How much does AI development cost?
Published ranges for AI projects are close to useless because they lump a weekend prototype together with a year-long platform. Here are the numbers we actually quote, and what moves them.
The short answer
A grounded chatbot answering from your documentation runs $4,000–12,000. A production RAG knowledge base with permissions and evaluation runs $10,000–30,000. A single-purpose agent with tool access runs $12,000–35,000. Document processing pipelines run $8,000–25,000 depending on document variety.
What moves those numbers is almost never the AI. It is integration count, data quality, permission complexity, and how many stakeholders must approve a decision. A project touching one clean system costs a fraction of the same feature touching four messy ones.
- Integration count is the biggest cost driver
- Poor data quality can double a timeline
- Permission and compliance requirements add weeks
- Evaluation is 10–20% of budget and worth every cent
- Running costs are usually tens to hundreds per month
- The cheapest project is the one correctly scoped down
Why quotes for the same brief differ by 10x
Ask five suppliers to build a support assistant and you will get numbers between four thousand and eighty. That spread is not mostly greed. It is that each of them scoped a different project, and the brief did not say which one.
The cheap quote usually assumes a wrapper around a model API with your FAQ pasted into a prompt. It will demo well. The expensive one assumes retrieval over your real documentation, permission handling, an evaluation suite, monitoring, and a plan for the day the model provider changes something. Both are honest answers to different questions.
The fix is to specify the parts that drive cost before asking for a price: how many systems must be integrated, what accuracy bar the output has to clear, who reviews wrong answers, and what happens at peak load. A brief containing those four things gets quotes within a comparable range.
The costs that appear after launch
Model usage is the obvious one, and usually smaller than people fear — most business workloads run in the tens to low hundreds of dollars a month. What surprises people is everything around it.
Re-indexing when your documents change. Re-running evaluations when a provider ships a new model version, which they do without asking. Someone reviewing the cases the system flagged as uncertain. Hosting, monitoring and the vector store. And the periodic work of looking at what users actually asked, which is where most of the improvement in year two comes from.
Budget somewhere between fifteen and twenty-five per cent of the build cost annually to keep an AI feature working properly. A system nobody maintains degrades quietly, because the world it was measured against moves and it does not.
Where the money actually goes
On a typical production feature, the model integration is rarely more than a fifth of the work. Getting at the data and cleaning it is often a third. Evaluation, monitoring and the interface take most of the rest.
This is why 'we already have the AI part working' is a misleading position to negotiate from. A notebook that produces good answers on ten hand-picked examples has completed the cheapest fifth of the project.
Frequently asked questions
Why do quotes vary so much between developers?
Usually because they are quoting different things. One is quoting a prototype, another a production system with evaluation, monitoring and hand-over. Ask specifically what happens when the AI is wrong — the answer tells you which you are being quoted.
What are the ongoing costs?
Model usage, hosting and any vector database. For most business applications that totals tens to a few hundred dollars a month. High-volume consumer products are a different conversation and need unit economics modelled before launch.
Is it cheaper to use an off-the-shelf tool?
Very often, yes, and we will tell you when. Custom work earns its cost when the workflow is specific to you, the data cannot leave your systems, or the capability is the product itself.
How can we reduce the cost?
Cut integrations, accept human review on edge cases rather than automating them, and start with one workflow instead of five. Scope is the only real lever.
Tell us what you are building.
Send a short description of the problem and we will reply within one business day with an honest view of scope, cost and whether we are the right person for it.
Or email directly: contact@hire-ai-dev.com