What Does an AI Support Agent Cost to Build vs License?

Three numbers, not one

AI support agent cost comes in three parts, and comparing them fairly is the whole exercise: the licence fee for a SaaS bot, the one-time build cost for a custom agent, and the monthly run cost that applies either way and never quite reaches zero.

Licensing a SaaS agent

Usually priced per resolved conversation or per seat. The attraction is no build cost and no maintenance. The catch:

• It scales with volume, and spikes in peak season — per-resolution pricing at Black Friday traffic is a large, forecastable line.
• Transactional actions are limited. You're often paying for FAQ deflection, not refunds executed under your rules.
• Little control over money-action guardrails, audit depth, or escalation tuning.

Building a custom agent: cost drivers

• Tools and integrations. Each API the agent calls — order lookup, returns, refunds, address change — is an integration, a schema, and a test suite. This is the biggest variable. Two tools is a small project; eight across three systems is not.
• The escalation console and helpdesk handoff. Building or integrating the surface your human agents pick up escalations in, with context attached.
• Conversation design and an evaluation harness. The harness that measures resolution and error rates on your historical tickets — you need it to tune safely.
• Safety and audit for money actions. Idempotency, approval thresholds, an audit log of every action.

The run cost — for either option

• Model inference per conversation. A support conversation is multi-turn with tool round-trips, so it's several model calls, not one. Multiply by monthly conversation volume, then again for the peak-season spike.
• Infrastructure. Async serving to hold many concurrent slow connections, plus a queue and datastore.
• Escalation staffing. The human share doesn't disappear. Whatever isn't contained still needs people.
• Ongoing tuning. Catalogue and policy changes mean the agent needs maintenance between peaks.

How to estimate yours

1. Monthly model cost ≈ conversations/month x average model calls per conversation x token cost per call. Add a multiplier for peak months.
2. Add infra and the escalation staffing for the share you expect to escalate.
3. Compare to a licence quote at your real volume, including peak.
4. Compare both to the loaded cost of the support hours displaced — wage plus overhead plus seasonal recru/training.
5. Payback on a build = build cost / (monthly labour saved − monthly run cost), measured across seasons.

What makes it pricier than expected

• Transactional tools — each one is bespoke integration and testing.
• The peak multiplier on model cost.
• You still staff escalation; containment isn't 100%.

What makes it cheaper than expected

• Reusing APIs you already have cuts the largest build driver. On our support agent project the agent was built on the client's existing order and returns APIs and their existing infrastructure — Postgres and a Redis queue they already ran — which kept the integration surface small.
• A narrow scope — order status plus returns — deflects a lot of volume for a modest build, with FAQ handled by cheaper retrieval.

A worked structure (variables, not numbers)

Put your own figures in:

• C = conversations per month, Cp = the same for a peak month.
• k = average model calls per conversation (a support conversation with two or three tool round-trips is usually 4–8).
• t = model cost per call at your token sizes.
• Monthly model cost ≈ C x k x t, and budget Cp x k x t for each peak month.
• Add infra (roughly flat) and escalation staffing = (1 − containment) x C x cost-per-handled-contact.
• Licence comparison: get the vendor's quote at C and at Cp and line it up against the above plus amortised build.

The number that swings it is containment — the share the agent closes without a human. Everything else you can bracket up front; containment you confirm with a pilot.

Where this stops being right

• Low volume. Build and run don't pay back; licence or use your helpdesk's built-in AI.
• Mostly informational tickets. A retrieval bot over a good help centre is far cheaper and enough.
• Estimates age fast. Re-quote model pricing before committing — multi-call conversations are sensitive to per-token cost.

FAQ

Is licensing always cheaper than building? At low-to-moderate volume, usually. As volume and transactional share grow, per-resolution fees can overtake the amortised cost of a custom agent.

What's the hidden cost people miss? The run cost — per-conversation model inference at peak volume, plus the escalation staffing that never goes to zero.

Can I predict run cost before building? Within a range: conversations x calls-per-conversation x token cost, plus infra. The uncertain input is containment rate, which a pilot narrows.

ISTRALLEN scopes and builds support agents against real conversation volume, sizing build and run cost up front; see AI for E-commerce.