Anthropic
Integration · AI models
Building business agents on Anthropic’s Claude models
Anthropic is a model provider: its Claude models, used through the API, do the reading, reasoning and writing inside an agent while your own system handles tools, permissions and records. Choose a model tier per task, understand the retention terms and measure cost per completed job.
What Anthropic provides, and what you build
Anthropic’s API gives your software access to the Claude family of models. It does not include your business logic. The agent we build decides which tools Claude may call, such as “look up an order” or “draft an invoice”, checks each action against your rules and records what happened. Claude decides how to use those tools to finish the task.
For a business, that might be an agent that reads an incoming purchase order, checks stock and pricing through tools connected to your systems, and prepares a quote for a person to approve. See quote generation and AI research assistant for two common patterns.
We use Claude heavily ourselves. The lessons on Find Tutor Web were written, illustrated, simulated and test-marked by Claude agent pipelines, which scripted each lesson, built it and audited it against quality gates before publishing. The audit stage mattered as much as the writing: a pipeline that checks its own output is what made 921 lessons possible without a content team.
Choosing the right Claude model for each job
Anthropic offers Claude models in several tiers, from small, fast and inexpensive to large and more capable. Names and versions change, so think in tiers and check the current line-up when you build.
- Break the workflow into steps: read the input, decide, call tools, write the result, check it.
- For each step, collect real examples with agreed correct outputs.
- Start with the smallest model and move up only for steps where it fails your examples.
- Keep the larger models for steps that need judgement across long documents or many tool calls.
- Re-test when a new model version arrives, since a cheaper model may now pass where it failed before, and pin the version you use in production.
This is the same discipline described in how to scope an AI agent, applied to model choice.
How does Anthropic handle API data?
At the time of writing, Anthropic states that by default it does not use inputs or outputs from its commercial products, including the API, to train its models, unless you report feedback or otherwise choose to allow it. Its default is to delete API inputs and outputs from its back end within 30 days, with exceptions: content flagged for usage policy violations may be kept longer, and data you store with features such as file storage is kept until you delete it. Zero data retention is available by agreement with Anthropic for eligible products.
Before you connect Anthropic: a checklist.
Connecting Anthropic safely
0 of 8 done
Keeping cost per job under control
Usage is billed per token, with input and output priced differently and larger models costing more. Agents that call tools in a loop can use far more tokens than a single question, so set limits on the number of steps per task and alert when a task exceeds its normal cost.
- Prompt caching reduces the cost of resending the same long instructions or documents on every call.
- Batch processing suits work that can wait, such as overnight document runs.
- Smaller models for routine steps, as above.
- Shorter context: send the relevant records, not the whole history.
On User Evaluation’s research pipeline, which uses both Anthropic and OpenAI models, matching models to tasks, shortening prompts and caching repeated retrieval cut per-job cost by 38% with no regression on their internal evals, and made median synthesis 2.1 times faster. See the AI agent cost guide for budgeting the rest.
Running Claude alongside other providers
There are good reasons to keep more than one provider available: a fallback during outages or rate limits, the best model for each task, and price changes. The cost is a little more engineering, because prompts and tool definitions need testing on each model. Our open-source FreeLLMAPI router shows the routing side of this: requests go to the best available model and move to the next route when a provider rate-limits, times out or fails. For business agents we usually route per step, with a tested fallback, rather than switching providers at random. See OpenAI for the same decisions on that side.
Keep the Anthropic account and API keys in your organisation, with separate keys per application and spending limits set, so access can be revoked cleanly.
Common questions
Is Claude better than GPT models for business agents?
Neither is better for every task, and rankings change with each release. Claude models are strong at following detailed instructions, working with long documents and using tools, but so are their competitors in different ways. Test both on your own examples for each step and choose on quality, cost and speed rather than reputation.
Can Claude take actions in our systems?
Only through tools your developers define. Claude can request a tool call, such as “create a draft invoice”, and your system decides whether to run it, checks permissions and records the result. Anything with financial, legal or customer impact should go through an approval step. See our guide to human-in-the-loop AI for how to set that up.
Is Anthropic available through cloud platforms as well?
Yes. Claude models are offered through Anthropic’s own API and through major cloud platforms, including Amazon Bedrock and Google Cloud Vertex AI. Using a cloud platform can simplify billing and keep data handling under an agreement you already have, though model availability, features and regions can differ, so compare them before you choose.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.