Short answer: AI integration costs from about $3,000 (roughly £2,400) for one feature added to a system you already run, $15,000 to $45,000 (roughly £12,000 to £36,000) for a production deployment wired into two or three business systems, and $60,000 upwards for multi-model work across a whole operation. Most companies should start at the low end, because the return comes from picking the right process, not from the size of the model.
You almost certainly do not need an AI product. You have a CRM, a helpdesk, an accounts package, maybe a custom application that runs your operation. The question is not whether to buy something new. It is whether AI can be put inside what you already run, without a rebuild and without breaking the things that currently work.
That is what AI integration services are for. Below is what the work actually involves, what it costs to build and to run, where it tends to fail, and how to tell whether your first project is worth starting.
What AI integration actually means
The phrase covers three quite different levels of work, and the price difference between them is large.
Level one: switching on a feature you already pay for
Your existing tools ship AI features. HubSpot, Salesforce, Zendesk, Microsoft 365, Google Workspace and most modern platforms now include summarisation, drafting and basic classification. Turning these on, configuring them, and training the team costs you a few days of consulting and whatever the vendor charges per seat. If nobody has audited what your current stack already does, start here, because it is the cheapest AI you will ever deploy.
Level two: connecting a model to your systems
This is the bulk of real AI integration work. A language model or a purpose-built service sits between your systems and does a specific job: reading inbound emails and routing them, extracting fields from invoices into your accounts package, drafting replies inside your helpdesk, qualifying and enriching leads before they reach a salesperson. The model is rented from a provider. The value you are paying for is the plumbing, the prompts, the guardrails and the testing.
Level three: a custom AI layer
A retrieval system over your own documents, a fine-tuned or heavily evaluated model for a domain-specific task, or several models orchestrated into a workflow with human review built in. This is custom software development that happens to use AI, and it is priced accordingly.
Most businesses asking about AI integration need level two. A minority genuinely need level three. Almost everyone skips level one, which is a mistake.
The five jobs that pay back first
After a lot of these builds, the same handful of use cases keep producing a return, and the same handful keep disappointing people.
- Document and data extraction. Invoices, purchase orders, delivery notes, applications, ID documents. The work is repetitive, the input is structured enough to check, and the output goes straight into a system. This is the single most reliable payback in the list.
- Inbound triage and routing. Reading an email, a form or a ticket, deciding what it is about and how urgent it is, then routing it and drafting a first reply for a human to approve. Fast to build, easy to measure.
- Lead qualification and enrichment. Scoring an enquiry, pulling in company information, and writing a short brief so the salesperson opens the record already knowing who they are talking to. Pairs naturally with CRM work.
- Internal knowledge search. A system that answers staff questions from your own policies, contracts and procedure documents, with citations back to the source. Valuable where staff currently interrupt each other to find answers.
- Reporting and first drafts. Turning raw numbers into a written summary a manager can read in a minute, or drafting the routine documents your team rewrites weekly.
What tends to disappoint: fully autonomous customer-facing chat with no human in the loop, anything that needs to be right one hundred percent of the time with no way to check it, and predictive work on data you do not actually have enough of.
What AI integration costs
| Scope | What it includes | Typical cost (USD) | Typical cost (GBP) | Timeline |
|---|---|---|---|---|
| Audit and enablement | Review of the stack, switching on existing AI features, prompts and team training | $1,500 to $4,000 | £1,200 to £3,200 | 1 to 2 weeks |
| Single feature | One job, one system, one model, with testing and a fallback path | $3,000 to $12,000 | £2,400 to £9,600 | 2 to 4 weeks |
| Production deployment | Two or three systems connected, guardrails, evaluation, monitoring, admin controls | $15,000 to $45,000 | £12,000 to £36,000 | 6 to 12 weeks |
| Custom AI layer | Retrieval over your own data, multi-step workflows, human review, role-based access | $45,000 to $120,000 | £36,000 to £96,000 | 3 to 6 months |
These are industry-typical ranges for 2026. An agency with design, engineering and evaluation in-house sits at the upper end of each band. A single contractor sits at the lower end, and usually leaves the monitoring and evaluation work undone, which is where these projects tend to fail six months later.
What the build actually involves
The model is the least interesting part. When we quote an AI integration, most of the estimate is the following.
Access to your data. Reading from and writing to your CRM, database, document store or line-of-business system, with credentials handled properly and permissions respected. If your data lives in five places and two of them have no API, this is where the budget goes.
The prompt and model layer. Choosing a model for the job, writing and versioning the instructions, handling context limits, and deciding what happens when the model returns something unexpected.
Guardrails. Rules about what the system is allowed to do on its own and what needs a human. Any action that sends a message, moves money, changes a record or talks to a customer should have an approval step until it has earned the right not to.
Evaluation. A test set of real examples with known correct answers, run every time the prompt or the model changes. Without this you have no way of knowing whether last week's tweak made the system better or quietly worse. Skipping evaluation is the most common reason an AI feature that demoed beautifully is switched off within a quarter.
Monitoring and cost control. Logging what went in, what came out, what it cost and how long it took, with alerts when any of those move.
Fallback behaviour. What the system does when the provider is down, rate limits you, or returns nonsense. It should degrade to the old manual path, not break the process.
The running costs people forget
AI integration is not a one-off purchase. Budget for the following every month.
| Item | Typical monthly cost |
|---|---|
| Model usage (provider tokens) | $50 to $2,000 depending on volume |
| Vector or search infrastructure, where used | $30 to $400 |
| Monitoring and logging | $20 to $200 |
| Re-evaluation and prompt maintenance | 2 to 8 hours of engineering |
The usage figure is the one that surprises people. It scales with how much text goes in and out, not with how many users you have, so a single document-heavy workflow can cost more than a chat feature used by the whole company. Any partner quoting an AI build should give you a modelled monthly running cost before you start, and should design the system to fail cheaply rather than expensively.
Where AI integration goes wrong
Automating a broken process. If the underlying process is unclear, AI will do the wrong thing faster and more confidently. Fix the process first. Our guide on how to automate your business covers how to map one properly.
No human in the loop on the first release. Put a person between the model and the customer for the first few weeks, watch what it gets wrong, then remove the checkpoint where the evidence supports it.
Sending data you should not send. Know which provider processes your data, where it is processed, whether it is retained, and whether your customer contracts and regulator allow it. Sort this before the build, not after.
No definition of success. Decide the number that has to move: minutes per ticket, hours per week of data entry, response time on new enquiries. If you cannot name it before you start, you will not be able to tell whether the project worked.
Buying capability nobody asked for. The winning projects come from watching what a team repeats every day, not from a list of things AI can theoretically do.
Build, buy, or switch on what you own
Buy when a vendor already solves your exact problem for a per-seat fee and you have no unusual requirements. It will be cheaper than anything custom, and it will keep improving without you paying for it.
Switch on what you own when your existing platforms have the feature and nobody has configured it. Cheapest option available, and frequently the right first move.
Build when the job depends on your own data, your own process or your own systems, when the vendor options force you to change how you work, or when the process is valuable enough that owning it matters. That is the point where an AI and machine learning engagement earns its cost.
How to choose a partner
Ask for a worked example of something they put into production and what it cost to run, not a demo. Ask how they evaluate quality, and expect to hear about a test set rather than a general reassurance. Ask what happens when the model is wrong, and listen for whether a human is in the path. Ask who owns the prompts, the code and the data, and confirm the answer is you. Ask for the modelled monthly running cost alongside the build price.
Be suspicious of anyone who leads with the model name rather than your process, promises full autonomy on the first release, or cannot tell you what the thing will cost to operate.
Bottom line
AI integration is worth doing when it sits inside a process you already understand, on data you already have, with a human checkpoint until it earns its independence. Start with one job that is repetitive, measurable and low risk. A first feature at $3,000 to $12,000 that saves a few hours a week will teach you more about where AI belongs in your business than a six-figure programme designed in a boardroom.
Book a free call and we will look at your current stack, name the one process worth automating first, and give you a fixed price to build it along with what it will cost to run.