Cost per Call: A Simple Model for Pricing Call Automation

"How much does a call cost us?" is a question many businesses can't answer, and it's the starting point for deciding whether automation makes sense. This article builds a simple model you can fill in with your own numbers. Every figure below is an illustration, not a benchmark.

The basic formula

Cost per call = total cost of handling calls in a period / number of calls handled in that period

The work is in getting "total cost" right.

What goes into the cost

Labor (usually the biggest part)

Don't use hourly wage alone. Use fully loaded cost:

• Wages
• Payroll taxes and benefits
• Paid time off and holidays
• Training and onboarding time
• Supervisors, QA and workforce management, spread across agents
• Turnover cost: recruiting and ramp time for replacements

Then account for utilization. Agents don't spend every paid minute on calls. Breaks, meetings, training and idle time between calls all count. If an agent is paid for 8 hours and spends 5.5 handling calls, every call carries a share of the other 2.5.

Technology

Phone system, dialer, CRM seats, recording and QA software, headsets, and any other per-agent tools.

Facilities

For in-house teams: desk space, utilities, equipment. For remote teams: stipends and equipment.

Telecom

Per-minute charges, numbers and trunks.

A worked example

Suppose a team of 10 agents, each with a fully loaded cost of $5,000 per month (an illustration; use your own figures).

• Labor: 10 × $5,000 = $50,000
• Technology, telecom and facilities: $8,000
• Total monthly cost: $58,000

The team handles 12,000 calls a month.

Cost per call = $58,000 / 12,000 = $4.83

You can also express it per minute. If average handle time (see average handle time) is 6 minutes, that's 72,000 handled minutes, or about $0.81 per handled minute.

Cost per call vs cost per resolved call

Cost per call treats a call that resolves the issue the same as one that causes a callback. If 25% of issues need a second call, the true cost of resolving an issue is higher than cost per call suggests.

Cost per resolution = total cost / issues resolved

This is the better number for comparing options, because automation that's cheap per call but resolves less can cost more overall. See first call resolution for measuring the resolution part.

Modeling automation

The mistake most business cases make is applying an overall average to the calls being automated. Those calls are usually the simplest and cheapest. Build the model by call type instead.

Step 1: Pick the call types to automate

From your call reasons, choose the types you'd hand to an AI agent first, say order status and appointment changes. Note their monthly volume and their handle time. Simple calls often have well below average handle time.

Step 2: Estimate the agent's containment

Containment is the share of those calls the agent resolves without a person. Be conservative at first, and validate it with a pilot. Don't assume 100%.

Step 3: Calculate both sides

For the automated call types:

• Current cost: volume × handle time × cost per handled minute
• New cost: (AI agent cost for that volume) + (calls not contained × handle time × cost per handled minute) + (review and maintenance time)

Remember that calls handed off by the agent often take a little less human time than before, because the person receives a summary. Don't count on that until you've measured it.

Step 4: Add one-time costs

Setup, integration work, testing, and your team's time during the pilot.

Step 5: Decide what you'll do with freed capacity

Savings only appear on the books if you change something: reduce overtime, slow hiring, stop using an overflow vendor, or move people to higher-value work like sales or retention calls. Freed capacity that nobody plans for turns into slightly shorter queues, which is nice but different from a cost saving.

Things that make the model wrong

• Using wage instead of fully loaded cost. Understates the cost of people.
• Using average handle time for simple calls. Overstates savings.
• Ignoring calls that still reach people. Overstates savings.
• Ignoring review and maintenance time. Someone needs to read transcripts and update the agent.
• Ignoring revenue effects. Answering more calls after hours, or calling leads faster, can bring revenue that a pure cost model misses.
• Ignoring quality. A cheaper call that annoys customers has costs that don't show up in this spreadsheet.

A simple template

A worked automation scenario

Let's put the steps above together with illustrative numbers. Suppose:

• Fully loaded cost per handled minute for staff: $0.80
• Order status calls: 4,000 per month, average 3.5 minutes each
• Current monthly cost of those calls: 4,000 × 3.5 × $0.80 = $11,200

A pilot shows an AI agent resolves 70% of order status calls without a person. The remaining 30% are transferred with a summary, and those calls take people an average of 3 minutes instead of 3.5.

• Calls handled by people after launch: 1,200 × 3 minutes × $0.80 = $2,880
• AI agent cost for 4,000 calls: suppose $0.25 per minute at an average of 2 minutes per call = $2,000
• Review and maintenance: 6 hours a month at $50 per hour = $300
• New monthly cost: $2,880 + $2,000 + $300 = $5,180

Monthly saving on this call type: about $6,000, before one-time setup costs. Change any assumption and the result changes, which is the point: the model shows you which assumptions matter most.

Sensitivity: which numbers move the result

In most automation models, three inputs dominate:

1. Containment rate. The share of calls the agent fully resolves. Every point matters, so validate it with a real pilot.
2. Handle time of the automated calls. Simple calls are usually short, which caps the savings per call.
3. Volume. Fixed costs like setup and maintenance are spread over more calls as volume grows.

Run your model with low, expected and high values for each. If the business case only works with the most optimistic containment figure, start smaller or pick a different call type.

Don't forget revenue

Cost models capture savings on calls you already answer. They miss the value of calls you currently lose. For many businesses, especially small ones, the bigger number is revenue from:

• After-hours calls that used to go to voicemail
• Peak-time calls that abandoned in the queue
• Web leads that got called back hours late

If you can estimate how many of those calls you lose and what an average customer is worth, add a revenue line to the model. Keep it conservative and separate from the cost line, so the case doesn't depend on it.

Presenting the model

When you take a model to leadership, show:

• The assumptions, clearly labeled, with sources (pilot data, system reports, estimates)
• Low, expected and high scenarios
• One-time costs and the payback period
• What will happen to the freed capacity
• How you'll measure results after launch

A transparent model with modest assumptions earns more trust than an aggressive one, and it's easier to defend when actual results come in.

Frequently asked questions

What's a typical cost per call?

It varies widely by industry, location and call complexity. Your own number is the only one that matters for decisions.

Should I include calls in the IVR?

If you're comparing options for the same calls, yes. Include whatever system handles them today.

How long should a pilot run?

Long enough to see normal variation, often four to eight weeks, on a defined slice of calls.

Written by the Voxvencer editorial team. We build and run AI voice agents for call centers and small businesses, and we write about what we see on real phone lines. Questions or corrections: info@voxvencer.com.