To automate a business process with Claude, pick one repeated, rule-bound task, map it, choose the lightest tool that can do it, put a person at every risky step, test it on real cases, log every run and measure the hours yourself. This guide gives that plan in eight steps, a table to choose the right layer, and a pre-launch checklist.
Devvista is a software development and growth agency that builds AI automation, workflow automation and custom integrations for business owners. We wrote this for owners who need an order of operations, not another demo.
01 Step 1: Pick One Process That Is Worth Automating
The most common mistake is starting with a tool ("we should use Claude for something") instead of a process. Start with the work that annoys your team every week and score it on four points.
| Question | Good candidate | Wait or keep it human |
|---|---|---|
| Frequency | Happens daily or weekly, in the same shape each time | Happens a few times a year |
| Rules | You can write down how a good person decides | The answer depends on a relationship or on instinct |
| Data | Inputs arrive in a place software can reach (email, form, CRM, spreadsheet) | Inputs live on paper, in someone's head or in a system nobody has access to |
| Risk | A mistake is cheap and visible, and can be undone | A mistake sends money, signs a contract or reaches a regulator |
Pick a process that scores well on all four. High frequency and low risk is the sweet spot for a first project. Inbound enquiry triage, invoice and document intake, meeting follow-ups, report drafting and support-ticket sorting usually qualify. Our post on AI automation use cases for companies lists more.
02 Step 2: Map the Process on One Page Before You Touch a Tool
Write the process as the people doing it today would explain it to a new hire. One page is enough. You need five things on it:
- The trigger. What starts the work: an email, a form, a new CRM record, a date.
- The inputs. What information arrives, and in what messy forms.
- The steps and rules. What a good person does, in order, and what they check.
- The outputs. Where the result goes and who relies on it.
- The exceptions. The cases that go to a manager, and why.
The exceptions list is the most valuable part. It tells you where a person must stay in the loop and what the tool should do when it is unsure: stop and ask, not guess. Anthropic's engineering guidance on agents makes the same point from the other side: agents suit open-ended problems, while predefined paths suit well-defined tasks, so a clearly mapped process often needs less AI than you expect.
03 Step 3: Choose the Layer: Claude.ai, API, Claude Code, Agent, Workflow Platform or Custom Code
"Automate with Claude" can mean six different things. Anthropic advises finding "the simplest solution possible, and only increasing complexity when needed". Use the table to find the lightest layer that fits your process.
| Layer | Use it when | What you own afterwards | Watch out for |
|---|---|---|---|
| Claude.ai chat, with connectors | One person, one task at a time, a human reads every result | A saved way of asking; nothing runs by itself | It depends on someone opening it; connectors need access reviews |
| Claude API (tool use) | A single step inside your own software: classify, extract, draft | An application that calls the API | Your code runs the tools and returns results, so you own the error handling |
| Claude Code | You need scripts, integrations or an internal tool built and changed quickly | Ordinary code in your repository | Someone must review the code and run it somewhere |
| An agent (Agent SDK or similar) | A multi-step job where the next step depends on what the last one found | An agent program with tools, permissions and logs | Costs more and errors can compound, so limits and checkpoints matter |
| Workflow platform (Make, n8n) | Connecting apps with triggers and steps you can see on a canvas | Scenarios or workflows in the platform | Platform limits, plan terms and where your data is processed |
| Custom code | High volume, strict security, or logic a canvas cannot express cleanly | A system you fully control | Needs real engineering and maintenance |
Claude Code is an agentic coding tool that reads a codebase, edits files and runs commands, and it can also be run from scripts with claude -p. The Agent SDK exposes the same tools and agent loop in Python and TypeScript. Make's automations are called scenarios, and its Help Center describes storing failed scenario runs as incomplete executions. n8n describes itself as a workflow automation tool that combines AI features with business process automation, in cloud or self-hosted form.
Most real systems combine layers: a workflow platform for the triggers and app connections, an API call for the judgement step, and custom code where the volume or the security demands it. Our guides to Claude Code for business automation and AI agents for business go deeper on two of these layers.
04 Step 4: Add Approval Points Where a Mistake Would Cost Real Money
An automation that can send, pay, delete or promise something needs a person at that step, at least until it has a record. Design the approvals before you build, not after the first incident.
- Draft, then approve. Let the tool prepare the reply, the invoice or the record. A person clicks send.
- Read freely, write carefully. Give the tool read access to most things and write access to very few.
- Thresholds. Anything over a value you choose, or from a customer you flag, goes to a person.
- A safe stop. When inputs look wrong, the automation parks the item and tells someone, instead of guessing.
The tools support this. Claude Code's manual mode starts read-only and asks before edits or commands (your default may differ, so check), Claude connectors let you set each tool to Always allow, Needs approval or Blocked, and n8n documents a human-in-the-loop option that requires approval before an AI agent runs specific tools. One caution from Anthropic's own documentation: instructions you write for Claude are context, not enforced configuration, so a rule that must never be broken belongs in a hook, a permission rule or the integration code, not only in a prompt.
05 Step 5: Test on Real Cases, Including the Ugly Ones
A demo on three clean examples proves nothing. Collect 30 to 50 real past cases (our rule of thumb, not a documented figure) from your own records, including the messy, incomplete and unusual ones, and write down what the right answer was for each. Run the automation on that set before it touches live work.
Then run it in shadow mode: it processes live inputs but a person still does the real work, and you compare the two. Fix the disagreements, then let it handle the easy cases while a person checks the rest. Anthropic recommends "extensive testing in sandboxed environments" for agents, and the reason is practical: you want its mistakes to happen where they cost nothing.
Keep your test set. Every time you change the prompt, the model or the workflow, run it again. That is how you find out a change made things worse before a customer does.
06 Step 6: Log Every Run So You Can See What Happened
If you cannot see what the automation did, you cannot trust it, fix it or explain it to an auditor. Every run should leave a record of:
- What triggered it and when.
- What input it saw and what it decided.
- What it changed, in which system, and who approved it.
- Any error, and what happened to the item afterwards.
Workflow platforms help here. Make stores failed runs as incomplete executions that you can resolve, and n8n documents execution views and error workflows. If you build with code or the Agent SDK, the logging is yours to write. Add an alert to a person for any failed or parked item, so a silent failure cannot sit in a queue for a month.
07 Step 7: Measure the Hours Yourself
We do not promise hours saved, and neither should any vendor, because the number depends on your process, your volume and your team. You can measure it in a few weeks without any special tool.
- Baseline before you start. Time the process by hand on a typical week: how many items, how many minutes each, how many errors or reworks.
- Measure after, the same way. Count the items, the minutes people still spend on approvals and exceptions, and the errors.
- Include the new work. Reviewing parked items and fixing the automation is work too.
- Add the running cost. Subscriptions, API usage and platform fees. Claude Code's JSON output reports a cost figure, but the documentation describes it as a client-side estimate, so use your real invoices.
If the net result is small, that is useful information. Automate something else or leave the process as it is.
08 Step 8: Maintain It, Because Everything Around It Changes
Automations break at the edges: an app changes a field, a password expires, a new type of enquiry arrives, a model behaves differently after an update. Plan for it from the start.
- Name an owner. One person is responsible when it breaks, and has access to fix it.
- Review monthly. Read a sample of logged runs, check the parked items and re-run your test set.
- Keep accounts in your name. Platform, API and cloud accounts belong to the business, never to a freelancer or an employee personally.
- Write down how it works. A short document that someone new can follow.
09 A Worked Process: Triage of Inbound Enquiries
Here is the plan applied to a process most businesses have: enquiries arriving by form and email. This is a description of the design, not a result, and it contains no numbers on purpose, because yours will differ.
| Plan step | For enquiry triage |
|---|---|
| Pick the process | Daily volume, clear categories (sales, support, supplier, spam), low risk if a message is mis-sorted and caught |
| Map it | Trigger: new form or email. Inputs: sender, message, any order number. Rules: how your team tells a buyer from a vendor. Exceptions: complaints, legal terms, large accounts |
| Choose the layer | A workflow platform for the trigger and the CRM connection, one Claude call to classify and draft a reply |
| Approval points | Sales and support replies are drafts a person sends. Anything flagged as a complaint or legal goes straight to a manager |
| Test | Run it on a set of past enquiries with known correct routing, then in shadow mode for a few weeks |
| Log | Each enquiry, the category chosen, the draft, who approved it, and the time |
| Measure | Your own before and after: time to first reply, minutes of triage per day, mis-routed items |
| Maintain | Monthly sample review; update the categories when the business changes |
This is close to work Devvista has delivered. In one case study on our work page, a Make.com workflow captures a long intake form, syncs it between a CRM and a database, uses Claude to draft a document for each lead, and holds delivery behind a Slack review so a person approves before anything is sent. The client is confidential, and that case study describes the system, not your results.
10 When to Do It Yourself and When to Bring in a Team
You can run a first, low-risk automation yourself with Claude.ai, a workflow platform and the steps above. Bring in a team when the process touches money, customer records or regulated data, when volume makes failures expensive, when several systems must be connected through their APIs, or when nobody on your side can own the maintenance.
A team earns its fee on the unglamorous parts: credentials, retries, monitoring, staging and handover. Our decision guide to building AI automation yourself or hiring a team sets out the trade-offs honestly, and our AI automation, workflow automation, API development and AI agents services cover the build side.
Have a process in mind? Contact us or book a call. We scope it with you first, and quote after that short call.
11 The Pre-Launch Checklist
Use this before any automation touches live work. If an answer is "no", that is the next thing to fix.
- Is the process frequent, rule-bound and low-risk enough to be a first project?
- Is it mapped on one page, with the exceptions written down?
- Is the layer the simplest one that does the job?
- Does a person approve every step that sends, pays, deletes or promises?
- Does the tool have the least access it needs, and are the accounts in the business's name?
- Was it tested on at least 30 real past cases (our rule of thumb), then in shadow mode?
- Does every run write a log, and does a failure alert a named person?
- Do you have a before-and-after measure that your own team recorded?
- Is there an owner, a monthly review and a short how-it-works document?
- Do you know what it costs to run each month?
Rolling this out across several countries? Each market has its own AI automation page with local rules: United Kingdom, United Arab Emirates and Australia.
? Frequently Asked Questions
→ Related Resources
Keep reading on automation with Claude, and the services that build it:
- Claude Code for Business Automation
- The CLAUDE.md File Explained
- AI Agents for Business: What They Can Automate
- MCP for Business: How Claude Connects to Your CRM, Email and Accounting
- Build AI Automation Yourself or Hire a Team?
- What Is AI Automation for Business?
- Business Process Automation Services: What Gets Automated and What Does Not
- How to Automate Business Processes With AI
- AI & Automation Services
- Workflow Automation Services
- API Development and Integration
- AI Agents
- Case Studies
→ Sources
Opened on 11 November 2026. Documentation changes, so check the current page before you rely on a detail.
- Claude Code overview: what Claude Code is and where it runs.
- Anthropic: Building effective agents: workflows versus agents, start simple, cost and compounding errors, sandbox testing and human checkpoints.
- Tool use with Claude: client tools run in your application; your code runs the call and returns the result.
- Agent SDK overview: the Agent SDK, the Client SDK and Managed Agents compared.
- Run Claude Code programmatically: claude -p for scripts and CI, JSON output, allowed tools and permission modes.
- Claude Code permissions: which actions ask for approval and how allow and deny rules work.
- Claude Code memory (CLAUDE.md): CLAUDE.md is context, not enforced configuration; use a hook to block an action.
- Claude Code hooks: a PreToolUse hook can block a tool call.
- Claude connectors (Anthropic Help Center): connectors use remote MCP servers; per-tool Always allow, Needs approval or Blocked.
- n8n documentation: n8n is a fair-code workflow automation tool with AI features, in cloud or self-hosted form.
- n8n documentation index: lists n8n pages on MCP, human-in-the-loop for AI tools, error workflows and execution logs.
- Make Help Center: scenarios: Make's automations are called scenarios, and failed scenario runs are stored as incomplete executions.