What you will be able to do
- Break an existing process into individual tasks before deciding what Claude should do in it
- Tell good candidates for Claude (high-volume, repetitive, text-heavy, well-specified) apart from steps that should stay with a person
- Choose between augmenting a process by automating some of its parts and redesigning it into a structured multi-step workflow, starting with the simplest option that works
Key concept
Decompose before you delegate — You don't bring Claude into a process all at once. First break the process into its individual tasks, then decide task by task what Claude does and what stays with a person.
1.Start from the process you already run
Most organisations don't bring Claude into an empty space. They bring it into something that already works, even if it works slowly: a support queue, a document review, a month-end checklist. The first step is not to pick a tool. It is to describe the process as it should run. Anthropic's customer support guide starts by asking you to outline an ideal interaction, because that outline is what shows you what the solution needs to do.
Next, you break that outline apart. A single customer conversation looks like one job, but the guide points out that it bundles several different kinds of work. Some are informational, like answering questions. Some are retrieval, like looking things up in a knowledge base. Some are actions, like getting a quote from an internal system. Each kind has a different risk profile and needs its own evaluation. Splitting the process this way is what lets you decide one piece at a time instead of saying yes or no to the whole thing.
| Task group | What it includes |
|---|---|
| Greeting and general guidance | Greet the customer; give general information about the company |
| Product information | Explain electric vehicle coverage, answer follow-up questions, offer links to sources |
| Conversation management | Stay on topic; redirect off-topic questions |
| Quote generation | Ask eligibility questions, submit the details to a quote API, present the quote |
Sources1
2.Good candidates and steps that stay with a person
Once the tasks are listed, some are clearly better suited to Claude than others. The support guide lists indicators that point to automating parts of a process. They share a pattern: lots of similar requests, a lot of text to search and combine, demand that spikes or never stops, and a need for a consistent voice. The guide frames this as automating portions of the process, not all of it. Claude takes the repetitive volume so that people have more time for the complex cases.
The same sources also show which steps should not go to Claude. Anthropic's commerce blueprint lets its shopping agent search, compare, and fill a cart, but it gives the agent no way to buy anything: the backend interface has no method that places an order or moves money, and the customer confirms checkout in the app. Those are irreversible steps, so they stay with a person. The legal summarization guide makes the same point about accountability. Errors in AI summaries can create legal liability, so legal professionals remain the reviewers.
| Step looks like… | Where it goes | Why |
|---|---|---|
| A large volume of similar questions | Claude | Handled efficiently, which frees people for complex issues |
| Retrieving and combining information from large knowledge bases | Claude | Faster than people researching several sources |
| Sudden spikes in volume | Claude | Absorbed without hiring and training extra staff |
| Placing an order or moving money | Person | Irreversible; the customer confirms it in the app |
| Signing off on output that carries legal liability | Person | Needs review by an accountable professional |
A bank wants Claude to prepare a recommendation for each loan application, but a compliance officer must still approve or reject every decision before it takes effect. Which integration pattern does this describe?
Correct answer: A — A human-in-the-loop pattern where Claude assists with a recommendation and a person retains final decision authority
- A. Correct. Requiring a compliance officer to approve or reject every Claude-generated recommendation before it takes effect is a human-in-the-loop integration, where Claude augments the decision process but a person retains authority.
- B. Incorrect. The scenario explicitly keeps a human approval step, so the decision is not finalized autonomously by Claude.
- C. Incorrect. The scenario does not describe a weekly bulk process; it describes ongoing review of individual recommendations.
- D. Incorrect. Nothing in the scenario has Claude writing directly to the ledger; the compliance officer's approval step sits between the recommendation and any action.
3.Augment the process, or redesign it?
After you know which tasks Claude should do, you decide how much of the process to change. Augmenting keeps the existing process and has Claude do some of its steps better or faster. Redesigning changes the structure itself, for example by splitting one step into a chain of smaller ones, or by adding a sorting step at the front that sends each kind of request down its own path.
Anthropic's engineering guidance favours restraint: use the simplest solution that works and add complexity only when it's needed. Often a single well-prompted call to Claude with the right reference material is enough. If more structure is justified, predefined workflows give predictability and consistency for well-defined tasks. That makes them a good fit when the steps of a process are fixed, for example because each one is required by regulation.
| Pattern | What it changes | Fits when |
|---|---|---|
| Single optimized LLM call | Nothing structural; Claude does one step | Retrieval and in-context examples are enough |
| Prompt chaining | One step becomes a sequence, with checks between steps | The task splits cleanly into fixed subtasks |
| Routing | Inputs are classified and sent to specialised paths | There are distinct categories that are better handled separately |
Augmenting. The required steps stay where they are, and Claude speeds up the work inside them, such as drafting, checking and summarising. A fixed, predefined workflow is the natural match for steps that can't move. Removing or reordering the required steps would not be a valid redesign.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Integrating Claude means deciding whether Claude should take over the whole process.Why is that wrong?
A process bundles several different kinds of task. You break it into those tasks and decide for each one, so part of the process can go to Claude while other parts stay with people.
Covered in Start from the process you already run
2.The most autonomous, most complex setup is always the strongest way to integrate Claude.Why is that wrong?
Anthropic recommends starting with the simplest solution and adding complexity only when it's needed. Sometimes that means building no agentic system at all.
Covered in Augment the process, or redesign it?
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Outline an ideal customer interaction to define how and when you expect the customer to interact with Claude.”
↩︎ Start from the process you already run“Customer support chat is a collection of multiple different tasks, from question answering to information retrieval to taking action on requests”
↩︎ Start from the process you already run“Here are some key indicators that you should employ an LLM like Claude to automate portions of your customer support process”
↩︎ Good candidates and steps that stay with a person“Claude excels at handling a large number of similar questions efficiently, freeing up human agents for more complex issues.”
↩︎ Good candidates and steps that stay with a person“Claude can handle sudden increases in query volume without the need for hiring and training additional staff.”
↩︎ Good candidates and steps that stay with a person“This outline will help to determine the technical requirements of your solution.”
↩︎ Augment the process, or redesign it?“break down your ideal customer interaction into every task you want Claude to be able to perform.”
↩︎ Key concept“Customer support chat is a collection of multiple different tasks, from question answering to information retrieval to taking action on requests”
↩︎ Exam trap 1 - 2.
“No method on that interface places an order or moves money.”
↩︎ Good candidates and steps that stay with a person“Checkout stages a summary the customer confirms in your app.”
↩︎ Good candidates and steps that stay with a person - 3.
“Understand the legal implications of errors in the summaries, which could lead to legal liability for your organization or clients.”
↩︎ Good candidates and steps that stay with a person - 4.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“For many applications, however, optimizing single LLM calls with retrieval and in-context examples is usually enough.”
↩︎ Augment the process, or redesign it?“workflows offer predictability and consistency for well-defined tasks”
↩︎ Augment the process, or redesign it?“Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one.”
↩︎ Augment the process, or redesign it?“Directing different types of customer service queries (general questions, refund requests, technical support) into different downstream processes, prompts, and tools.”
↩︎ Augment the process, or redesign it?“we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.”
↩︎ Exam trap 2