How Much Does a Custom AI Feature Cost to Run Per Month?

Three parts, whatever the domain

Teams price the build and get surprised by the bill that arrives every month afterwards. AI feature run cost breaks into the same three parts regardless of what the feature does: model inference at your volume, the supporting infrastructure, and the human-in-the-loop staffing that never quite reaches zero.

Model inference

Per-call cost times calls per unit of work times monthly volume. The trap is "calls per unit of work" — it's rarely one:

• A support conversation is a multi-turn exchange with tool round-trips: several model calls, not one.
• A product search is an embedding call plus a re-ranking call.
• A fraud transaction might be a single reasoning call — but on every transaction.
• A voice minute is continuous generation.

Know your calls-per-unit, then multiply by volume, then add a multiplier for peak — traffic spikes multiply the bill exactly when you're leaning on the feature hardest.

Supporting infrastructure

A feature store's reads, a vector index, a message broker, async serving to hold many slow connections, logging and observability. Mostly flat, scaling in steps rather than smoothly. Cheaper when it reuses infrastructure you already run — a vector extension on a database you already operate costs far less than a dedicated vector service, and a queue you already run for other jobs costs nothing extra.

The line people forget: keeping it current

An AI feature isn't build-and-forget. Someone re-embeds the catalogue when it changes, retrains the model as fraud patterns or packaging shift, tunes the prompts and thresholds as the business changes, and reviews the escalation buckets weekly. It's a fraction of a role rather than a headcount, but it's a standing cost, and a feature with no one maintaining it degrades within months.

Human-in-the-loop

The escalation queue, the review band for ambiguous cases. It doesn't disappear. Roughly (1 − containment rate) × volume × loaded cost per handled case. On the projects in our portfolio, containment ran around two-thirds for the support agent — meaning a third of conversations still went to people, and that staffing is a real monthly line.

The model-choice lever

Routing the calls that don't need a frontier model to a smaller, cheaper one is often the single biggest saving — a small model can be an order of magnitude cheaper per call. The fraud engagement uses a lightweight reasoning model precisely because the call runs on every transaction.

A worked estimate

Put your own figures in:

• C = tasks per month (conversations, searches, transactions), Cp = the same for a peak month.
• k = model calls per task (a support conversation with two or three tool round-trips is usually 4-8; a search is 2; a fraud transaction may be 1).
• t = model cost per call at your token sizes.
• r = containment rate — the share the system handles without a human.

Monthly model cost ≈ C × k × t, and budget Cp × k × t for each peak month. Add roughly flat infra. Add escalation staffing ≈ (1 − r) × C × cost-per-handled-case. The number that swings the total is r, which a pilot pins down; everything else you can bracket up front. Re-quote per-token prices before committing — they move.

Where the estimate goes wrong

• Forgetting the peak multiplier. Black Friday or a fraud attack wave multiplies C and the model bill together.
• Forgetting escalation staffing. Containment is never 100%, and the remainder is a real monthly cost.
• Quoting a stale token price. Model pricing changes often enough that a six-month-old quote is a different number.

Where this stops being right

• Trivial volume — the fixed infrastructure dominates and the estimate is mostly "what's the smallest instance."
• A fully offline or batch feature — no per-request cost, just a scheduled job.
• When run cost is small relative to the value — don't over-model it.

FAQ

What's the biggest line usually? Model inference at volume, or the human escalation staffing that never reaches zero — which one leads depends on your containment rate.

Can I predict it before building? Within a range. Volume and calls-per-unit are knowable; the containment rate is the uncertain input, and a pilot narrows it.

Does a cheaper model always save money? On the calls that don't need frontier reasoning, yes. Routing the simple calls to a small model is one of the largest levers you have.

ISTRALLEN sizes the monthly run cost of an AI feature against real volume before the build starts — see what we do.