Many AI budgets start with the licence or the API price and stop there. That leaves out the build, the data preparation, the hosting, the monitoring, the upkeep and the staff time. None of these are hard to estimate once you name them.
We build AI integrations through our AI integration service, so weigh that. Every fact below about Anthropic and Cloudflare was checked against their own documentation in October 2026.
Short answer: There’s no honest generic price for “an AI integration”, because the cost depends on your task, data and systems. Budget it as seven lines: build effort, API usage, hosting, monitoring, maintenance, model migrations and staff time. API usage is the easiest to estimate, because Anthropic publishes per-token prices: in our example below, classifying 3,000 support tickets a month on Claude Sonnet 5.5 costs about $15 in API fees. The build and the people around it usually cost far more. Ask for each line separately, in a written scope.
What the Vendor Docs Say About Running Costs
As of October 2026, from Anthropic and Cloudflare documentation:
| Item | What the documentation says |
|---|---|
| Claude Sonnet 5.5 | $2 per million input tokens, $10 per million output |
| Claude Haiku 5.5 | $0.10 input, $0.50 output per million, for prompts up to 100,000 tokens |
| Batch API | 50% off input and output; most batches finish within an hour |
| Prompt caching | Cache reads cost 0.05x base input on Sonnet 5.5 and Opus 5.5 |
| Tokenizer | Claude 4.7 and later models produce about 30% more tokens for the same text |
| Token counting | Free to use, with its own rate limits |
| Spend caps | Monthly cap per tier; you can set a lower limit, also per workspace |
| Model retirement | At least 60 days’ notice for publicly released models |
| Workers Paid | $5 a month minimum, 10 million requests included |
Sources: Anthropic pricing, Batch processing, Token counting, Rate and spend limits, Workspaces, Model deprecations, Cloudflare Workers pricing.
Why Generic Budget Ranges Mislead
Online cost ranges rarely describe your project. A range built from other companies’ builds can’t know whether your CRM has an API, how clean your data is or who will review the output. Treat any range as a prompt for questions, not as a budget.
Small Business: Sizing Your First Integration
For a small business, the best first budget covers one workflow: processing inbound documents, answering support questions from your knowledge base, or drafting product descriptions for a WooCommerce catalogue. One workflow gives you one input, one output and one thing to measure.
Size it from the task, not from your revenue. Write down how many items run each month, how long a person takes on each and what a wrong answer costs. That’s your scope and your baseline. If you can’t describe the task that specifically yet, scope it first.
Pilot vs. Production Cost Realities
A pilot is a scoped test on one workflow, in staging, with real users judging real output. Production is the same integration running daily with monitoring, error handling, logging and trained staff. The second usually needs work the first skipped.
Ask for them as two separate budget lines. A proposal that prices the pilot and calls it the total leaves out the part you’ll live with.
Where the Money Actually Goes
Split the budget into one-off and running costs. One-off: scoping, data preparation, the build and testing. Running: API usage, hosting, monitoring, maintenance and staff time. The running costs get forgotten, because they arrive after sign-off.
Software Licenses Are Only Part of the Number
The subscription or API price is the line you can see. Connecting it to your CRM, cleaning the inputs and managing the outputs is where most of the effort goes, and none of that is on a pricing page. Budget every line.
Line Items That Often Go Unlisted
The costs that surprise businesses most:
- Data preparation: Your data is probably not in the format the AI needs. Cleaning and structuring it is labour, sometimes more than the integration itself.
- Integration work: Connecting AI to your WordPress site, CRM or order system takes development time, even when a vendor says it “connects easily”.
- Staff training: when to trust the output, when to flag it and what to do when it fails.
- Ongoing support and monitoring: Your data, your customers and the models all change. Someone has to own that.
A Worked API Usage Example
This is our own example, with assumed numbers. A support inbox gets 3,000 tickets a month. Each call sends 1,500 input tokens (1,000 of instructions, 500 of ticket) and gets back 200 output tokens: a category and a short reason.
- Input: 3,000 × 1,500 = 4.5 million tokens. At $2 per million on Claude Sonnet 5.5, that’s $9.
- Output: 3,000 × 200 = 600,000 tokens. At $10 per million, that’s $6.
- Total: about $15 a month (source).
Three levers change that. If tickets don’t need an instant answer, the Batch API charges 50% of standard prices: $7.50 (source). Claude Haiku 5.5 brings the same workload to about $0.75, if it handles your tickets well enough (source). Prompt caching cuts the cost of the shared instructions: on Sonnet 5.5 a cache read costs $0.10 per million instead of $2, and a 5-minute cache write costs 1.25x the input price (source). The cache refreshes each time it’s used, so savings depend on how your tickets arrive (source).
Don’t guess token counts. Anthropic’s token counting endpoint is free (source). Allow for retries and test runs, and check volumes against rate limits.
Hosting and Monitoring
The code that calls the model has to run somewhere. As one example, Cloudflare’s Workers Paid plan has a $5 monthly minimum, includes 10 million requests a month and charges $0.30 per additional million, with no charge for duration (source). Other hosts price differently.
Logging is a separate line: on that plan, Workers Logs includes 20 million log events a month (source). The bigger monitoring cost is usually a person checking the output is still right.
You can also cap API spend. Each Anthropic usage tier has a monthly spend cap, and you can set a lower limit (source). Workspaces add spend limits, alerts and separate cost tracking per project (source).
Maintenance and Model Migrations
Models retire. Anthropic gives at least 60 days’ notice before retiring a publicly released model, and requests to a retired model fail (source). Budget time to test the replacement on real inputs.
A migration can change running costs too. Claude 4.7 and later models produce about 30% more tokens for the same text (source), and Anthropic advises recounting prompts with the new model (source).
Why AI Budget Forecasts Go Wrong
Forecasts go wrong when they’re built from a demo, which shows ideal conditions instead of your systems and data.
The Misestimation Problem
Scope is often set at the demo stage, not the scoping stage. Legacy systems, data quality and internal workflows all create friction, and friction costs time. Start with a technical scoping exercise, not a product demo. If a vendor won’t scope before you sign, treat that as a signal.
Legacy System Integration Adds Cost Before You Write a Prompt
An old CRM, a custom ERP or a booking system with no webhooks may not have been built for API access. Connecting to it can take more work than the AI part.
Before you commit, ask for a systems audit listing every connection point, how each will be integrated and an effort estimate for each. If that document doesn’t exist before the contract is signed, the contract number is a guess.
How to Build a Realistic AI Budget for 2026
Start with the task, not the tool. “Classify inbound support tickets and route them to the right person” can be scoped, measured and tested against a break-even threshold. “Use AI to improve operations” can’t.
The Break-Even Test: Set Your Own Threshold
Add up the one-off costs. Divide by the monthly saving minus the monthly running costs. That’s your months to break even. Use measured baselines, not estimates.
Decide in advance how many months you’d accept. If the result is longer, build the case for why it’s still worth doing.
What a Fixed Tech Budget Actually Buys You in AI Tooling
With a fixed technology budget, one integration done properly usually beats several started at once. Production needs support, monitoring and an owner. A pilot needs testers. If someone proposes parallel AI projects, ask for the plan for each: they compete for the same people.
The Contingency Line
Add a contingency line sized to what you don’t know yet: unknown data quality, an untested system connection, a review process nobody has tried. More unknowns, bigger line. Build it in before sign-off, not after the first change request.
When AI Integration Is Not Worth the Budget
Some tasks don’t need AI. If volume is low, the rules are simple or a wrong answer is costly and hard to catch, a script or a person may do better.
Signs a Vendor Pitch Doesn’t Match Your Actual Workflow
The clearest signs:
- The demo uses clean, formatted sample data. Yours is not.
- The vendor can’t name the integration method for your systems.
- The ROI claim uses industry benchmarks, not your volumes.
- There’s no mention of what happens when the output is wrong.
- The proposal has no data preparation phase.
Any of these should trigger a pause, not a signature.
Questions to Ask Before Signing Any AI Integration Contract
- What is the exact integration method for our stack, and has it been tested against our system version?
- What does data preparation involve, and is it in the quoted cost?
- What do monitoring and error handling look like after launch, and who pays for them?
- What will monthly API and hosting costs be at our volumes, and how were they calculated?
- How will we measure whether it works, and what happens if it doesn’t?
Vague answers here show up as budget lines later.
Related Posts
- True cost of AI integration: what the build quote leaves out
- AI model updates and your integration: planning for model changes
- Custom AI tool maintenance: what ownership takes after launch
- AI automation cost vs. manual: the break-even arithmetic
Frequently Asked Questions
How much should a small business budget for AI integration in 2026?
It depends on the task, your data and your systems, so we don’t quote a standard figure. Budget seven lines: build effort, API usage, hosting, monitoring, maintenance, model migrations and staff time. API usage is often the smallest: in our example, 3,000 tickets a month on Claude Sonnet 5.5 cost about $15 at Anthropic’s published prices.
What are the hidden costs of AI implementation?
Data preparation, connecting older systems, staff training, monitoring and model migrations. Anthropic gives at least 60 days’ notice before retiring a publicly released model, and a newer model can use more tokens for the same text, so each migration needs testing and a fresh estimate.
How long does it take to see ROI from AI integration?
Divide the one-off costs by the monthly saving minus running costs, using measured baselines. That gives months to break even for your task. Decide in advance what payback you’d accept.
Is there a reliable AI project failure rate?
We don’t have a figure we can stand behind, and a global percentage says little about your project. The causes are more useful: scope set at the demo stage, underestimated integration work and poor data. All three can be checked before you spend.
Is it better to use off-the-shelf AI tools or custom AI integration for an SMB?
Off-the-shelf tools start faster and cost less upfront, but you pay a subscription, depend on the vendor and often still need integration work. A custom integration costs more upfront and gives you ownership of the workflow. Compare both for your task.
What questions should I ask an AI agency before signing a contract?
Ask for the integration method for your systems, whether data preparation is in scope, what monthly API and hosting costs to expect and how they were calculated, and how performance will be measured. Vague answers tend to become overruns.