AI automation does not have one universal price. A workflow that summarizes a form is different from a customer-facing system connected to billing, identity, private data, and multiple business applications.
The most reliable estimate starts with the process, not the model or software brand.
/ 01
The five cost categories
A complete budget includes more than build time. Separate one-time implementation, recurring provider charges, recurring service or maintenance, and internal staff time. Otherwise an inexpensive API can disguise an expensive operating process.
- Discovery: observation, process mapping, risk and access decisions
- Implementation: interface, integration, data model, prompts or rules
- Infrastructure: hosting, database, APIs, storage, email, voice, and monitoring
- Readiness: test data, security checks, documentation, training, and rollback
- Operations: review time, exception handling, maintenance, and future changes
/ 02
Complexity changes the estimate
Costs rise when inputs are inconsistent, several systems claim to be authoritative, permissions differ by role, information is regulated, volume is high, or actions are difficult to reverse. A narrow internal drafting workflow is usually less expensive than a public agent with identity, payments, calendar access, and write permissions.
Ask what must be true for the workflow to be production-ready. Authentication, audit evidence, signed webhooks, rate limits, backups, error queues, data retention, and incident response are real work even when the first demo took one afternoon.
/ 03
Keep infrastructure in your name
Whenever practical, the business should own its domain, hosting, database, API accounts, and billing relationships. This keeps control of data and assets with the client and makes provider costs visible. Management convenience can be offered separately, but ownership should never be ambiguous.
/ 04
Compare cost to the current process
Use one shared model: monthly current cost = volume × minutes per case ÷ 60 × loaded hourly cost, plus measurable rework, delay, and error cost. Expected monthly benefit = time actually removed × loaded cost, plus conservatively valued recovered outcomes. Do not count time that employees cannot realistically redeploy.
Example: 300 requests × 8 minutes at $36 per loaded hour is $1,440 per month. If a reviewed workflow safely removes five minutes from 70% of cases, gross labor capacity is $630 per month—not $1,440. Compare that conservative benefit with implementation, provider, review, and maintenance costs.
/ 05
Use a 12-month total-cost worksheet
Write one-time implementation in month zero, then forecast monthly software, usage, support, review labor, and contingency. Model low, expected, and high usage rather than one precise guess. Include an exit cost: data export, credential rotation, and handoff documentation.
A credible proposal identifies assumptions and shows which provider charges go directly to the client. Price certainty comes from a narrow scope and visible assumptions, not from pretending future volume and exceptions are known.
- One-time build and setup
- Monthly fixed providers
- Usage-based low / expected / high
- Human review and exception time
- Maintenance or partner service
- 10–20% contingency for unknowns
- Ownership and exit plan
/ Practical takeaway
Ask for a 12-month estimate that separates implementation, providers, internal review, maintenance, assumptions, and asset ownership—then compare it with a measured baseline.
Primary references
These sources support the factual and technical guidance in this article. Product decisions still require review against your own systems, policies, and risk.
OpenAI business leader’s guide: establish time, cost, and accuracy baselines NIST Secure Software Development Framework