Ask ten agencies what an AI project costs and you’ll get ten different non-answers. That’s not evasion — AI and ML work genuinely doesn’t compress into a single number the way a website or an MVP does, because the biggest cost driver is something you don’t know yet: whether the model is actually good enough for the job.
Here’s a straight explanation of why, and what a sensible process actually costs.
Why there’s no rate card
A booking form and an AI-assisted document review pipeline can both be called “an AI feature,” and their costs can differ by a factor of ten. What decides it isn’t lines of code — it’s:
- Whether the underlying task is even a good fit for AI. Some workflows look automatable and aren’t, once failure cost and edge cases are accounted for.
- Data readiness. Clean, accessible, well-labelled data is cheap to build on. Data scattered across systems with no clear ownership adds discovery and cleanup work before any model gets touched.
- Acceptable failure cost. A tool that drafts an internal summary can tolerate more error than one that makes a compliance or financial decision. Higher failure cost means more evaluation, human review, and guardrail work.
- Integration depth. A standalone assistant is simpler than a model embedded in an existing workflow with permissions, audit trails, and legacy system connections.
Because of this, credible AI providers scope in two stages rather than quoting a fixed project price upfront.
Stage one: validation
Before committing serious budget, the sane first step is a bounded validation phase — similar in shape and cost to a standalone discovery engagement, not an open-ended research project. It should answer, on a fixed budget and timeline:
- Is this task actually a good fit for AI, given the failure cost involved?
- Is the data available and clean enough to build on?
- What would a task-specific evaluation set and acceptable failure boundary look like?
If the answer is no, you’ve spent a bounded, small amount to avoid an open-ended one. That’s the point of treating validation as its own priced step rather than folding it into a full build quote nobody can actually stand behind.
Stage two: build and integrate
Once validated, the build phase is scoped against the real inputs: which model or provider, how much retrieval and grounding infrastructure is needed, how deep the integration goes, and how much human review the workflow requires in production.
As a rough market orientation — not a quote — narrow, well-scoped AI features (a grounded search or summarisation tool over existing content) tend to sit in the low tens of thousands. Workflow automation with multiple integrations, evaluation infrastructure, and human-in-the-loop review climbs from there, and regulated or high-failure-cost use cases climb further still because of the validation and control work involved.
What should be included regardless of size
- Use-case validation before build commitment, not after.
- An evaluation set and defined failure boundaries, agreed before anyone calls a demo “production-ready.”
- Human controls and fallbacks for outputs below the quality threshold, not just a raw model call.
- Cost and quality monitoring once it’s live, since model behaviour and usage patterns both drift over time.
A quote missing the first two items is usually cheaper for a reason — it’s pricing a demo, not a production feature.
Questions worth asking before you commit budget
- What specifically will the validation phase tell us, and what happens if the answer is “this isn’t a good fit”?
- What’s the evaluation set, and who defines what “good enough” means?
- Can this run inside our own cloud and data boundary, using our access controls?
- What’s the fallback when the model gets it wrong in production?
- How is ongoing cost monitored once real usage starts?
If you have a workflow you think AI could improve, book a strategy call and we’ll tell you honestly whether it’s a good validation candidate — see the full approach on the AI and Automation service page.
