What production AI actually costs
A plain-English breakdown of the cost of shipping AI to production - build, run, governance, and the line items most vendors quietly leave out of the first quote.
"How much does it cost to ship an AI system?" is the question we hear most often in first calls, and it is harder to answer honestly than it looks. Not because the numbers are secret - they aren't - but because most of the public numbers are for the build, and most of the cost in a year-long programme is not the build.
This piece is a plain-English breakdown of where the money actually goes when you deploy AI into an enterprise, with rough ranges for each line. Ranges, not promises - your mileage will vary. The point is to put all the line items in one place so a first conversation can be grounded.
Build cost: the part you usually get quoted
A competent delivery team can ship a first production AI workflow - agent, automation or narrow product - in eight to twelve weeks. That build typically costs between $80,000 and $250,000 depending on scope, integration surface and governance requirements.
Inside that envelope:
- Design and architecture (~10–15% of build cost). Scoping, UX, data model, integration map. Often under-weighted and later regretted.
- Engineering (~45–60%). Frontend, backend, model layer, data layer.
- ML / model work (~10–20%). Prompting, tool design, fine-tuning if needed, eval harness. A bigger share on agent work; smaller on pure automation.
- Governance and observability (~10–15%). PII redaction, audit trail, evaluation harness, incident runbook.
- Launch and hypercare (~5–10%). The first thirty days after go-live matter more than almost anything else in the build.
A quote that doesn't explicitly include governance and hypercare is a quote that's about to get a change order.
Run cost: the part you don't get quoted
Build is roughly a one-off. Run is a permanent line on the P&L. Run cost has three components that most first-draft budgets under-estimate.
Inference cost
Model usage. For a narrow automation that processes a few thousand events a day, inference cost can be under $500 per month. For a voice agent handling a million calls a month, it can be $50,000 per month. The range is that wide because inference cost scales with use and with model tier.
Useful rules of thumb:
- A reasoning-tier model costs 10–30× a fast-tier model per token.
- A single agent call with tool use is typically 5–20 calls to the model, not one.
- Caching and batch inference can cut bills by 40–70% if designed in early.
- Fine-tuned or distilled small models are usually cheaper and faster in production than large frontier models.
Data cost
Warehouses, pipelines, storage, egress - the underlying data platform that serves every workflow above it. For a first workflow, typically $2,000–$10,000 per month. For a platform supporting a programme of half a dozen workflows, $10,000–$40,000 per month. Governance and lineage tools add a further 20–30%.
Operations cost
The humans who run the thing. On-call rotation, weekly evaluation review, monthly regression work, quarterly model refresh. For a single workflow delivered to production-grade, budget 2–3 engineer-days a month to keep it honest. For a full programme, closer to a full-time operator. This is the line most pilots never budget for, which is the primary reason most pilots never make it to year two.
Governance cost: non-negotiable
Governance is not an appendix - it is a first-class system. In regulated industries (finance, healthcare, public sector, education), governance is typically 15–25% of year-one programme cost. That includes the evaluation harness, red-team probes, PII/PHI redaction, audit logging, policy-as-code and the external review the CISO office will insist on.
If you try to save money by deferring governance, you will spend it back three times over in the next external audit. This is something we have seen happen at a dozen clients.
A worked example
To make this less abstract, here is a representative year-one budget for an enterprise that ships one production AI workflow and one agent:
| Line item | Range | |---|---| | Scoping engagement | $25,000 – $75,000 | | Workflow build (10 weeks) | $120,000 – $200,000 | | Agent build (8 weeks) | $100,000 – $180,000 | | Governance & eval harness | $60,000 – $120,000 | | First-year inference | $20,000 – $100,000 | | First-year data platform | $25,000 – $100,000 | | First-year operations | $60,000 – $150,000 | | Total, year one | $410,000 – $925,000 |
The wide range reflects how much depends on scale, integration surface, and regulatory requirements. A startup shipping a narrow automation can be at the low end. A public-sector programme with on-prem deployment, sovereign data handling and external audit can be above the top.
What to cut, safely
If the year-one number is higher than your envelope, the right cuts are:
- Narrow the scope - one workflow at a time, not three in parallel.
- Pick fast-tier models by default; reserve reasoning-tier for the 10% of calls that actually need it.
- Re-use platform investments where governance and data plumbing can serve later bets.
The wrong cuts are governance, evaluation and hypercare. Skipping those does not save money - it defers and amplifies it.
Related reading
Frequently asked
How much does it cost to deploy AI in an enterprise? Year-one programmes typically run between $400,000 and $1,000,000 including build, governance, infrastructure and operations. The exact figure depends on scope, scale and regulatory requirements.
What is the ongoing run cost of an AI system? Run cost is a combination of inference, data platform, and operator time. Budget 15–30% of the build cost annually as a rough planning number, adjusted for scale.
Is it cheaper to build AI in-house or through an agency? Build cost is comparable either way. The real difference is speed and governance maturity. An experienced partner reaches production faster and usually ships governance that an in-house team takes a further six months to reach.
How should I budget for governance? Budget 15–25% of year-one programme cost for governance. Less, if you are not in a regulated industry; more, if you are in public sector or healthcare.
For a grounded first-read of your own programme budget, book a scoping call - we'll give you ranges for your specific scope, not a generic quote.