AIPricingBudgetIntegration
AI integration cost: what it takes to add AI to business software, with worked usage numbers
10 min read
Adding AI to business software costs two different things: a one-off build and a monthly usage bill. The build depends on the data you must prepare, the systems you must connect and the controls you must put in place. Usage depends on the model you pick and the number of requests, and it is often far smaller than people fear. What really costs money is making the feature reliable.
Most pages on this topic offer price bands from $10,000 to several hundred thousand with no source. This one prices both halves with model prices checked on 7 November 2026 and two worked examples you can redo with your own volumes. For the use cases themselves, read our guide to adding AI to a business application. For an agent that acts across several systems, our article on custom AI agent costs covers the heaviest case.
Three levels of integration, three budgets
| Level | Example | What you pay for |
|---|---|---|
| Switch on AI in a tool you use | The assistant built into your email suite or your packaged CRM | A per-user monthly add-on set by the vendor |
| Add an AI feature to your software | Reading purchase orders, summarising a case file, drafting a reply | Targeted development, then model usage |
| Build a connected assistant or agent | Answering from internal documentation, preparing actions across systems | A full project: data, permissions, controls, monitoring |
The first level is the cheapest and fastest, but it only sees the data in the tool that hosts it. As soon as AI must work on your own business data, inside your software, under your rules, you are at level two or three.
"ChatGPT integration" vs API integration
A common confusion: a ChatGPT or Claude subscription for your staff and an API integration are two different products. The subscription gives people a chat window, billed per seat. The API lets your own software call a model, billed per token, with no seat fee. When a vendor quotes an "OpenAI integration", they mean development work around the API: connecting it to your data, building the screens, adding controls. The API calls are usually the smallest line.
The build cost lines
| Cost line | What it covers | What pushes it up |
|---|---|---|
| Scoping | Choosing the task and defining a good answer | A vague goal ("add some AI") |
| Data | Gathering, cleaning and chunking documents or records | Scanned files, conflicting versions |
| Connectors | Reading from and writing to your existing systems | Legacy software without a documented API |
| Interface | The screen where users see, correct and approve | Several roles, multi-step approval |
| Test set | Real cases with expected answers, to measure quality | Cases that need a scarce expert to judge |
| Guardrails | Limiting what the model can read and do, blocking abuse | Automatic actions with no human approval |
| Permissions | Showing each user only what they are allowed to see | Fine-grained rights per team or client |
| Logging | Keeping a record of requests, answers and decisions | Strict audit requirements |
Two lines are missing from most quotes: the test set, without which nobody knows whether the feature works, and guardrails. The OWASP Top 10 for LLM applications ranks prompt injection first: a document or email can carry hidden instructions that the model may follow. Any AI feature that reads outside content must be designed with that in mind.
Model prices, checked on 7 November 2026
Providers bill per million tokens (a token is roughly three quarters of an English word), input and output priced separately. US dollars, before discounts:
Two mechanisms lower the bill, both documented on those pages: batch processing, billed at half price by both providers for requests that can wait, and prompt caching of repeated instructions, which Anthropic bills at a fraction of the normal input price when reused. Prices move often, so check them when you build your estimate.
Worked example 1: reading purchase orders
Assumption: each PDF purchase order is about 3,000 input tokens (instructions included) and 500 output tokens (the extracted fields). Your documents will differ, so measure a real sample.
| Model | Cost per document | 5,000 documents a month |
|---|---|---|
| Claude Haiku 5.5 | $0.00055 | $2.75 |
| Claude Sonnet 5.5 | $0.011 | $55 |
| Claude Fable 5.1 | $0.055 | $275 |
The working for Sonnet 5.5: 3,000 × $2 / 1,000,000 = $0.006 for input, plus 500 × $10 / 1,000,000 = $0.005 for output. If orders can wait until the end of the day, batch processing halves those amounts.
Even on the most expensive model, usage is small next to the manual entry time it saves. The real cost sits elsewhere: the connector to your ERP or CRM, the review screen where a person checks uncertain fields, and the test set that proves extraction is reliable. The same pattern applies to "AI in the CRM" features such as summarising an account before a call.
Worked example 2: an assistant on internal documentation
Assumption: 50 users each ask 10 questions per working day, about 10,500 questions a month over 21 days. Each question sends roughly 8,000 tokens (the question plus retrieved document passages) and gets 600 tokens back.
| Model | Cost per question | Per month |
|---|---|---|
| Claude Haiku 5.5 | $0.0011 | $11.55 |
| Claude Sonnet 5.5 | $0.022 | $231 |
| Claude Fable 5.1 | $0.11 | $1,155 |
Here the model choice moves the bill by a factor of a hundred. The right method: build a set of about fifty real questions with their correct answers, test several models, and keep the cheapest one that meets your quality bar. How this kind of assistant works, and its permission pitfalls, is covered in our guide to RAG for business documents.
What it costs after go-live
- Monitoring: tracking answer quality, errors and daily spend, with alerts when something drifts.
- Human review of sensitive cases, which is a staffing cost, not a software cost.
- Model changes: providers retire older models on a schedule, as Anthropic's model deprecations page shows. Each migration means rerunning your test set.
- Usage growth: a feature people like gets used more, and the bill follows. Set a spending cap and an alert.
Compliance: plan it into the budget
Sending data to a model means handing processing to a third party. In the EU and the UK, data protection law applies: send only the data you need, check where it is processed, and put a proper processing agreement in place. The UK regulator's guidance on AI and data protection is a practical starting point. If you serve EU users, the EU AI Act adds transparency duties or stricter requirements depending on the use case; our overview of the EU AI Act for business software helps place your project. These points are cheap at scoping and expensive after launch.
Ways to pay for it
The build can be priced like any software work: a fixed price on a scope made visible upfront, or time and materials, where you carry any overrun. Usage is better passed through at cost, with a monthly report, so you see exactly what each feature consumes. A third option folds everything into one monthly subscription covering build, hosting, monitoring and model usage up to an agreed volume; in that case, rights to the code are defined case by case in the contract.
Mistakes that cost money
- Starting from the technology ("we need a chatbot") instead of a precise, measurable task done every day.
- Defaulting to the most powerful model without a side-by-side test.
- Letting AI act alone on decisions that commit the company: prices, payments, contractual replies.
- Skipping the correction screen: a person must be able to see, correct and approve, and their corrections feed the test set.
- Leaving usage spend uncapped.
- Training a model when passing the right documents with each request would have done the job.
A checklist for reading an AI integration quote
- Is the task described precisely, including what a good answer is?
- Is there a test set, and who supplies the real cases?
- Is monthly usage estimated, with volume assumptions stated?
- Can the model be swapped without rebuilding the application?
- Are guardrails against prompt injection described?
- Does the AI respect the same permissions as the application?
- Do monitoring, maintenance and model migrations have their own line?
- Where is data processed, and under which agreement?
Where to start
- Pick one task that is repetitive, time-consuming and easy to check.
- Collect twenty to fifty real cases with their correct outcome.
- Estimate usage with public prices and your own volumes, as above.
- See the feature before you fund it.
We build a free visual prototype of the feature and present it on a video call within 5 working days: you see where AI steps in and what the user approves, and the quote rests on that scope. The approach is explained in see your software before you pay.