The Cost Curve of Autonomous Email Resolution Over 24 Months

Almost every cost claim in this category is a two-point line. Here is what you pay a human, here is what you pay us, look at the gap. The problem is that nobody buys software at two points in time, and the shape between those points is where most of the surprises live.

So this is the shape. Twenty-four months of cost per resolved email ticket, modelled month by month, including the part at the beginning where the number goes up.

How these numbers were built

These figures are a model, not a measured study. They are composites assembled from deployment patterns we see repeatedly, expressed as medians with quartile ranges so you can see the spread rather than a single flattering number.

The reference profile is a company running 12,000 to 400,000 email tickets a year across e-commerce, SaaS, fintech, logistics, telecom, travel, insurance, edtech and marketplaces. Everything below is loaded cost, which means salary plus benefits plus tooling plus management overhead plus QA time, divided by tickets actually resolved. Not tickets touched. Resolved.

If your loaded cost per ticket differs from the baseline here, the absolute numbers shift but the curve shape holds. The shape is the point.

The baseline

A human-only email support operation in this profile runs about $6.20 per resolved ticket. That sits inside the $5 to $15 range we covered in more detail in our breakdown of cost per email ticket benchmarks, toward the lower end because email is asynchronous and agents batch it efficiently compared to voice.

The curve, month by month

Blended cost per resolved ticket across the whole queue, including tickets the AI handles and tickets that escalate to a person:

• Month 1: $7.90
• Month 2: $6.80
• Month 3: $5.90
• Month 4: $5.10
• Month 6: $4.20
• Month 9: $3.40
• Month 12: $2.90
• Month 18: $2.40
• Month 24: $2.30

End to end that is a 63% reduction. But read the first row again.

Month one is more expensive than doing nothing.

That is not a rounding artifact or a pessimistic assumption. It is what happens when you run two systems at once. Your agents are still working the queue, the AI is handling a small slice, implementation cost is being amortised, and somebody senior is spending hours a week reviewing drafts. Every honest deployment has this month. Vendors who tell you otherwise are quietly excluding implementation cost from the denominator.

The paired resolution rate

Cost falls because autonomy rises, so the two curves belong side by side:

• Month 1: 31% autonomous resolution
• Month 3: 52%
• Month 6: 66%
• Month 12: 74%
• Month 18: 78%
• Month 24: 79%

Notice how flat it goes after month 12. You gain 5 points across the second year, against 43 points in the first. Anyone modelling a straight line from their month-six number to 95% by year two is going to miss badly.

Two crossover points, and only one of them matters to finance

The monthly number drops below the $6.20 baseline in month 4. That is the one people celebrate.

The one your CFO cares about is cumulative breakeven, where total spend since kickoff finally equals what you would have spent staying manual. That lands around month 7, because months one through three were running a deficit that has to be paid back before the savings are real.

Three months of gap between "it's cheaper now" and "it has paid for itself" is worth naming out loud in your business case. We've watched teams present the month-four number to a board, get asked about cumulative position, and have no answer ready. If you're building the model yourself, our ROI calculation walkthrough covers the inputs in more depth.

Where the money actually goes

The blended number hides a composition shift that explains the whole curve. Cost per resolved ticket, broken into components:

Month 1

• Human handling of escalations: $5.05
• Ops, KB work and amortised implementation: $1.90
• Platform credits: $0.71
• QA and eval overhead: $0.24

Month 24

• Human handling of escalations: $1.31
• Platform credits: $0.62
• Ops and KB maintenance: $0.29
• QA and eval overhead: $0.08

Platform credits barely move. They fall from $0.71 to $0.62, roughly 13%, while total cost falls 71%. The software line is not what you are optimising. Human time on escalated tickets is 64% of month-one cost and it drops by three quarters, and that single line accounts for most of the savings on the chart.

Which is worth sitting with if you are evaluating vendors mainly on price per interaction. You are negotiating hard over the component that contributes least to the outcome. Our notes on how credits-based pricing works lay out why we structure it that way rather than per seat.

The number that goes the wrong way

Here is the finding that changes how you should read every resolution-rate chart you see.

Cost per escalated resolution rises over the two years, from $6.10 in month one to $9.40 in month 24. Up 54%.

The logic is not complicated once you see it. Your agent takes the tractable work first: order status, password resets, invoice copies. What remains in the human queue is the residue. Multi-issue threads, angry customers, edge cases that need a judgment call, technical problems requiring log access nobody has granted the agent yet. Those tickets take longer, need more senior people, and carry more rework.

So the blended average falls while the marginal ticket gets more expensive. Both things are true at the same time, and only one of them shows up in a vendor deck.

Two practical consequences. First, your remaining support team gets harder work, not less work, which is a staffing and retention question rather than a headcount question. Second, chasing the last five points of autonomy costs far more per point than the first fifty did. There is a rational stopping point, and it is usually earlier than the roadmap suggests.

The second drop nobody plans for

Most teams expect one improvement curve. The model shows two, and the second one arrives around months 14 to 16.

It does not come from more knowledge base content. By month 14 the KB is in decent shape and adding articles produces diminishing returns. The second drop comes from write access: letting the agent issue the refund rather than explain the refund policy, change the shipping address rather than describe how to change it, apply the account credit rather than route the request to someone who can.

An agent that can only read is a very good search interface. An agent that can act closes the ticket. The gap between those two is worth about 40 cents per resolved ticket in this model, and it is the single largest lever available after month twelve. We go deeper on this in our look at write-access integrations, which is where the mechanics live.

The reason it lands so late is usually organisational rather than technical. Somebody has to approve the agent moving money.

The spread is wider than the average

Medians are comfortable. The quartiles are more useful.

The top quartile hits $2.30 per resolved ticket by month 11. The bottom quartile is still at $3.90 at month 24, having taken twice as long to reach a number 70% worse. Same software, same category, same broad ticket mix.

Two things separate them, and neither is the vendor:

• Knowledge coverage depth. Not article count. Whether the answer to a real customer question exists in retrievable form, including the awkward exceptions that live in a senior agent's head and nowhere else.
• Permitted actions. How many things the agent is actually allowed to do without asking. Teams that cap this at "draft for review" indefinitely never get past roughly 55% autonomy, and their cost curve flattens early and high.

A useful diagnostic, if you're twelve months in and stuck above $3.50: it is almost never a model problem. Check what your agent is permitted to do before you check what it knows.

Where this curve does not hold

The model assumes a queue with enough repetition to learn from. Several situations break that assumption, and it is worth checking whether you are in one before you commit a number to a spreadsheet.

Very low volume. Below roughly 800 email tickets a month, the fixed costs of knowledge work, QA and eval do not amortise. The curve still slopes down but it flattens around $4, and the honest answer is that the savings case is weak even when the experience case is strong. Faster replies at 2am have value that this chart does not capture.

Genuinely bespoke queues. If most of your inbound is unique, the intent distribution has a long tail and no meaningful head, and there is nothing for the agent to get repeatedly right. Enterprise B2B accounts with deep custom implementations often look like this. Autonomy tops out in the 30s and the cost curve barely moves.

Hard seasonality. Businesses with a 6x November peak see the curve compress during the spike, because that is when deflection is worth the most, then partially rebound in the quiet months when human agents have spare capacity anyway. Annualise before you draw conclusions from any single month.

Organisational refusal to grant actions. Covered above, but it belongs in this list. A team that will not move past draft-review is not going to get this curve, and no amount of model quality changes that.

What this means if you are still building the business case

Model the dip. A business case that shows savings starting in month one will lose credibility the moment month one arrives, and it makes the genuinely good month-seven story look like a miss.

Set the target at the right number. Two-year steady state in this model is $2.30 against a $6.20 baseline. That is a strong result. It is not the 90% reduction that occasionally shows up in marketing material, and a business case built on 90% fails at month nine when reality shows up.

Budget for the second-year work separately, because the month 14 to 16 drop requires integration effort and approval cycles that nobody scoped in the original project. Teams that treat month twelve as the finish line stall at exactly the point where the remaining upside was cheapest to capture.

And build the escalation cost line into your staffing plan from the start. Your human queue is going to get smaller and harder at the same time. If you're weighing this against adding people instead, our hiring versus AI cost comparison works through the headcount side of the same question.

The curve is not a straight line and it never was. It's closer to a swimming pool: a shallow entry that gets worse before it gets better, a steep drop in the middle, and then a long flat bottom where most of the remaining decisions are about what you're willing to let the agent do rather than what it's capable of doing.

Frequently Asked Questions

What is a realistic cost per resolved email ticket with AI?

In this model, blended cost settles around $2.30 per resolved email ticket at 24 months, against a human-only baseline near $6.20. That is a 63% reduction, reached gradually rather than immediately. Month one typically runs higher than baseline because of dual running and implementation cost. Your own figure depends heavily on loaded agent cost, ticket mix and how many actions the agent is permitted to take without human approval.

How long before AI email support pays for itself?

Two dates matter and they are not the same. Monthly cost per resolution drops below the human baseline around month 4. Cumulative breakeven, where total spend since kickoff equals what staying manual would have cost, lands closer to month 7, because the first three months run at a deficit that has to be recovered. Present both in a business case, since finance will ask for the cumulative position.

Why does cost per escalated ticket go up over time?

Because the AI resolves the easy work first. What stays in the human queue is the hard residue: multi-issue threads, judgment calls, angry customers, technical problems needing access the agent lacks. Those take longer and need more senior people. In this model, cost per escalated resolution climbs from $6.10 to $9.40 over 24 months while the blended average falls. Both movements are real and they happen simultaneously.

Does the software licence drive most of the savings?

No, and this surprises most buyers. Platform credits move from $0.71 to $0.62 per resolved ticket across two years, roughly 13%, while total cost falls 71%. The savings come almost entirely from reduced human handling time on escalated tickets. Negotiating hard on per-interaction price optimises the smallest component. Integration depth and permitted actions matter far more to your final number.

Why do two companies using the same AI get different costs?

The top quartile in this model reaches $2.30 by month 11; the bottom quartile is still at $3.90 at month 24. Two factors explain most of the gap: knowledge coverage depth, meaning whether real answers including exceptions exist in retrievable form, and permitted actions, meaning how much the agent can do without asking. Teams that keep the agent in draft-review mode indefinitely tend to plateau near 55% autonomy.

Ready to model what autonomous email resolution would actually cost your team? Robylon AI resolves 60–80% of customer emails autonomously with agents that take action across Zendesk, Shopify, Stripe, Salesforce and 60+ other integrations. Start free at robylon.ai