Per-resolution AI pricing beats seasonal hiring on a spike because the bill follows the curve and a headcount plan cannot. A seasonal agent costs money to recruit, costs money for two to four weeks before producing anything useful, costs money to wind down, and none of those three line items shows up in a cost-per-ticket comparison.
The short version: if your volume is flat for ten months and 4x for six weeks, buy capacity that meters by outcome rather than by seat or by head. A seat-based tool makes you buy twelve months of peak capacity to use it for six weeks. A seasonal cohort makes you commit recruiting and training spend eight to ten weeks before the peak, against a forecast. A per-resolution charge is the only one of the three that costs you nothing in the ten quiet months.
The numbers, before the argument
Metric | Figure | Source |
|---|---|---|
Peak season volume against normal | 4x to 6x | Zendesk, 2025 |
Average cost per hire | $4,700 | SHRM |
Human ticket handling cost | $6 to $12 per ticket | published by Intercom's Fin |
Median AI resolution rate, 195 deployments across 38 vendors | 70% (P25 56%, P75 80%) | My AskAI, May 2026 |
Live chat resolution across 220M+ interactions | 44.8% | Comm100 2026 Benchmark Report |
Lorikeet published rate, Start plan | $0.95 per chat, email or SMS resolution | lorikeetcx.ai/pricing |
Intercom Fin published rate | from $0.99 per Fin outcome | intercom.com/pricing |
Flex chat volume during rent week | 4x the rest of the month | lorikeetcx.ai customer story |
Define the words before you compare the percentages
Peak-season claims are unusually slippery because four different measurements get reported as one number. Fix the definitions first or the comparison is meaningless.
Deflection. The customer never opened a ticket. Measured by absence, which means it is measured badly: a help centre article someone read and then contacted you anyway still counts in most deflection reports.
Containment. The conversation ended inside the AI without reaching a human. An abandoned chat is contained. Abandonment rises during a spike, so containment flatters itself exactly when you are least able to check.
Automation rate. The share of conversations handled without a person involved. This is the number most case studies publish, including Hnry's roughly 70% at peak.
Resolution. The customer's issue was actually closed out. The strictest of the four, and the only one worth paying for.
The gap between the labels is measurable. In My AskAI's May 2026 dataset of 195 deployments across 38 vendors, figures reported under "Resolution" averaged 72.5% while figures reported under "Automation" averaged 61%: a 12-point swing produced by the word alone, not by the software. Comm100's 2026 benchmark, drawn from more than 220 million live chat interactions, puts chat resolution at 44.8%. That is the number to hold in your head whenever a vendor quotes you 80%.
The arithmetic of a six-week 4x spike
Cost-per-ticket comparisons assume a flat month and almost nobody has one. Here is a team modelled against a curve instead of an average. Replace every input with your own.
Inputs
Input | Value |
|---|---|
Baseline volume | 2,500 tickets per month |
Peak multiplier | 4x |
Peak duration | 6 weeks |
Tickets per agent per working day | 40 |
Working days inside the peak window | 30 |
Fully loaded agent cost | $25 per hour |
Ramp before an agent is productive | 3 weeks |
Recruiting cost per seasonal hire | $1,500 (SHRM puts the all-hire average at $4,700) |
Paid hours per seasonal agent, ramp plus window | 342 |
Outputs: the seasonal hiring path
Output | Working | Result |
|---|---|---|
Tickets in the peak window | 6 weeks at 10,000 per month | 15,000 |
Baseline tickets in the same window | 6 weeks at 2,500 per month | 3,750 |
Incremental tickets | 15,000 minus 3,750 | 11,250 |
Seasonal agents required | 11,250 / (40 x 30) | 10 |
Recruiting | 10 x $1,500 | $15,000 |
Wages across ramp and window | 10 x 342 x $25 | $85,500 |
Supervisor and QA cover | 1 x 342 x $45 | $15,390 |
Total | $115,890 | |
Cost per incremental ticket | $115,890 / 11,250 | $10.30 |
That $10.30 lands near the top of the $6 to $12 per-ticket range Intercom's Fin publishes for human handling, which is a sanity check on the model. Note what the ramp does to the timing: to have ten productive agents on day one of the peak you start recruiting eight to ten weeks earlier, on a forecast. If the peak comes in at 3x rather than 4x, the recruiting money and most of the wages are already spent.
Two costs are missing from that table because they are hard to price and easy to ignore: wind-down, meaning offboarding, licence clawback, notice periods and churn among people who leave before the peak ends; and the load half-trained temporary staff put on your permanent agents at the worst possible moment.
Outputs: the per-resolution path on the same curve
Take the same 11,250 incremental tickets and assume the agent closes out 60% end-to-end. On Lorikeet's published Scale plan, chat, email and SMS resolutions meter at 0.80 credits and one credit is one dollar.
Output | Working | Result |
|---|---|---|
Incremental tickets resolved by the agent | 11,250 x 60% | 6,750 |
Credits consumed in the peak window | 6,750 x 0.80 | 5,400 |
Dollar value of those credits | 1 credit = $1 | $5,400 |
Incremental tickets still needing a person | 11,250 minus 6,750 | 4,500 |
Seasonal agents still required | 4,500 / 1,200 | 4 |
Seasonal agents avoided | 10 minus 4 | 6 |
Hiring cost avoided | 6 x ($1,500 + 342 x $25) | $60,300 |
Now argue against that number
The $5,400 is a marginal figure drawn against a pool you have already bought, and quoting it alone would be dishonest. Run it the other way. Scale is $4,000 a month paid annually, so $48,000 a year, and this team's full-year volume is 41,250 tickets. At 60% automation that is 24,750 resolutions consuming 19,800 credits against 48,000 included, an effective $1.94 per resolution, well above the published $0.80. The published rate is what the meter charges; the effective rate is what you pay, and the two converge only when you consume the pool.
If peak cover is the only thing you want, you are buying twelve months of platform to solve six weeks, which is the same error as buying twelve months of seats. The whole difference is that the platform works the other forty-six weeks and the ten seasonal agents have gone home. Judge a platform on the full year of work it absorbs, never on the spike alone.
Ranked: how four charging models behave under a 4x spike
This ranking is on one criterion only: how closely the bill tracks a demand curve that is flat for ten months and 4x for six weeks. It is deliberately different from a flat-volume ranking, and it is not a ranking of product quality. All prices are list prices from each vendor's own public pricing page, checked on 4 September 2026.
1. Intercom Fin: from $0.99 per outcome, and no seats required
Fin has the most elastic bill in this set, so on the stated criterion it ranks first. Intercom's pricing page lists Fin at from $0.99 per Fin outcome, and the same page offers Fin on your existing helpdesk with "no seats required". Nothing to pre-buy, nothing to commit to in advance, nothing to wind down. The bill goes up 4x in the peak weeks and back down afterwards, which is exactly the behaviour you want from a spike instrument.
The catch is what sits underneath it. If Intercom is also your helpdesk, seats are $29, $85 or $132 per seat per month on Essential, Advanced and Expert, and those seats behave like every other seat under a spike. Fin's elasticity applies to the AI layer, not the people layer.
Model it as pure variable cost: peak outcomes times the outcome price, no annual floor.
If Intercom is also your helpdesk, model the seat layer separately on the same curve.
Test whether Fin clears your peak tickets before assuming elasticity becomes capacity.
2. Lorikeet: an annual credit pool, $0.95 or $0.80 per resolution, and no per-seat charge
Lorikeet ranks second on elasticity and first on what happens to your roster. Published pricing is $1,500 a month on Start with 18,000 credits a year, or $4,000 a month on Scale with 48,000 credits a year: chat, email and SMS resolutions meter at 0.95 and 0.80 credits, voice at 1.50 and 1.20 for resolutions up to three minutes, and there are no per-seat charges at all. The pool is annual rather than monthly, so a six-week draw-down at four times the usual rate is simply a draw-down. There is no purchase decision to make at the moment your queue is on fire.
The commercial promise on that page is quotable verbatim: "We only charge for successfully resolved tickets. If you're unhappy with how Lorikeet handled a ticket, you don't pay for that ticket." That matters more under a spike than in a flat month, because peak is when quality slips and when you are least able to audit it.
Where it ranks below Fin: the pool is bought up front, so a team that does not consume it pays for capacity it did not use, exactly as a seat-based buyer does. Where it ranks above everything seat-based: adding six human agents for the peak costs nothing on the Lorikeet line. One caveat that is frequently glossed over: Lorikeet runs on top of your helpdesk rather than replacing it, so your helpdesk's seat economics still apply.
Work out your annual resolution count first and check it against the credit pool, not the monthly ticket band.
If your peak month crosses into the next band, ask which plan you will be quoted on. The published bands describe steady-state volume and your curve is not steady state.
Price voice separately if your peak is a phone peak: 1.50 credits on Start, 1.20 on Scale, for resolutions up to three minutes.
3. Freshdesk: $19 to $89 per agent per month, plus day passes at $2 and AI sessions at $49 per 100
Freshdesk is seat-based, which normally sinks a tool on this criterion, and it ranks third anyway because it is the only vendor here publishing a purpose-built spike instrument. Its pricing page lists Growth at $19, Pro at $55 and Enterprise at $89 per agent per month billed annually, and separately lists day passes at $2 per pass, described there as a way to give occasional agents temporary access and manage a sudden spike in support. The AI layer is metered too: Freddy AI Agent includes the first 500 sessions, and additional sessions are $49 per 100, roughly $0.49 each, with a session defined on Freshworks' own pricing FAQ as a unique interaction between an end user and the AI agent.
Run the day pass arithmetic before assuming it wins. A Growth seat at $19 a month across roughly 20 working days is about $0.95 a day, so a $2 pass pays off only for agents working fewer than about ten days a month. Across a six-week cohort, 45 working days of passes is $90 per agent against $42.75 for a Growth seat, and the seat wins. Day passes suit overflow cover, weekend surges and on-call staff, not a full-time seasonal cohort.
Use day passes for irregular overflow, monthly seats for full-time seasonal staff.
Model Freddy sessions separately from seats: 500 included, then $49 per 100.
Remember a session and a resolution are different units. A session is an interaction, not an outcome.
4. Zendesk: $55 to $115 per agent per month, and a monthly allowance that does not carry over
Zendesk ranks last on this criterion, and only on this criterion. Its pricing page lists Suite Team at US$55 and Suite Professional at US$115 per agent per month paid yearly, Enterprise on request. AI agents are included on every Suite and Support plan, charged per automated resolution, but that rate is not published, so you cannot model a spike from public information alone. Zendesk's own help documentation describes a resolution allowance sized by plan type and agent seat count that does not carry over month to month, with additional allowance purchased on top.
A non-carrying monthly allowance is the worst possible meter for a spiky curve. You forfeit unused allowance for ten months and overshoot it in the six weeks that matter, and because the allowance is a function of seat count, the only structural way to raise it is to buy more seats. A team with 12 permanent agents needing 22 at peak, on Suite Professional at US$115 paid yearly, pays $30,360 a year for 22 seats against $16,560 for 12. That $13,800 difference buys ten seats for twelve months to use them for six weeks, and 105 of those 120 seat-months, worth $12,075, are capacity you never touch.
Ask what your per-resolution rate actually is before modelling anything, because it is not on the public page.
Check whether your contract lets seat count fall back after the peak or only rise. Ask what happens in month seven, not month one.
Check whether unused monthly allowance carries. If it does not, either size on the peak month and accept ten months of waste, or size on baseline and budget for the overshoot.
Summary: the four models against the same curve
Vendor | How it meters | Behaviour under a 4x six-week spike | Best suited to |
|---|---|---|---|
Intercom Fin | From $0.99 per outcome; optional seats $29 to $132 | Tracks the curve; no floor on an existing helpdesk | Any spike shape, if it clears your spiking tickets |
Lorikeet | $1,500 or $4,000 a month, annual credit pool, 0.95 or 0.80 credits per chat resolution, no seats | Pool absorbs the spike with no purchase decision; floor must be consumed to hit the published rate | Recurring or staggered peaks, tickets needing system actions |
Freshdesk | $19 to $89 per agent per month; day passes $2; Freddy sessions $49 per 100 after 500 | Day passes flex; full-time seasonal seats do not | Short simple spikes, irregular overflow staff |
Zendesk | US$55 to US$115 per agent per month paid yearly; per-resolution rate not public | Allowance is seat-linked and does not carry month to month | Flat or steadily growing volume, not a spiky curve |
Elasticity is not the same as capacity
Everything above measures whether the bill tracks the curve. It says nothing about whether the software can clear the work, and under a spike that second question decides the outcome. The tickets that spike are rarely the ones a knowledge-grounded answer resolves: they are account-specific, they need something looked up or changed in a system, and they come from customers who are already stressed. Three published examples, three different spike shapes.
A monthly spike. Flex lets renters split monthly payments, so its peak arrives every rent week. Its published story records 4x chat volume during rent week against the rest of the month, plus 2x CSAT compared with its previous support tool and a 50% decrease in median conversation duration to resolution. Flex ran a head-to-head test against Decagon before choosing Lorikeet, having decided to migrate off Ada. The commercially interesting part is the shape: a monthly spike consumes an annual pool twelve times a year rather than once, which is why a recurring peak fits a pooled model far better than an annual one.
Staggered annual spikes. Summ, formerly Crypto Tax Calculator, handles cryptocurrency tax reporting across Australia, the US, UK, Canada and Europe, so it faces several tax seasons at different points in the year. Its published story records first response times falling from around 30 minutes with human agents to under a minute, described as 97% faster resolutions, plus refund workflows automated down to eligibility checks and escalation of approved cases. You cannot hire, train and wind down a cohort four times a year in four jurisdictions, which is where a headcount plan has no answer at all.
One brutal week. Hnry deployed on Australian tax, the hardest jurisdiction it operates in, going live in mid-May 2026 from a baseline of zero automation and handling more than 17,000 support conversations in the first month. By the peak week of the Australian end of financial year it was automating around 70% of conversations, up from roughly 58% across the full period. Use Hnry's wording carefully: that is an automation figure, not a resolution rate, and per the definitions above they are not the same number. The directional point is what matters. Automation share went up as load went up, the opposite of what happens to a roster of tired people, and the extra capacity arrived without a scheduling decision.
For a wider look at which tools suit different peak shapes, see our rundown of the best AI support tools for seasonal peak volume, which covers a broader vendor set than the four here.
Who this is not for
A genuinely once-a-year spike on simple, low-stakes tickets
If your peak is annual and the spiking tickets are FAQ lookups, order status and password resets with no system actions behind them, do not buy a platform priced for complex action-taking work. Temporary staff plus a cheap deflection bot beats it on arithmetic. Here is where the flip happens.
Lorikeet's Start plan is $1,500 a month paid annually, so $18,000 a year, and includes 18,000 credits at one credit to the dollar. Chat, email and SMS resolutions meter at 0.95 credits, so the pool buys 18,947 resolutions a year, about 1,579 a month. You only pay the published $0.95 if you consume the pool. At 60% automation that takes roughly 2,632 tickets a month; at 50%, roughly 3,158. Below that line you pay the same $18,000 for less work and the effective rate climbs.
Monthly ticket volume | Resolutions per year at 50% automation | Effective cost per resolution on Start |
|---|---|---|
550 average (400 baseline, 4x for six weeks) | 3,300 | $5.45 |
1,500 | 9,000 | $2.00 |
2,500 | 15,000 | $1.20 |
3,158 or above | 18,947 | $0.95, the published rate |
A team on 400 tickets a month that quadruples for six weeks handles 6,600 tickets a year and pays an effective $5.45 per resolution. Freshdesk publishes additional Freddy AI Agent sessions at $49 per 100 after the first 500, about $0.49 each. The two units are not identical and it matters: a Freddy session is a unique interaction, a Lorikeet credit is charged on a successful resolution. On simple tickets they are doing comparable work, and the cheaper unit wins. Lorikeet's price assumes the resolution involved calling your systems and changing something; if it did not, you are overpaying. Lorikeet's own documented buyer profile agrees, excluding companies under roughly 1,000 to 2,000 tickets a month and buyers whose only goal is the lowest cost per ticket.
Teams already committed to a helpdesk-native stack
If you are on Zendesk or Intercom, want the AI working the same tickets in the same data model, and need native reporting, knowledge base sync and an agent-assist copilot in the inbox, the native add-on is one vendor and one contract. Lorikeet is behind on reporting and observability, knowledge base management and customer memory, and deliberately does not build an agent-assist copilot. Those are exactly the tools a support manager reaches for during a peak.
Setup-speed buyers, voice-led peaks and pure B2B support
Lorikeet trades plug-and-play for configurability, so if your peak is eight weeks away and you have no engineering support to connect the systems the spiking tickets touch, pick something live within the hour and deliberately shallower. If the constraint at peak is telephony rather than reasoning, judge the voice stack on its own terms; Lorikeet targets parity there rather than leadership. And if your support is pure B2B, account-managed and high-touch, a spike is a scheduling problem for named account managers, not a metering problem.
Five ways peak planning goes wrong
Sizing on the peak instead of on the curve. Both the seat model and the headcount model make you buy for the highest point on the graph and then hold it. The question that matters is what that capacity costs in the ten months you do not need it.
Treating ramp as free. Two to four weeks before a seasonal agent is productive means the hire decision happens eight to ten weeks before the peak, on a forecast. In the model above, ramp is $28,500 of the $115,890, spent before a single peak ticket arrives.
Buying seats you cannot hand back. Adding seats mid-term is easy at every vendor here. Removing them before renewal usually is not. Ask the question in the negotiation, not in month seven.
Confusing peak volume with peak concurrency. Four times the tickets across six weeks is a staffing problem. Four times the tickets arriving in a two-hour window on the first of the month is a concurrency problem, and no human roster solves it at any price you would pay. Look at your hourly distribution, not your monthly total, before deciding which one you have.
Assuming the AI takes the hard tickets. It usually does not, and the criticism deserves a straight answer: AI absorbs simple volume and leaves humans a harder average mix, and at peak that concentration is sharper because simple tickets are exactly what spikes. Plan for your human queue to be worse per ticket than before, staff it with experienced people rather than temporary hires, and measure handle time on the human residual separately so you can see it happening.
Key takeaways
Peak seasons run 4x to 6x normal volume, per Zendesk's 2025 figure, and a flat monthly average hides that entirely.
Ten seasonal agents for a six-week 4x window model out at $115,890, or $10.30 per incremental ticket, near the top of the $6 to $12 human handling range Intercom's Fin publishes.
SHRM's $4,700 average cost per hire is the ceiling on seasonal recruiting, not the floor, but recruiting plus a three-week ramp is committed spend whether or not the peak arrives.
On bill elasticity, Intercom Fin's per-outcome charge ranks first, Lorikeet's annual credit pool second, Freshdesk's day passes third, Zendesk's seat-linked non-carrying allowance last.
Published rate and effective rate are different numbers. Lorikeet's $0.95 on Start becomes $5.45 for a team doing 3,300 resolutions a year, because the $18,000 annual floor dominates.
Below roughly 2,600 to 3,200 tickets a month, on simple tickets with no system actions, temporary staff plus a cheap deflection tool is the better buy.
Automation rate is not resolution rate. Hnry's roughly 70% at peak is an automation figure in Hnry's own wording, and Comm100's 44.8% across 220 million chat interactions is the counterweight to any 80% claim.
What to do next
Pull last year's ticket export and bucket it by week rather than by month. Two numbers fall out: your true peak multiplier and the real length of your peak window. Re-run the two tables above with your own throughput, loaded hourly cost and automation share.
If your peak weeks account for more than about 15% of annual volume, and the tickets that spike touch your systems rather than your help centre, per-outcome models are worth quoting and worth testing on your real peak tickets rather than a demo script. If your spike is FAQ volume once a year, price a deflection bot and two temporary agents first. Our guide to handling seasonal support spikes with AI covers the operational preparation, which is the half of this that pricing cannot solve.
