The fastest way to lose money on AI customer support is to buy it on deflection rate. The fastest way to make money on it is to do the per-resolution math against your fully loaded human cost and then measure the outcomes that actually move your P&L.
AI customer support ROI is the net financial return a high-volume team earns when an AI agent resolves tickets that would otherwise consume human agent time, measured against what the AI costs to run. In 2026, the calculation is simpler than it used to be because pricing has moved from per-seat to per-outcome. A human-handled ticket costs roughly $1.25 to $4.00 in fully loaded agent time at high volume. A Lorikeet AI resolution costs about $0.80 on chat, email, and SMS, and about $1.00 on voice, with escalations not charged. The gap between those two numbers, multiplied across hundreds of thousands of tickets, is the return.
The human baseline for a high-volume team is roughly $1.25 to $4.00 per ticket once you load in wages, benefits, tooling, QA, management, and attrition.
Lorikeet prices per resolution: about $0.80 on chat, email, and SMS, about $1.00 on voice, with Coach QA at about $0.10 per ticket and escalations not charged.
The customer defines what counts as a resolution, which removes the incentive a vendor has to inflate the metric you pay on.
Deflection rate is the most common vanity metric in this category. A ticket can be deflected without being resolved, and a deflected-but-unresolved ticket usually comes back more expensive.
The numbers worth tracking are cost per resolution, true resolution rate, contact rate, CSAT on AI-handled tickets, and reopen rate, not raw deflection.
Last updated: June 2026
High-volume support teams sit on a cost structure that does not scale gracefully. Every new tier of volume needs more agents, more managers, more QA, and more recruiting to backfill attrition, and each of those layers adds cost without adding much leverage. AI changes the unit economics, but only if you evaluate it correctly. Most teams that are disappointed by AI support bought it on a deflection number a vendor controlled and never built the per-resolution model that would have told them whether the thing was actually paying for itself. This guide walks through the cost math, a worked example for a high-volume team using clearly hypothetical numbers, the drivers that determine payback speed, what to measure, and the pitfalls that make ROI look better on a slide than it does in your finance system.
The Cost Math: Human Ticket Cost vs AI Resolution Cost
ROI is a comparison between two unit costs. To get it right you have to load both sides honestly, because the headline numbers on each side are misleading on their own.
Fully loaded cost per ticket: the total cost of resolving one ticket with a human agent, including wages, benefits, software seats, QA, management overhead, recruiting, and the productivity lost to attrition and ramp time, divided by tickets resolved.
Cost per resolution: what you pay an AI vendor for one ticket the AI actually resolved, as opposed to one it merely touched or deflected.
What a human-handled ticket really costs
The wage is the smallest part of the number. A support agent earning a modest salary looks cheap until you load in benefits and payroll taxes, the helpdesk and telephony seats they need, the QA function that samples their work, the team leads and managers who supervise them, the recruiters who replace them when they leave, and the weeks of reduced output while a new hire ramps. For a high-volume consumer team, those layers push the fully loaded cost into roughly the $1.25 to $4.00 per ticket range, with simple chat tickets at the low end and complex voice or regulated cases at the high end. The range is wide because it depends on geography, channel mix, and how complex your tickets are, so the right move is to compute your own number rather than borrow one.
The number also moves the wrong way as you grow. Attrition in high-volume support is chronic, every departure carries a recruiting and ramp cost, and the QA and management layers grow roughly in step with headcount. Volume spikes, like a product launch or an outage, force you to either overstaff for the peak or accept a backlog and a CSAT hit during the surge.
What an AI resolution costs with outcome-based pricing
Lorikeet prices on resolutions rather than seats: about $0.80 per resolution on chat, email, and SMS, and about $1.00 per resolution on voice. Coach, the analytics and automated QA agent, runs at about $0.10 per ticket and can be deployed on its own. Two design choices in that model matter more than the headline rate. First, the customer defines what counts as a resolution, so the metric you are billed on is not one the vendor can quietly inflate. Second, escalations are not charged, so when the AI correctly hands a hard case to a human you do not pay for a resolution that did not happen.
As a concrete anchor, Lorikeet's Scale plan is 48,000 resolutions for $48,000 per year, which works out to about $1.00 per resolution blended across channels. Compare that to the same 48,000 tickets handled by humans at, say, $2.50 fully loaded, which is $120,000, and the structural gap is visible before you model anything else.
Why outcome pricing changes the ROI calculation
Per-seat pricing rewards a vendor for the size of your team. Per-resolution pricing ties the vendor's revenue to work that actually got done. The distinction matters for ROI because it removes a layer of guesswork: you are not amortizing a platform fee across an uncertain volume, you are paying a known amount per outcome. The one thing to verify in any outcome-priced contract is the definition of the outcome. If the vendor defines resolution, the number you pay on is theirs to optimize. When the customer holds the veto on what counts, the pricing model and your interests point the same direction.
A Worked ROI Example for a High-Volume Team
The numbers below are illustrative and hypothetical, chosen to show the shape of the calculation rather than to describe any specific customer. Plug in your own figures before making a decision.
Picture a high-volume consumer team handling 50,000 inbound tickets a month, or 600,000 a year, across chat, email, and voice. Assume a fully loaded human cost of $2.50 per ticket, which sits in the middle of the typical range. That is a $1,500,000 monthly support cost, or $18,000,000 a year, before any AI.
Step 1: establish the human baseline
At 50,000 tickets a month and $2.50 fully loaded per ticket, the all-human baseline is $125,000 a month. This is the number AI has to beat, and it is the number too many ROI models skip in favor of a wage figure that understates the real cost by half.
Step 2: apply a realistic resolution rate
Assume the AI resolves 65% of inbound volume end-to-end, a deliberately conservative figure for a mature deployment, and escalates the remaining 35% to human agents. That is 32,500 AI-resolved tickets and 17,500 human-handled tickets a month. Higher automation rates are achievable, and some regulated deployments report figures in the mid-80s, but modeling on a conservative rate keeps the business case honest.
Step 3: cost out the AI-resolved tickets
At a blended $0.90 per resolution across the channel mix, 32,500 AI resolutions cost $29,250 a month. Add Coach QA at $0.10 per ticket across all 50,000 tickets, which is $5,000, and the AI layer costs about $34,250 a month. Escalations are not charged, so the 17,500 handed to humans add nothing to the AI bill.
Step 4: cost out the remaining human tickets
The 17,500 escalated tickets still cost $2.50 each fully loaded, which is $43,750 a month. These are, by design, the harder cases, so the per-ticket cost may run higher in practice, which is an argument for measuring your escalated-ticket cost separately rather than assuming it matches the blended average.
Step 5: compare totals
The combined cost with AI is about $78,000 a month ($34,250 for AI plus $43,750 for human-handled escalations), against the $125,000 all-human baseline. That is roughly $47,000 a month saved, or about $564,000 a year, on this hypothetical volume, before counting any revenue retained by faster resolution or the avoided cost of hiring and ramping to cover growth. The savings come from two places: the cost gap on each AI-resolved ticket, and the headcount you do not have to add as volume grows.
The point of the exercise is not the specific dollar figure, which depends entirely on your inputs. It is the structure: baseline minus the sum of AI cost and residual human cost, with escalations excluded from the AI bill and resolution defined by you. Build the model with your own ticket volume, channel mix, and fully loaded cost, and the answer will be specific to your business.
What Drives Payback Speed
Two teams with identical ticket volume can see very different payback periods. The variables below explain the spread.
True resolution rate, not deflection rate
Payback is driven by how many tickets the AI actually closes, end-to-end, without a human picking them up later. A high deflection rate that hides a low resolution rate produces reopened tickets, which cost you twice. The single biggest lever on payback is genuine end-to-end resolution, which is why the resolution definition in your contract is a financial term, not a technicality.
Channel mix
Voice resolutions cost more than chat, email, and SMS, both for the AI and for humans. A team with a heavy voice mix has a higher AI cost per resolution but also a higher human baseline to beat, so the gap usually still favors automation. The mix matters for modeling, not for whether the case closes.
Time to first production resolution
Every week the platform is in configuration is a week of baseline cost with no offsetting savings. Platforms with a forward-deployed implementation model and a sandbox you can stand up quickly start returning value sooner. Lorikeet's typical path is a sandbox in 20 to 30 minutes and a production-operational deployment in around a month, which compresses the period before savings begin.
Escalation accuracy
An agent that escalates too eagerly leaves savings on the table; one that escalates too rarely creates bad outcomes you pay for in CSAT and reopens. Because escalations are not billed under outcome pricing, accurate escalation is purely a quality lever, and getting it right is what lets you push the resolution rate up safely over time.
Volume growth you avoid hiring for
The clearest ROI in high-volume support is often the headcount you never add. If volume grows 30% and the AI absorbs the increase, the savings is the fully loaded cost of the agents, managers, and QA you did not have to hire and ramp. This avoided-cost line is frequently larger than the direct per-ticket savings and is the one most ROI models forget.
What to Measure
The metric you choose decides whether your ROI story survives contact with finance. These are the ones that hold up.
Cost per resolution
The denominator must be resolutions, not contacts or deflections. Track it separately for AI-resolved and human-resolved tickets, because the blended number hides whether the AI is taking the easy work and leaving humans the expensive long tail. Watch the trend over time as the agent handles more ticket types.
True resolution rate
Measure the share of tickets closed end-to-end with no human touch and no reopen within a defined window, such as 7 days. This is the number that should sit next to deflection on every dashboard, because it is the one that maps to dollars.
Reopen and recontact rate
A reopened ticket is a resolution that was not. Tracking reopens within a window catches the deflection-rate illusion, where tickets look closed but bounce back at a higher cost. If reopens climb as deflection climbs, the deflection number is fiction.
CSAT on AI-handled tickets
Cost savings that come with a CSAT drop are not savings, they are deferred churn. Segment CSAT by AI-handled versus human-handled so you can see whether automation is holding quality. The strong outcomes in this category pair automation with equal-or-better CSAT, not a quality trade.
Contact rate
Total contacts divided by active customers tells you whether you are resolving the root cause or just processing symptoms faster. A falling contact rate means the AI, and the analytics around it, are surfacing and fixing the issues that generate tickets. Coach-style automated QA on 100% of tickets is what makes this measurable at scale rather than from a 2% sample.
Pitfalls: The Deflection Vanity Metric and Other Traps
Most disappointing AI support deployments fail on measurement, not on technology. These are the traps that make a good ROI look bad and a bad one look good.
Buying on deflection rate
Deflection counts tickets the AI kept away from a human. It says nothing about whether the customer's problem was solved. A high deflection rate with a low resolution rate means customers are being turned away and coming back angrier and more expensive. Deflection is a vanity metric precisely because it is easy to inflate and disconnected from outcomes. Anchor on resolution instead, and make sure you, not the vendor, define it.
Letting the vendor define resolution
If the metric you pay on is defined by the party you pay, the incentive is to count generously. An outcome-priced contract is only as good as its definition of an outcome. The protection is a customer veto on what counts as a resolution, which keeps the billing metric and the business outcome aligned.
Ignoring the fully loaded human cost
Comparing an AI resolution price to a bare agent wage understates the human side by roughly half and makes the AI look less valuable than it is. Load benefits, tooling, QA, management, recruiting, and attrition into the human number before you compare. The wage is the floor, not the cost.
Forgetting escalation and long-tail cost
The tickets the AI escalates are the hard ones, and they cost more per ticket than your blended average. An ROI model that costs escalated tickets at the average understates the residual human spend. Measure escalated-ticket cost on its own, and treat the fact that escalations are not billed by the AI vendor as the relief it is rather than assuming those tickets are free.
Modeling on a peak resolution rate
Vendor case studies quote the resolution rate at maturity. Building your business case on that number sets an expectation you will miss in the first quarter. Model on a conservative rate, plan for it to climb as the agent learns more ticket types, and let the upside be a pleasant surprise rather than a missed forecast.
Where Lorikeet Fits
Lorikeet is an AI customer support platform built for complex and regulated businesses, and its commercial model is built around the ROI math above rather than against it. The pricing is per resolution, about $0.80 on chat, email, and SMS and about $1.00 on voice, with Coach QA at about $0.10 per ticket and escalations not charged. Crucially, the customer defines what counts as a resolution, so the number you are billed on is the number that matters to your finance team.
On the quality side, Lorikeet resolves issues end-to-end across chat, email, voice with sub-1-second latency, SMS, and WhatsApp on a single workflow engine, combining deterministic structured workflows with natural-language workflows. Its defence-in-depth approach (pre-launch adversarial simulations, inbound message checks, outbound guardrails, and 100% post-facto QA through Coach) is what lets a regulated team push the resolution rate up without the CSAT or compliance risk that usually accompanies automation. Anonymized proof points include a regulated fintech reaching roughly 85% automation with equal-or-better CSAT, which is the combination that actually drives ROI: more tickets resolved, quality held.
An honest limitation: a platform built for regulated depth carries a more involved implementation than a drop-in chat widget. Lorikeet uses a forward-deployed PM and engineer and a roughly one-month path to production, which is faster than most enterprise agent deployments but slower than pasting a script tag onto your site. For a high-volume team where the hard tickets are the expensive ones, that depth is the point. For a small team with simple FAQ deflection needs, a lighter tool may be enough.
If you are building the ROI case for AI support on a high-volume team, book a Lorikeet demo and bring your real ticket volume, channel mix, and fully loaded cost so the model is yours, not a generic slide.
Key Takeaways
ROI is the gap between your fully loaded human cost per ticket (roughly $1.25 to $4.00 at high volume) and your AI cost per resolution (about $0.80 on chat, email, SMS and $1.00 on voice with Lorikeet), multiplied across volume.
Load the human side honestly: wages, benefits, tooling, QA, management, recruiting, and attrition, not just the wage.
Outcome pricing only protects you when the customer defines the resolution and escalations are not charged, which keeps the billing metric aligned with the business outcome.
Payback speed is driven by true resolution rate, channel mix, time to first production resolution, escalation accuracy, and the headcount growth you avoid.
Measure cost per resolution, true resolution rate, reopen rate, CSAT on AI-handled tickets, and contact rate. Do not buy on deflection, which is a vanity metric you can inflate without solving anything.
Conclusion
The ROI of AI customer support for a high-volume team is real and structural: per-resolution costs sit well below fully loaded human costs, and the gap compounds across volume and across the headcount you avoid hiring as you grow. But the return only shows up in your finance system if you model the human side fully, pay on a resolution you define rather than a deflection the vendor controls, and measure outcomes rather than activity. Build the model with your own numbers, anchor on resolution and CSAT instead of deflection, and the platforms worth shortlisting are the ones whose pricing and quality controls are designed to make those numbers true rather than to make them look good.
