching and will pay for it later in quality and attrition.
The most useful exercise is to break the number down. The single composite shrinkage figure obscures whether the problem is structural (holiday, training) or behavioural (adherence, ACW). Each needs a different response.
-
What’s a normal attrition rate for a contact centre?
The honest answer is that the headline number is the wrong one. Year-one attrition is typically 35%–55% in retail and service operations; tenured attrition is 8%–15%. The composite hides the actual problem. The operations that meaningfully reduce attrition do it by segment, almost always starting with year-one.
The follow-on question, usually, is what the attrition actually costs. The number most operations carry on the slide (£1,500–£2,000 per leaver) captures about a quarter of the real cost — the all-in figure including vacancy, ramp-up, manager time and quality drift is closer to £6,000 per leaver.
Read: what’s a normal attrition rate? →
Read: the true cost of attrition → -
Is Erlang C still the right tool?
For most voice-only operations, yes. Erlang C is still the right tool when arrivals are roughly Poisson, the queue is single-skill, you have at least 15-minute intervals, and abandonment isn’t a meaningful share of contacts. When abandonment is material, reach for Erlang A or Erlang X instead — they model patience, busy-tone blocking and redials, and usually need slightly fewer agents.
It’s the wrong tool for chat and async work (different concurrency physics), for skills-routing problems (Erlang C assumes everyone can take everything), for bursty arrivals (the Poisson assumption breaks), and for outbound (genuinely different model). When in doubt, simulation is the answer.
Read: what is Erlang? (plain English) →
Use: the Erlang C calculator →
Use: the Erlang X calculator (abandonment + redials) →
Read: abandonment and caller patience →
Read: chat staffing isn’t Erlang → -
How accurate should my forecast be?
More useful than the accuracy target is the question “against what?” A weekly forecast at total-operation level should hit 3%–5% MAPE on a stable operation. A daily interval-level forecast is closer to 8%–12%. A 30-minute interval forecast on a small queue can be 15%–25% and still be doing a useful job — because the noise floor at that granularity is high.
The mistake most operations make is benchmarking forecast accuracy against an aspiration instead of against a naive baseline. If your fancy model isn’t beating “same day last week” by a meaningful margin, the sophistication isn’t adding value.
Read: forecast accuracy metrics →
Read: your forecast is probably more accurate than you think →
Read: Poisson and natural noise → -
How much will AI improve my forecast?
Less than vendors claim, more than sceptics expect. A well-tuned ML model with external drivers typically delivers a 15%–25% relative MAPE improvement over a strong statistical baseline. Foundation models add a small further increment. The bigger gains usually come from getting the data clean, not from a better algorithm.
The honest order of operations is: fix the data, beat a moving-average baseline with a Holt-Winters or Prophet model, layer in driver-based regression for the variables that genuinely matter, then consider AI/ML. Operations that start at the AI step and skip the foundations usually find the model is no better than what they had.
Read: AI for forecasting — what works and what doesn’t →
Read: gen-AI for planners (the everyday productivity use) → -
Will a chatbot or AI let me cut headcount?
Less than the business case says — because of a second-order effect almost nobody models. Automation deflects the simple, short contacts first, so the ones left for your agents are harder and longer. Your average handle time goes up even as volume falls, and the workload drops by less than the headline deflection rate suggests. A 30% deflection might cut workload by only 15%.
The fix is to plan on deflected workload (volume × AHT), not deflected volume, and to rebuild AHT for the residual mix rather than reusing your old number — the moment a bot is in front of the queue, your historical AHT is obsolete.
Read: when AI deflects the easy contacts, your AHT goes up →
-
Does adding agents always improve service level?
Yes, but not linearly. The curve is heavily non-linear: at low staffing, one more agent transforms service; at high staffing, one more agent barely moves it. The most important intuition for a finance conversation is that 14-to-15 agents matters more than 24-to-25 — and that intuition makes the headcount conversation tractable.
This is also why “just throw bodies at it” is a poor strategy past a certain point. Once you’re past the steep part of the curve, scheduling shape, AHT, and shrinkage move SL more than headcount does.
Read: what is Erlang? (the S-curve explained) →
Try: change agents by 1 in the Erlang C calculator →
Read: the power of one → -
Should breaks be fixed or flexible?
Fixed breaks deliver more predictable intraday coverage; flex breaks deliver better workforce experience and lower attrition. The honest answer is a hybrid — fix breaks tightly at the operation’s known coverage troughs (lunch peak, mid-afternoon dip) and flex them elsewhere.
Operations that pick one of the two pure models leave value on the table. The schedule-design question more broadly is one of the cheapest service-level levers most planners under-invest in.
-
What should I do when finance gives me a headcount that’s too low?
Operate honestly inside the cap, but document the gap quarterly. Five levers genuinely move outcomes inside a constrained headcount: shrinkage, schedule shape, AHT, channel/demand-shaping, and targeted SL compromise by segment. Avoid the four traps that look like solutions but aren’t — permanent overtime, the unstaffed late shift, hiding in averages, and quiet quality drift.
The political move is to present the gap as a trade-off rather than a deficit. “At this FTE, here’s the SL we’ll deliver; at the modelled FTE, here’s what we’d deliver. The difference costs £X in payroll and delivers £Y in retained customer value.” That framing wins more conversations than arguing the maths.
-
What does a planning team look like in a small contact centre?
Below about 60 agents the planning function is usually one person doing forecasting, scheduling, and real-time across the day. 60–150 agents typically supports a 2-person team split between long-range (forecasting/scheduling) and real-time. 150+ supports a proper specialist function with separate leads for forecasting, scheduling, and real-time, plus an MI/insight role.
The structure matters less than the discipline. Even a one-person team can produce serious work if the process is right; a five-person team can produce theatre if it isn’t.
Read: the planning team in a small contact centre →
Read: the minimum viable planning team → -
What’s the difference between adherence and conformance?
Adherence asks whether an agent was where the schedule said they should be at every interval — a timing measure. Conformance asks whether they worked the total hours their schedule required, regardless of timing. An agent can hit 100% conformance with poor adherence by working all their hours but shifting them around.
For voice operations with real-time service-level targets, adherence is the metric that matters — because the schedule is engineered for half-hour coverage. For back-office and async work where the target is turnaround rather than wait time, conformance is often the right one. Many operations report both and over-act on adherence in contexts where conformance would be the better signal.
-
What’s the right occupancy target?
For a sustainable voice operation, 80%–85% occupancy is the upper end. Above 88% you start seeing measurable attrition impact within a quarter; above 90% the link to sickness and disengagement is well-established. Below 70% suggests over-staffing or schedule shape that doesn’t match demand.
Occupancy is one of the most under-watched leading indicators in contact centres — most operations only notice it via attrition six months later. The discipline is to track it weekly, by queue, and to treat sustained spikes as a planning problem (schedule shape, AHT, demand mix) before it becomes a people problem.
-
How should I think about vulnerable customers in planning?
Vulnerable customers need a distinct planning track, not a sub-category of main-queue planning. Three disciplines: a named segment in your forecast (don’t bundle into general volume), a separate service-level target (typically higher than the main queue), and a routing rule that gets them to trained agents quickly.
Under UK Consumer Duty, financial services firms must evidence outcomes for vulnerable customers specifically — meaning segment-level SL is no longer optional. The composite headline SL hides exactly the disparity the regulator wants to see.
Read: Consumer Duty — evidencing outcomes for vulnerable customers →
-
What’s the right span of control for a team leader?
Most CC operations carry 12–15 agents per TL, but that number tells you almost nothing. The right span is whatever divides the TL’s coaching hours among their team so each agent gets the minimum coaching that produces improvement — usually about 30 minutes of one-to-one per agent per fortnight.
Working backwards from coaching hours rather than inheriting an org chart usually produces a span of 8–12 for new agents and 14–18 for tenured. Operations that widen the span to save TL cost typically pay it back in attrition and quality drift within two quarters.
-
How do I plan back-office work?
Back-office work is a queue, not a roster. It has its own arrival pattern, its own AHT, its own service-level target (often a turnaround time rather than a wait time), and its own staffing model — typically not Erlang.
The two biggest mistakes are treating it as overflow capacity for voice (it isn’t — the skills are different and the cognitive switching cost is real) and assuming the backlog acts like a live queue (it doesn’t — backlog work can be paused and resumed, so the staffing model is closer to a workshop than a switchboard).
Read: planning email and processing work →
Read: back-office and blended work → -
What’s a sensible service level for chat or digital channels?
Don’t import voice SL targets into chat — the physics are different. Chat is planned on concurrency and response time, not 80/20. A reasonable starting point is 90% of chats answered within 30 seconds, with a maximum concurrency of 3 active sessions per agent for transactional work and 2 for complex.
Async messaging (WhatsApp, in-app) needs a turnaround SLA — typically “first response within an hour” rather than a queue-wait SL — because open conversations can stay idle for hours without being abandoned in the voice sense.
Read: chat staffing isn’t Erlang →
Read: planning async messaging → -
What KPIs should I report to my exec team?
Five metrics, no more: service level (or a customer-outcome metric like FCR), forecast accuracy, shrinkage, attrition, and a leading indicator of customer experience (CSAT, complaint volume, or theme analysis from speech analytics).
Most execs glaze over past five metrics. The hardest discipline is removing the ones you’ve always reported because someone once asked for them. Every metric on the slide should drive a different executive action; if two metrics drive the same response, one of them is decoration.
Read: all lagging, no leading — the leading indicators worth adding →
-
How do I plan for AI / generative AI in the contact centre?
Plan for it as a capability with its own operating model — not as a magical headcount lever. The disciplined sequence: name the operational decision the AI supports, choose a calibrated capability (not the shiniest), pilot with human-in-the-loop, measure both intended and unintended effects (AHT, abandon rate, complaint volume), then scale only if the evidence holds.
Most planning failures come from skipping the operating-model question — who owns it, who monitors it, what triggers a rollback. AI without an operating model isn’t an AI failure, it’s a planning failure.
Read: AI in the contact centre — the 2026 landscape →
Read: the operating model your AI capability deserves →
Read: the eight-step AI implementation framework → -
How do I plan for peak season?
Start in March, not October. Peak planning has four phases — demand modelling (March–May), capacity planning and recruitment lead time (May–August), schedule design and training (August–October), and execution with daily replanning (October–January).
The single biggest mistake is treating peak as a scale-up of normal operations. The demand pattern, AHT, channel mix and attrition profile all shift at peak, and a roster scaled from October patterns will misfire in December. The disciplined planner rebuilds the forecast for peak from peak data, not from a multiplier on the rest of the year.
Read: retail CC planning — peak as the year’s organising event →
Question we haven’t answered?
Suggest one in the discussion on LinkedIn or message ccPlanning on LinkedIn. We’ll rotate the list as new questions come in and as new articles publish.
Page reviewed regularly as the article library grows. Last reviewed: May 2026.