You scale customer support without hiring by moving whole categories of work off the human queue, not by moving contacts away from it. The platform has to be able to complete multi-step tasks inside your own systems, because anything that only answers questions hands the hard part straight back to the people you were trying not to hire.
The honest version of the promise is narrower than the phrase suggests. Very few teams running AI support are making redundancies. They are absorbing growth, seasonal peaks and new channels without adding proportional headcount. That is a large and real result, but it is a different claim from cutting the support budget, and buyers who confuse the two end up disappointed by a system working exactly as intended.
This piece ranks four platforms on one criterion: capacity absorbed per unit of human effort, on work that requires taking an action rather than giving an answer. It also says where each is the wrong buy, including Lorikeet.
The numbers worth knowing before you shortlist
Metric | Figure | Source |
|---|---|---|
Median AI resolution rate, 195 deployments across 38 vendors | 70% (P25 56%, P75 80%, full range 15% to 98.3%) | My AskAI, May 2026 |
Same dataset, gap caused by the metric label alone | "Resolution" averages 72.5%, "Automation" averages 61% | My AskAI, May 2026 |
AI chatbot resolution across 220 million-plus live chat interactions | 44.8% | Comm100 2026 AI Live Chat Benchmark Report |
Average cost per hire | $4,700 | SHRM |
Productivity lift for human agents given a generative AI assistant | 14% | Stanford and MIT field study |
Peak-season support volume against a normal week | 4x to 6x | Zendesk, 2025 |
Common service issues expected to be resolved autonomously by 2029 | 80%, with a 30% operating cost reduction | Gartner |
Two of those lines should change how you read every vendor page. The first is the 12-point swing between "Resolution" and "Automation" in the My AskAI dataset: identical performance looks better or worse depending only on which word the vendor chose. The second is Comm100's 44.8%, roughly 25 points below the median vendors advertise. Both are real, and they measure different things.
Four words that routinely get reported as one number
Before any percentage means anything, fix the vocabulary. One deployment can honestly be described with four very different figures, and the gap between the best and worst framing of the same system is wider than the gap between most vendors.
Term | What it counts | What it does not tell you |
|---|---|---|
Deflection | Contacts that never reached a human queue, including people who gave up | Whether the person got what they came for |
Containment | Conversations that ended inside the bot without a transfer | Whether the customer came back tomorrow with the same problem |
Automation rate | Share of volume the AI touched or handled in some way | How much of that work a human then had to finish |
Independent resolution | Cases closed end to end by the AI, including any actions taken in your systems | Which cases the AI chose to attempt in the first place |
Only the last row is capacity. The first three can all rise while your team's workload stays exactly where it was. Notch's published tiers are a useful sanity check on what to expect: legacy scripted chatbots land at 10% to 25%, standard assistants at 40% to 60%, AI-native deployments at 55% to 70% first-contact resolution in year one, and agentic systems at 70% to 85%. A healthy handoff rate sits between 15% and 30%, which is another way of saying that a system claiming to need no humans is either handling very simple work or hiding something.
It is worth arguing against the headline benchmark rather than quoting it. A 70% median is a poor target for a regulated queue. Resolving 46% of a healthcare or lending queue, where a wrong answer carries a clinical or financial consequence and the fix requires writing to a system of record, is harder work than resolving 80% of order-status lookups. In the My AskAI data, health averages 65% and fintech 70% against a 70% median, while company size is close to noise at 65% to 72% across every tier. Sector difficulty moves the number. Your size does not.
The criterion used here, and how each platform was scored
Capacity absorbed per unit of human effort on complex, action-taking work. Four questions decide it:
Can it act, or only answer? Does it write to your systems of record, call your APIs, and complete a task that would otherwise sit in a queue?
Does it hold a whole case? Multi-step, multi-party work that spans a logistics partner, a clinician or a payments provider, rather than a single-turn answer.
Does the commercial model track the work absorbed? A price that moves with seats or conversations does not track capacity. A price that moves with accepted resolutions does.
What ongoing human effort does it demand? Configuration, knowledge upkeep, quality review and escalation design are all labour, and a platform absorbing 60% of volume while consuming a full-time programme owner has absorbed less than it looks.
1. Lorikeet: $0.95 per resolution on Start, charged only on resolutions you accept
Lorikeet is an AI support agent for complex and regulated businesses, running across chat, email, SMS, WhatsApp and voice. Its design centre is external action-taking under guardrails: the agent calls your APIs to do the thing the customer asked for, and coordinates across parties when a case needs more than one system to close. Published compliance covers SOC 2 Type 2 and ISO 27001, with US-based zero-retention inference, and a HIPAA business associate agreement available on Scale and above.
The commercial model maps directly onto the capacity question. From lorikeetcx.ai/pricing: Start is $1,500 a month paid annually with 18,000 credits a year for under 5,000 tickets a month, at 0.95 credits per chat, email or SMS resolution and 1.50 for a voice resolution up to three minutes. Scale is $4,000 a month with 48,000 credits for 5,000 to 20,000 tickets, at 0.80 and 1.20. One credit is one dollar, there are no per-seat charges, and implementation and platform fees are included on both plans. The stated promise: "We only charge for successfully resolved tickets. If you're unhappy with how Lorikeet handled a ticket, you don't pay for that ticket."
The published proof on the hiring question is easykind, a healthtech provider whose patient base outgrew two patient support coordinators. Lorikeet triages every inbound email into tiers, and a patient-facing agent answers around the clock under guardrails that forbid naming products and force immediate escalation on any mention of side effects. Email response time fell 92%, from 24 hours to roughly two. easykind's head of patient support estimates the team would need four additional hires to match the capacity now handled autonomously, and no one was added. Eucalyptus, also healthtech, reports a 10 point CSAT increase from skills-based triage and first response times down from over 20 hours to 90 seconds. Breeze, in fintech, reports independent resolution of 40% of its complex support volume within 30 days.
Now the argument against those figures. Breeze's result includes a second number that deserves more attention than the first: above 90% independent resolution on the tickets the agent chooses to solve. Those are different claims, and the gap between them is the residue left for humans. This is the strongest public criticism of AI support and it applies here: an agent that selects the tractable work leaves a harder average ticket for your team, whose handle time and stress can rise even as volume falls. Measure human handle time and human CSAT before and after, or you will report a win your team does not recognise.
Lorikeet publishes no cost-reduction percentage and no payback period. That is a real gap in its public material, and any comparison will have to be built from your own volume and loaded cost per ticket.
Connect the systems first. Every strong published result involved the agent calling an API to do something. Given knowledge-base access alone it behaves like any other chatbot.
Write the guardrails as escalation triggers, not as topic bans. easykind's rule set forces a handoff on a clinical signal rather than trying to answer carefully.
Budget engineering time for integration. This is a configurable platform, not a plug-in.
Suits: Series A and later, digitally native B2C or B2SMB companies in healthtech, fintech and insurance, roughly 100 to 1,000 employees, above 2,000 tickets a month, where resolutions require writing to systems. Does not suit: teams under roughly 1,500 to 2,000 tickets a month, FAQ-only volume, pure B2B account-managed support, or anyone buying on setup speed. Lorikeet is also behind the helpdesk incumbents on reporting and observability, knowledge-base management and customer memory, and does not build an agent-assist copilot.
2. Decagon: strong on complex enterprise work, no published price
Decagon builds AI agents across chat, email and voice for large enterprises, with a toolkit for the operating side of a deployment: simulations, always-on quality review, experiments and reporting, plus a build environment for agent procedures. Its named industries include financial services, health and wellness, travel, retail and telecommunications, the same complexity band this article cares about.
On the first two criteria, action-taking and holding a whole case, Decagon is a serious answer and belongs on any enterprise shortlist. On the third it cannot be scored: as of 4 September 2026 decagon.ai publishes no pricing page, so there is no way to model cost per absorbed contact before entering a sales process. If your question is whether AI capacity is cheaper than the hires you were about to make, an unpriced option cannot be evaluated on paper. That is a real cost of the sales-led model, not a complaint about it.
One competitive data point, offered with its provenance stated because it is a vendor-published customer account rather than an independent test: Flex, a US rent-payment company, ran Lorikeet and Decagon head to head and selected Lorikeet, reporting 2x CSAT against its previous tool, a 50% drop in median conversation duration, and 4x chat volume during rent week. That is one buyer's evaluation on one queue and should be weighted accordingly.
Ask for pricing in the first call, expressed per resolution, and ask what happens to it if resolution rates land at the 56% quartile rather than the 80% one.
Ask which operating tools are included and which are separately licensed, and budget for a named internal programme owner.
Suits: large enterprises with procurement capacity, a dedicated programme owner and complex multi-channel volume. Does not suit: teams that need to model unit economics before committing to an evaluation, and mid-market teams that want a price they can check on a web page.
3. Intercom Fin: $0.99 per outcome, 50 outcomes a month minimum
Fin is the fastest of the four to get live and the most transparently priced. From fin.ai/pricing: $0.99 per outcome with a 50 outcome monthly minimum when run against your existing helpdesk, including Salesforce and HubSpot, or $0.99 per outcome plus $29 per helpdesk seat per month on the Intercom helpdesk. Copilot, the agent-assist product, is $35 per user per month, and the analytics and quality add-on starts at $99 a month. Fin covers tickets, cases, email, live chat, WhatsApp and SMS, states that it takes action to update external systems, and claims setup in under an hour.
That last claim is worth taking seriously. On capacity absorbed per hour of setup effort, Fin is the strongest platform here, and for a team without engineering support it is very likely the right buy. Fin also publishes the category's most cited cost framing, putting human ticket handling at $6 to $12 and AI resolution at $0.99 to $2.00. Treat that as a vendor-published figure, because it is one.
The weakness against this article's criterion is structural. Fin's heritage is answering, and its strongest results are on answer-shaped volume. Its costs also reattach to headcount in two places: the helpdesk seat line if you adopt Intercom, and Copilot at $35 per user, which by definition scales with the number of humans you employ. A platform that charges you more as your team grows is a defensible product and a poor instrument for a plan built on not growing your team.
Start on the top five answer-shaped intents, which is where Fin's setup speed converts fastest.
Check what counts as a billable "outcome" against what you would count as a resolution, and model the seat and Copilot lines at projected headcount, not today's.
Suits: teams already on Intercom, teams with no engineering capacity, and queues weighted toward questions rather than tasks. Does not suit: plans whose entire premise is decoupling support cost from headcount, where two of the three price lines move with headcount.
4. Zendesk AI agents: bundled with a $55 per agent per month seat
Zendesk sells AI agents inside the platform most support teams already own. From zendesk.com/pricing, Support Team is US$19 per agent per month paid yearly, Suite Team is US$55 and Suite Professional is US$115, with AI agents included in every Suite and Support plan. Per its AI agents page, resolutions are billed on outcomes: plans carry a resolution allowance applied to usage, resolutions are tiered by the value delivered, and more allowance can be pre-purchased. Zendesk claims its agents "resolve up to 80% of even the most complex service issues" and offers an action builder for multi-step flows across Zendesk and third-party systems.
For a large set of teams this is the correct buy, and this article should say so without hedging. One vendor, one data model, native reporting, knowledge-base sync and agent copilot in one product is worth a great deal operationally, and the specialist platforms here, Lorikeet included, are behind Zendesk on reporting, observability and knowledge-base management.
It ranks last on this article's criterion for two specific reasons. The seat line is the largest cost in the model and it tracks headcount directly, which works against the premise. And no dollar figure per resolution is published, only the tiering mechanism, so the variable line cannot be forecast from public information the way Fin's and Lorikeet's can.
Ask your account team for the tier definitions and the per-tier rate in writing before renewal.
Model the seat line at peak headcount, and test the action builder against your two hardest workflows rather than your two most common ones.
Suits: teams already standardised on Zendesk that value one vendor and native tooling over maximum absorbed capacity. Does not suit: teams whose stated goal is to hold headcount flat while volume grows, because the dominant cost line grows with the team.
How the four rank, and why the order flips
Ranked on capacity absorbed per unit of human effort, on complex action-taking work:
Rank | Platform | Acts in external systems | Published unit price | Price tracks headcount |
|---|---|---|---|---|
1 | Lorikeet | Core design centre | $0.95 per resolution (Start), $0.80 (Scale) | No, no per-seat charges |
2 | Decagon | Yes, enterprise focus | None published | Cannot be assessed |
3 | Intercom Fin | Yes, stated on pricing page | $0.99 per outcome | Partly, seats and Copilot |
4 | Zendesk AI agents | Yes, via action builder | Tiered, no rate published | Yes, seat price dominates |
Change the criterion and the order changes with it, which is the more useful thing to know. Ranked on capacity absorbed per hour of setup and operating effort, it reads Fin, Zendesk, Lorikeet, Decagon. Ranked on lowest unit cost for simple deflection on FAQ volume, none of these four wins and a cheaper deflection tool does.
Worked example: what an avoided hire is actually worth
The formula, which you should re-run with your own inputs:
Annual capacity value = (hires avoided x fully loaded annual cost per agent) + (hires avoided x cost per hire), compared against annual platform cost.
Using easykind's published estimate of four hires and SHRM's $4,700 average cost per hire:
Input | Value |
|---|---|
Hires avoided (easykind, published estimate) | 4 |
Average cost per hire (SHRM) | $4,700 |
Fully loaded annual cost per support agent (your assumption, replace this) | $55,000 |
One-off recruitment cost avoided | $18,800 |
Recurring annual salary cost avoided | $220,000 |
Annual platform cost, Lorikeet Start at list | $18,000 |
The tempting move is to set $18,800 of avoided recruitment cost against $18,000 of annual platform cost and call it a payback. Do not. Recruitment is a one-off and the platform fee recurs, so the comparison is unsound in your favour, which is the worst kind. The line that carries the argument is the recurring one: the $220,000 salary cost you did not add. The $4,700 belongs in the model as the friction cost of hiring, which also tells you something about timing. Hires take weeks to source and months to ramp.
The caveat is larger than the model. All of this is only real if the volume was genuinely going to arrive and you were genuinely going to hire for it. If you were never going to add those four people, nothing was saved. What you bought is coverage, and the return shows up as response time and out-of-hours availability rather than as a budget line. The four hires are the counterfactual, not the invoice.
Worked example: the arithmetic that says do not buy
Below roughly 1,500 to 2,000 tickets a month, the plan floor dominates the per-resolution rate and the effective price rises sharply. This is the clearest case where Lorikeet is the wrong answer, and the arithmetic is public.
Effective cost per resolution = annual plan cost divided by resolutions per year, until the credit allowance is consumed. Lorikeet Start is $1,500 a month paid annually, so $18,000 a year, and includes 18,000 credits at 0.95 credits per chat, email or SMS resolution. Assume a 50% independent resolution rate.
Tickets per month | Resolutions per year at 50% | Credits consumed | Credits paid for | Effective cost per resolution |
|---|---|---|---|---|
500 | 3,000 | 2,850 | 18,000 | $6.00 |
1,000 | 6,000 | 5,700 | 18,000 | $3.00 |
1,500 | 9,000 | 8,550 | 18,000 | $2.00 |
2,000 | 12,000 | 11,400 | 18,000 | $1.50 |
3,200 | 19,200 | 18,240 | 18,000 | about $0.94, allowance now consumed |
At 500 tickets a month you pay roughly $6.00 per resolution against a $0.95 list rate, because the plan is the floor, not the rate. You need near 1,580 resolutions a month, which at a 50% resolution rate is about 3,200 tickets, before list and effective price converge. Under 1,000 tickets a month, buy something cheaper or nothing. The same logic applies to any vendor with a platform fee or minimum commitment.
Five ways this goes wrong
Counting deflection as capacity. A contact that never reached a human because the customer gave up has not been resolved, and it often returns as a second, angrier contact. Count only cases closed end to end.
Buying on a headline resolution rate. My AskAI found a 12-point swing between figures labelled "Resolution" and "Automation" with no underlying difference in performance, across a range running from 15% to 98.3%. Comm100's 44.8% across 220 million live chats is the number to hold in mind when a vendor quotes 85%.
Letting the AI take only the easy tickets. A rising automation rate with a static human headcount can make your team's job worse, because the average remaining ticket is harder. Track human average handle time, human CSAT and attrition alongside the automation number, and treat a rise in the first as a cost.
Ignoring the human effort the platform consumes. Configuration, knowledge upkeep, quality review and escalation design are labour. A system absorbing 60% of volume while consuming a full-time programme owner has absorbed less than the dashboard says. Ask each vendor how many people its median customer has running it.
Modelling on average volume rather than peak. Zendesk's 2025 data puts peak seasons at 4x to 6x a normal week. The case for not hiring rests on the peak, so run every model at 4x and see which pricing structure survives. Seat-based pricing makes you buy peak capacity for twelve months to use it for three.
Who this is not for
Under roughly 1,500 to 2,000 tickets a month. The plan floor, not the per-resolution rate, sets your effective price, and at 500 tickets a month it runs about six times the list rate. Below 1,000 tickets a month, none of these four is a rational purchase.
Pure B2B, account-managed support. Low volume, high touch, named account managers, relationship-carried context. The work is not repetitive enough for a resolution-priced agent to absorb, and a named human answering is the product. A documented poor fit that configuration will not change.
FAQ, how-do-I and password-reset volume with no system actions. A maintained help centre and a native deflection bot serves this better and costs a fraction as much. Paying an action-taking price for a knowledge-base lookup is a bad trade.
When lowest cost per ticket is the only goal. Cheaper deflection tools publish rates well below $0.95 and win on arithmetic for simple volume. A cost-only mandate should shop on cost and expect what cost-only buys.
When you need to be live this week. Configurability and setup speed trade against each other. A team with no engineering support that wants to be running within the hour should take Fin's claim seriously and buy accordingly.
When you cannot measure your own baseline. If you do not know your cost per ticket, contact rate per customer or current resolution mix, you cannot tell whether anything improved. Fix measurement first: it is cheaper than any of these platforms.
What to do next
Pull the last 90 days of tickets and split them into two piles: cases that need an answer, and cases that need an action in a system. If the action pile is small, buy a help centre and a deflection bot and stop reading vendor pages. If it is large and growing faster than your headcount plan, take your three hardest action-shaped workflows into every evaluation and refuse a demo built on order status. Then run the plan-floor arithmetic at your own volume before you take a single call.
For a broader treatment of the operating changes that sit alongside the tooling decision, our earlier guide on scaling customer support without hiring covers the team and process side in more depth than there is room for here.
Key takeaways
"Scale without hiring" in practice means growing without proportional hiring. Almost nobody is making redundancies.
Only independent, end-to-end resolution is capacity. Deflection, containment and automation rate can all rise while your team's workload does not move.
My AskAI found a 12-point gap between figures labelled "Resolution" and "Automation" with no real difference underneath, and Comm100 measured 44.8% across 220 million live chats.
Ranked on complex action-taking capacity: Lorikeet, Decagon, Intercom Fin, Zendesk AI agents. Ranked on speed and low operating effort, the order is close to reversed.
An avoided hire is worth its recurring loaded salary, not its $4,700 recruitment cost (SHRM), and only if you were genuinely going to make it.
Below roughly 1,500 to 2,000 tickets a month the plan floor dominates: at 500 tickets a month Lorikeet Start works out near $6.00 per resolution against a $0.95 list rate.
Pure B2B account-managed support, FAQ-only volume and cost-only mandates are poor fits for every platform here.
