"How much does an AI tool cost?" is a fair question with an annoying answer, because the phrase covers everything from a weekend script to a system that touches your billing. What follows is how the money actually breaks down, with ranges you can take into a budget conversation.
Three cost buckets, not one
Every honest AI proposal separates these. Be suspicious of any that does not.
1. Discovery and scoping
Mapping the process, finding where the data lives, deciding what "good" looks like, and agreeing what the system is not allowed to do. Typically $3,000–$12,000. Skipping it is the single most reliable way to overspend later, because you end up building the wrong thing accurately.
2. The build
Roughly, by ambition:
- Proof of concept — $8,000–$25,000. One workflow, your real data, no integrations, not production-hardened. The point is to answer "does this work well enough to be worth continuing."
- Production single-workflow tool — $25,000–$75,000. Authentication, logging, error handling, an admin view, a human review step, and integration with one or two systems you already run.
- Multi-workflow or customer-facing — $75,000–$250,000+. Several processes, multiple integrations, real availability requirements, and a compliance story.
3. Running it — the part that gets forgotten
This is monthly, it never stops, and it is where budgets get ambushed.
- Model API usage. Usually $50–$2,000/month for internal tools at normal volume. Genuinely high-volume or long-document workloads go higher, but this is rarely the biggest line.
- Hosting, storage, and search infrastructure. $100–$800/month for most builds.
- Maintenance. Budget 15–25% of the build cost annually. Models get deprecated on real timelines, your documents change, and accuracy drifts. This is the line vendors omit to look cheaper.
Why the API bill is smaller than you think
People anchor on model pricing because it is the only number published. In practice, for an internal tool used by a few dozen staff, inference is often a rounding error next to the engineering. The expensive parts are getting your data into usable shape, integrating with systems that were not designed to be integrated with, and the review workflow around the output.
If a vendor's cost story is dominated by token pricing, they are probably quoting you a thin wrapper.
What moves the number most
- Data condition. Clean, structured, in one place: cheap. Scattered across scanned PDFs, a shared drive, and someone's inbox: expensive. This is the single biggest swing factor.
- Integration count. Each system you must read from or write to adds cost, and the ones with poor or no API add the most.
- Accuracy bar. "Helpful draft a person reviews" is dramatically cheaper than "correct without supervision." Every nine of reliability costs real money.
- Compliance scope. HIPAA, CJIS, CMMC, or FedRAMP change the architecture, the hosting, and the documentation burden. See our note on security and authorization work for what that involves.
A realistic first year
For a mid-sized business automating one meaningful internal process:
- Discovery: $6,000
- Production build: $45,000
- Running costs: ~$400/month → $4,800
- Maintenance (partial year): $5,000
- Year one: roughly $60,000. Year two: roughly $15,000.
Whether that is a good deal depends entirely on what the process costs you today. If it consumes 20 hours a week of loaded labor, it pays back inside a year. If it consumes two, it does not, and no vendor should pretend otherwise.
How to keep the number honest
- Insist on a fixed-price discovery phase before any build estimate. Anyone quoting a build price before seeing your data is guessing.
- Ask what happens when the model version is retired, and who pays for that migration.
- Ask who owns the prompts, the evaluation set, and the code. Get it in writing.
- Require a measurable success criterion in the statement of work, not "improved efficiency."
Software procurement discipline applies here exactly as it does to anything else you buy — the same thinking behind pricing managed IT per employee works for AI: get the recurring cost on the table before you sign, not after.
Send us the process you have in mind and we will scope it properly, including the parts that are not worth building.