How CFOs Work Out the Payback Period on an AI Customer Support Platform (2026)

A CFO works out the payback period on an AI customer support platform by finding the first month in which cumulative net savings cover the one-time cost of going live. Net monthly saving is tickets x resolution rate x (avoidable cost of a human ticket minus the price of an AI resolution), minus fixed platform fees. On the illustrative inputs below, a team handling 15,000 tickets a month at $9.61 per ticket, with the AI resolving 50% once ramped, recovers $55,100 of one-time cost in the third live month.

Two inputs decide almost every answer: the share of tickets the AI genuinely resolves, and the share of human cost you actually stop paying. This guide gives you the formula, a month-by-month worked example with every input labelled, a pricing model conversion, and a sensitivity table, so you can rerun it on your own numbers. It goes one level deeper on payback than our CFO framework for the total cost of AI customer service, which covers the full cost stack.

Key takeaways

• Payback month = the first month where cumulative (resolved tickets x margin per resolution, minus fixed fees) is greater than one-time cost. Count months from go-live, and put any pre-live fees into the one-time cost.

• Use resolution rate, not deflection. A ticket that leaves the AI and comes back through another channel saves nothing, so it should not count in the model.

• Apply a realisation factor to cost per ticket. In the example, only 70% of the $9.61 human cost is treated as avoidable. At 50% realisation, payback slips from month 3 to month 4.

• Convert every pricing model to effective cost per resolution. A $2 per-conversation price is $4 per resolution if half of billed conversations resolve.

• Decision rule: approve only if payback lands inside the contract term at your conservative resolution rate. In the example, even a 25% steady rate pays back in month 5.

Which AI customer support platforms make payback easy to model?

The platforms that are easiest to model price on outcomes, because cost moves with the number of tickets actually resolved. Lorikeet charges a plan fee plus a rate per resolved ticket, $0.90 per chat, email or SMS resolution on its Scale plan, and does not charge for unresolved tickets or escalations to a person. Alternatives worth comparing:

• Fin (now part of Salesforce) charges $0.99 per outcome, where billable outcomes include resolutions and procedure handoffs, with a 50-outcome monthly minimum.

• Zendesk AI agents are included in every Suite and Support plan and billed per automated resolution, on top of seat-based helpdesk pricing.

• Salesforce Agentforce lists $2 per conversation, or Flex Credits at $500 per 100,000 credits, alongside per-user licensing.

How do CFOs work out the payback period on an AI customer support platform?

They build a monthly cash model with five inputs and run it until cumulative savings cross zero. The formula is:

Monthly net saving = V x R(month) x (C x A minus P) minus F

Payback month = the first month where the sum of monthly net savings is at least I

1. V, monthly ticket volume in the channels the AI will work (chat, email, SMS, voice priced separately).

2. R, resolution rate by month: the share of all tickets in scope closed by the AI with no human touch and no repeat contact. It ramps, so model it month by month.

3. C x A, avoidable cost per ticket: fully loaded human cost per ticket (C) times the share of it you will actually stop paying (A, the realisation factor).

4. P and F, the vendor's price: the per-unit price converted to a per-resolution figure (P), and fixed monthly fees such as platform or seat charges (F).

5. I, one-time cost: integration, knowledge clean-up, testing, training and any fees paid before the first live ticket.

Keep the model monthly rather than annual. An annual average hides the ramp, and the ramp is where most of the payback risk sits.

What is the fully loaded cost per ticket, and how much can you actually save?

Fully loaded cost per ticket is everything you spend to run support in a month divided by the tickets handled that month, and you will usually save only part of it. For reference, HDI reported an average cost of about $22 to resolve a level 1 ticket at North American IT service desks. IT desks and customer support teams differ in ticket mix and pay, which is why you should build the number from your own payroll.

A defensible build for a US team:

• Wage: the BLS median wage for customer service representatives was $21.53 an hour in May 2025.

• Benefits: in private industry, benefits were 30.0% of employer compensation costs in June 2026, so applying that share, $21.53 becomes about $30.76 an hour loaded.

• Overhead: add team leads, QA, workforce planning, software seats and facilities. The example uses an illustrative 25%, giving $38.45 an hour.

• Throughput: divide by tickets handled per paid hour, after breaks, training and idle time. At an illustrative 4 tickets an hour, cost per ticket is $9.61.

Then apply the realisation factor. Savings become cash only when you remove cost: hires you no longer make, overtime and contractor hours you stop buying, BPO volume you cancel, or attrition you do not backfill. Team leads, the tooling contract and the floor do not shrink when 50% of tickets go to an AI. The example treats 70% of cost per ticket as avoidable ($6.73). If your headcount is fixed for the next year, the honest factor may be much lower, and payback should be calculated on avoided hiring instead. Our guide to cost per ticket benchmarks breaks the baseline down further.

Why does resolution rate, not deflection, drive the payback month?

Resolution rate drives payback because only a ticket that is closed for good removes human work, while a deflected ticket can come straight back. A customer who leaves a chatbot without an answer and then phones in costs you the full human ticket plus the AI attempt. Count a ticket as resolved only if the customer did not contact you again on the same issue within a set window, such as 7 days. The difference is explained in resolution rate vs deflection rate.

Model the ramp, not the end state. Gartner predicts that agentic AI will autonomously resolve 80% of common customer service issues by 2029, leading to a 30% reduction in operational costs. That is a forecast for common issues three years out, not a month-one assumption. Real ramps start lower and climb as workflows are added. Carmoola, a UK car finance lender, had existing automation resolving about 30% of inbound questions; Lorikeet resolved 40% of inbound conversations end to end from day one, and that figure now sits at 60% across WhatsApp, chat and email.

How do pricing models change the payback math?

Pricing models change payback by changing what you pay for tickets the AI does not resolve. Convert each one to effective cost per resolution with the same formula: (fixed monthly fees + unit price x units billed) / tickets resolved. The table assumes the AI handles 15,000 conversations and resolves 7,500 of them (50%), for illustration only.

Two things to check in any contract. First, who defines a resolution and whether you can dispute one. Second, what sits outside the unit price: platform fees, minimums, seat licences and paid onboarding all belong in F or I. On a per-resolution plan with a platform fee, the fee is the fixed cost to cover. In the example below, $5,100 a month is covered once about 6% of tickets resolve. The per-resolution pricing explainer covers the variants.

Worked example: payback for a 15,000-ticket team (illustrative inputs)

Every input below is illustrative, chosen to show the method. None of them is a Lorikeet customer result or a forecast for your team. Prices are from Lorikeet's public pricing page.

Margin per resolution is $6.73 minus $0.90, or $5.83.

Payback lands in live month 3, or month 4 from signature if the first month is spent getting to the first live ticket. Net saving over the first 12 live months, after one-time cost, is $355,930. Escalated tickets carry no AI charge, so the human-handled half of volume adds nothing to the vendor line.

How sensitive is payback to resolution rate?

Payback is far more sensitive to resolution rate and realisation than to the unit price. Each row keeps the same ramp shape (40%, 60%, 80%, then 100% of the steady rate) and every other input from the example.

Halving the realisation factor costs more than dropping from 50% to 35% resolution. Pressure-test the cost side as hard as the vendor's resolution claims.

What hidden costs push payback out?

The costs that move payback are the ones that never appear on a vendor quote. Add each to I or F, or reduce A:

• Knowledge and workflow work. Writing down policies that live in senior agents' heads, cleaning help centre articles and mapping each workflow to the API it needs.

• Integration time. Engineering hours to expose order, account or billing actions the AI can call. Lorikeet's pricing page puts average time to first live ticket at under 29 days; budget your own team's hours either way.

• Quality review. Someone has to read AI conversations, especially early. Automated QA is priced separately on most platforms ($0.27 per ticket on Lorikeet Scale).

• Harder remaining tickets. The tickets humans keep are the complex ones, so handle time per human ticket rises. In the example, a 15% cost premium on the 7,500 human tickets is about $10,800 a month.

• Repeat contacts. Any resolution that reopens must come out of R, or the model overstates savings.

• Customer demand for humans. Gartner predicts that by 2028, regulatory changes related to AI will increase assisted service volume by 30% as customers ask for a person.

• Future unit prices. The same Gartner release predicts generative AI cost per resolution will exceed $3 by 2030. Lock rates for the contract term and rerun the model at renewal.

What does AI customer support ROI look like for high-volume support teams?

For high-volume support teams, AI customer support ROI improves with scale because fixed fees spread over more resolutions while the margin per resolution stays roughly constant. At 15,000 tickets, the $5,100 platform fee adds $0.68 per resolution at a 50% rate; at 5,000 tickets it would add about $2.04. Above 20,000 tickets a month, Lorikeet's Signature plan moves to custom per-resolution rates with a $75K platform fee, so rerun the conversion table on the quote you receive.

The return is not only fewer tickets for humans. Research on 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. That gain belongs on the human side of the model: it lowers C for the tickets people keep. To plan headcount rather than payback, see how to scale customer support without hiring, and for the operating view of the same numbers, our sibling guide on lowering cost per ticket with AI.

What still needs a human

A payback model is a finance tool, and it leaves several decisions with people:

• Which tickets are in scope. Complaints, vulnerable customers, hardship cases and anything a regulator treats as advice should route to people by rule, and those tickets belong outside R.

• Escalation design. The AI should hand over with full context when it cannot finish a ticket. Fast, clean escalation protects CSAT, which a cost model does not capture.

• Compliance sign-off. In regulated industries, a compliance owner should approve workflows before go-live and review samples after.

• Headcount decisions. Whether savings come from attrition, redeployment or hiring freezes is a leadership call, and it sets A.

• The 90-day check. Compare actual resolution, repeat-contact rate and cost per ticket with the conservative row of your sensitivity table, and decide to continue, expand or stop.

The quickest way to replace illustrative inputs with real ones is to measure resolution on your own tickets. You can start a 30-day free trial with Lorikeet and run the model on what it actually resolves.