On this page
“How much will it cost?” is usually the second question we hear about any AI project, right after “Can AI do this?”. The honest answer is that the AI model itself is rarely the biggest line in the budget. Most of the cost sits in the work around it: connecting systems, preparing data, testing, training people and keeping the system healthy once it is live.
This guide sets out what a realistic AI budget includes, so you can plan, compare proposals and avoid surprises.
Two budgets, not one
Every AI implementation has two kinds of cost.
- Build cost: the one-off effort to design, build, test and launch the solution.
- Run cost: the recurring cost to operate, monitor and improve it.
Proposals that only quote the first number are incomplete. A system that is not maintained will drift in quality as your data, processes and the underlying models change.
What goes into the build cost
Discovery and design
Mapping the current process, agreeing the success metric, choosing the right approach and getting security sign-off. On a well-run project this is a short phase, and skipping it is the most expensive mistake you can make.
Integration
Connecting the AI to the systems where data lives and where work happens: CRM, help desk, ERP, document storage, email and chat. This is usually the largest part of the build. Every system has its own access rules, data formats and limits.
Data preparation
AI is only as good as the information it receives. Duplicate CRM records, outdated documents and inconsistent fields all reduce quality. Some cleanup is almost always needed, and it often improves other processes too.
Testing and quality assurance
Building a set of real test cases, measuring accuracy, and adjusting until the agreed quality bar is met. This is what separates a demo from a dependable system. We describe the approach in how to test an AI automation.
Change management
Training the team, adjusting the process, and setting up review steps. Without this, even an excellent system goes unused.
What goes into the run cost
Model usage
AI providers usually charge by usage, based on how much text the model reads and writes. Costs scale with volume: a process that runs a few hundred times a month costs very little, while one that runs on every inbound email adds up.
Infrastructure and tools
Hosting, workflow tools, databases, and any third-party data services the solution uses, such as enrichment providers.
Monitoring and support
Someone needs to watch quality, handle exceptions, respond to errors and keep integrations working when connected systems change.
Continuous improvement
Processes evolve, new edge cases appear and better models are released. Budgeting a small amount of time each month for improvements keeps the system valuable rather than slowly degrading.
How to keep running costs under control
Good engineering choices can reduce running costs considerably without reducing quality:
- Use the right model for each step. Simple classification or extraction often works well with smaller, cheaper models. Reserve the most capable models for the hardest steps.
- Avoid unnecessary calls. Many processes run AI on every record when only some need it. Simple rules can filter first.
- Reuse results. If the same information is requested repeatedly, store and reuse the answer instead of generating it again.
- Set spending limits. Daily and per-task limits, with alerts, prevent a bug or a sudden spike from producing an unexpected bill.
A test set makes these decisions safe: you can try a cheaper model or a simpler design and see immediately whether quality holds.
Linking cost to return
The most useful budget conversations start from the business metric, not the technology. For example:
- Hours of manual work saved per month, multiplied by the loaded cost of that time.
- Faster response to inbound leads, and the resulting change in conversion. See speed to lead.
- Fewer errors in data entry or document processing, and the cost of correcting them today.
- More accounts covered per sales rep, with the same or better quality.
When the expected return is clear, the build and run costs can be judged properly, and the project has a clear yardstick for success.
Red flags in AI proposals
When you compare proposals from vendors or partners, be cautious if:
- The proposal covers the build but says nothing about running and maintaining the system.
- There is no mention of testing, quality measurement or how errors are handled.
- Integration with your existing systems is described as “simple” without detail.
- Success is defined vaguely, without a metric tied to your business.
- Pricing depends on the number of users when the value depends on volume of work, or the other way round, without explanation.
Our guide on how to choose an AI implementation partner has a full list of questions to ask.
We give clients a clear build and run budget before any work starts, tied to a business metric. Book a call to discuss your project.


