What AI Support Agents Actually Cost — and Save — in 2026

Every vendor deck shows the same math: agent handles 70% of tickets, humans cost $6 a ticket, therefore millions saved. The real model has more moving parts, and it is worth building honestly before you sign anything.

The costs that appear on no slide

Model spend is the smallest line. A well-routed support agent — cheap models for classification and retrieval, a reasoning model only for ambiguous cases — lands somewhere between $0.08 and $0.60 per resolved conversation depending on order complexity. If your vendor quotes flat per-seat pricing, ask what happens at 3x seasonal volume.

Integration is the real project. An agent that cannot read order status, shipment tracking and your returns policy is a deflection bot, and customers punish deflection. Budget for connecting the OMS, the shipping providers and the helpdesk — typically 60–70% of the initial build.

Evaluation is ongoing, not one-off. You need a labelled test set before launch and a regression suite forever after. Models change, your catalogue changes, your policies change. Teams that skip this discover quality drift from their CSAT reports, which is the expensive way to find out.

Where the savings are real

• First response time collapses. From hours to under a minute. This alone moves CSAT more than most teams expect, because waiting is the worst part of most support experiences.
• Resolution, not deflection. Order-aware agents with authority to act (within guardrails) fully close 55–70% of contacts in commerce. Deflection bots pretending to be resolution bots close far less and burn trust doing it.
• Seasonal elasticity. The agent handles Black Friday at the same cost per contact as a quiet Tuesday in February. Temporary hiring and training disappears.

The thresholds that matter

Below roughly 1,500 tickets a month, buy an off-the-shelf tool and move on — custom work will not pay back. Between 1,500 and 20,000, a configured platform with real integrations usually wins. Above that, custom orchestration with model routing starts beating platform pricing, often decisively.

One rule we hold in every AI build: refunds, cancellations and address changes get approval gates until the evaluation data says otherwise. Autonomy is earned with evidence, not granted at kickoff.

A number to take away

In our commerce engagements, the honest all-in cost per AI-resolved ticket — model, infrastructure, amortised build, ongoing evaluation — settles between $0.40 and $1.10. Against a $4–8 human cost per contact in the US and Western Europe, the payback is real. It is just 30–40% smaller than the vendor slide, and it arrives a quarter later than promised. Plan for that version and you will be pleased; plan for the slide and you will be explaining variance to your CFO.

If you want the spreadsheet behind this model, ask us for it — we share it freely, because clients who model honestly make better buyers.