The short version of support staffing math is one row of arithmetic: tickets per period, times average handle time, divided by productive minutes per agent, plus a buffer for shrinkage. The number that falls out is the floor. Most teams get it wrong in one of two ways: they forget shrinkage, or they staff to the average instead of the peak. Below is the formula, the inputs that actually matter for a small team, and where AI changes the math in 2026.
What does the staffing formula actually look like?
For a chat or email team, the working formula is straightforward:
agents needed = (tickets per month × AHT in minutes)
÷ (productive minutes per agent per month)Productive minutes is paid minutes minus shrinkage. If an agent is paid for 160 hours a month and your shrinkage is 30%, productive time is 112 hours, or 6,720 minutes. If your team handles 5,000 tickets at an average handle time of 6 minutes, that is 30,000 ticket-minutes, divided by 6,720 productive minutes, equals 4.46. Round up. Five agents is the floor.
For phone or synchronous chat, this is the simplified version of Erlang C, the formula developed by Agner Erlang in 1917 and still the industry standard for inbound staffing. Erlang C adds queueing math: given a target service level (say, 80% answered in 20 seconds), it tells you the minimum agents needed in a 30-minute interval. The free calculators from Call Centre Helper or Assembled are accurate enough for most small teams.
What inputs actually matter?
Three. The rest is noise.
Average Handle Time (AHT). The 2025 cross-industry benchmark is around 6 minutes 10 seconds for general customer service, per Sprinklr's data. Industry varies. Retail and ecommerce push AHT under 2 minutes. Tech support runs 7 to 10. Pick your own from a real sample of recent tickets, not an industry average. Six weeks of data is enough.
Shrinkage. The percentage of paid time agents are not available to handle contacts: breaks, meetings, training, sick days, holiday, the time it takes to write a wrap-up note. Industry average is 30 to 35%, per Loris and Calabrio's 2025 reports. High-performing centres hit 20 to 25%. A small team without a dedicated workforce manager is probably at 35% or higher. Measure yours; do not assume.
Volume by interval, not by month. The monthly average lies. A team that does 5,000 tickets a month does not do 167 a day evenly. There is a Monday spike. There is a 10 am surge. Staff to the peak interval at the service level you want, or the peak gets dropped.
Why do small teams systematically understaff?
Because the founder picks the number that feels right and the spreadsheet never gets touched. We see three failure patterns repeatedly on teams of five to fifty.
First, they staff to the average instead of the peak. The math gives the right answer for an average Tuesday and the wrong answer for Monday morning. Second, they ignore shrinkage. Paid hours and productive hours are not the same; a 35% shrinkage rate turns a "we have five agents" team into a "we have 3.25 effective agents at any moment" team. Third, they wait until CSAT cracks before hiring, by which point the queue has burned out the team they already have.
The cost of getting this wrong shows up at the back door. Call centre agent turnover ran 30 to 45% annually in 2025 per industry research, with over 60% of leavers citing stress as the top reason. Replacing one agent costs $10,000 to $20,000 in recruiting, training, and ramp time. A perpetually-understaffed team is also a perpetually-rehiring team. The savings from holding one open seat usually disappear three months later when you backfill the leaver who burnt out.
How does AI change the staffing calculation?
Two ways, neither of them headcount-zero.
It compresses AHT on routine tickets. Brynjolfsson, Li and Raymond's 2025 paper in the Quarterly Journal of Economics, observing more than 5,000 customer support agents at a Fortune 500 firm, found AI assistance raised productivity by 15% on average and roughly 35% for the least experienced agents. Industry data shows AI-assisted contact centres cutting handle time by 9% on average. Plug those into the formula. Same volume, lower AHT, fewer agents needed for the routine slice.
It deflects volume from the queue. If your AI agent answers a real share of tickets without escalation, that volume comes out of the formula entirely. The honest version: count only the tickets where the AI resolved the customer's question, not the ones it touched. Salesforce's 2025 State of Service survey found service teams estimate 30% of cases are currently handled by AI, projected 50% by 2027. The realistic small-team result is meaningfully lower, somewhere between 20% and 50%, depending on product complexity and source content quality.
Where AI does not change the math: complex tickets, regulated tickets, retention conversations, and anything the customer specifically asked a human about. Gartner's February 2026 research found only 20% of organisations reported reduced agent headcount due to AI, and predicted half of those that planned cuts will reverse the plan by 2027. The honest pattern is that AI absorbs the routine volume and your team does harder work, not that headcount goes to zero. The US Bureau of Labor Statistics projects customer service rep employment to decline 5% from 2024 to 2034 but adds about 341,700 annual openings from replacement need, which is the more useful framing.
What does the math look like for a small team?
Worked example, ecommerce store, before and after AI.
Volume: 5,000 tickets per month. AHT: 5 minutes (ecommerce, mostly WISMO and order-status). Shrinkage: 35%. Paid minutes per agent: 9,600 (40 hours × 4 weeks × 60 minutes). Productive minutes per agent: 6,240 (after 35% shrinkage).
Before AI: (5,000 × 5) ÷ 6,240 = 4.0. Round up. Five agents to cover average load with no buffer for peaks.
After AI, 40% of volume resolved by AI, AHT on remaining tickets unchanged: (3,000 × 5) ÷ 6,240 = 2.4. Round up. Three agents.
That swing from five to three is meaningful for a small team's payroll. But notice what it actually says: AI took out two seats of routine work. The team that remains does the harder work. That is the typical 2026 small-team result.
How Keloa approaches support staffing
Keloa's AI agents sit in front of your inbox, answer the routine slice grounded in your own content, and escalate cleanly into the unified inbox when the question needs a human. That means the volume the formula has to staff for is the post-deflection volume, not the inbound volume.
Per-reply pricing makes the math transparent: you pay for what the AI handles, not per seat or per "resolution" with disputed definitions. See the per-resolution pricing trap for why that matters when you are doing this calculation. And our first response time benchmarks for SMBs covers the interval-level math for service levels.
Frequently asked questions
What is a good shrinkage rate for a small team? Industry average is 30 to 35%, high performers hit 20 to 25%. Most small teams without a dedicated workforce manager are at 35% or higher. Measure your own from a real two-week sample before you assume. Halving shrinkage usually requires schedule discipline more than tooling.
Should we use Erlang C for a five-person email team? No. Erlang C is for synchronous channels where queueing matters. For email and async chat, the simpler formula (tickets × AHT ÷ productive minutes) is accurate enough. Use Erlang C the day you add live phone or live chat at volume.
How do I know if I am under-staffed right now? Three signals. First response time creeping up week over week. CSAT dropping on the channels where queue length is highest. Agents working through breaks. Any two of those three for two weeks running, you are short at least one seat.
Does AI actually let us cut headcount? Sometimes, but not usually. Gartner found only 20% of organisations reduced headcount due to AI in 2026, and half of those that planned cuts will reverse them by 2027. The honest pattern is that AI absorbs routine volume and the existing team does harder work without growing as fast as the business does.
What about peaks like Black Friday? Plan to the peak interval, not the average month. Add buffer for shrinkage spikes (sick days run higher near holidays). And lean on AI for the surge, because hiring contractors for a four-week peak is expensive and the ramp eats the gain. See our piece on peak season support without hiring.
Is "tickets per agent per day" a useful metric? Directionally, not for staffing decisions. The 2017 industry baseline of 21 tickets per agent per day still gets cited, and many teams using AI assistance push higher. Use it to compare your team to itself over time, not to a benchmark.