Hand repetitive work to AI agents, and keep the decisions with your team.
Reading inbound emails, updating the CRM, chasing documents, answering order questions: work that follows a pattern but still needs some judgment. I map one workflow with the person who owns it, automate the predictable steps, use an AI agent where judgment is needed, and put a person’s approval in front of anything that matters.
It’s a short form: what you need, a few lines about the project, and your email. I read every inquiry myself and reply by email, usually within 24 hours.
- Customer emails“Where’s my order?”
- Agent finds the orderIn the store, with the carrier’s status
- Agent drafts a reply
- A person approvesLate parcels and refunds go to a person
- Reply sent and logged
How I build it
Map the workflow.
Inputs, systems, actions, exceptions, and who decides what. If simple automation is enough, I’ll tell you; an agent isn’t always the right tool.
Connect the systems.
Connectors to your CRM, email, documents, store or database, with only the permissions the workflow needs.
Set the approval points.
You choose which actions need a person: emailing a customer, issuing a refund, changing a record.
Evaluate before launch.
The workflow runs against real past cases, with your permission, so we can see where it gets things right and where it doesn’t. After launch, the log shows every action.
Roll out gradually.
At first a person approves every action. Approvals are relaxed only where the results justify it, and only if you agree.
You get: the workflow running in your systems, the evaluation results, the approval and logging setup, and either a handover to your team or an agreement for me to keep operating it.
For Shopify and WooCommerce stores
Repetitive store work, for example order-status questions, returns intake, product data updates or supplier emails, follows the same approach: automate the routine part, and route anything unusual to a person.
An agent can also analyze store data and propose experiments to improve conversion. A person decides what gets tested, and the test design and evaluation are CRO work. Changes to the storefront itself are Shopify or WordPress work; an agent doesn’t make them on its own.
Example: “Where’s my order?” emails
A customer emails asking where their order is.
The agent finds the order in Shopify and the tracking status from the carrier.
It drafts a reply with the order details.
If the parcel is late, or the customer asks for a refund, the case goes to a person instead.
At first, every reply waits for approval. Later, if you choose, plain status replies can go out on their own.
The screen where a person approves these is software work: dashboards & approvals.
Before we start
- A process owner, and examples of the work as it’s done today.
- Access you’re allowed to give: APIs, accounts, data and permissions.
- Rules for escalation and approval, which we write down together.
- Running costs. AI model and API usage is billed as the workflow runs. We agree who pays it and how it’s watched.
- Your data. I work with the major AI providers and OpenRouter, and with open-source or on-premise models on request. Which provider processes your data, and on what terms, including whether it may be used for training, is agreed per project.
- Not promised: accuracy, unattended round-the-clock operation, or fewer staff. Agents make mistakes; that’s why consequential actions need approval and why the workflow is evaluated before and after launch.
Why me for this
I build AI agents and B2B software connected to CRMs, data and approval steps, and I design the approval points with the people who’ll use them.
My experience
Most recently I built an agentic system for conversion optimization. It works on ecommerce data in a closed loop: it reads the analytics, suggests changes, evaluates them, runs the tests and starts again from the results. It can run on its own, or with a person approving what gets tested.
How we’d work together
I usually work for a fixed weekly rate. The work is scoped around a goal rather than a set number of days, and we agree the scope and terms for each project. They’re open to discussion. Discovery, the build and ongoing operation are agreed separately. After launch, I can keep operating the workflow or hand it to your team.
Questions
Do we need an AI agent, or just automation?
Often the answer is some of each. Steps with fixed rules are cheaper and more reliable as ordinary automation. An agent earns its place where the input varies, such as free-text emails or documents in different formats.
What happens when the agent gets something wrong?
Consequential actions wait for a person’s approval, so a wrong draft gets corrected before it goes out. Every action is logged, and the workflow is tested against past cases before launch.
Can it work with our CRM?
If the CRM has an API and you can give access, usually yes. I check before we agree scope.
Who runs it after launch?
Your choice. I can hand it over with documentation, or keep operating and improving it under a separate agreement.
Further reading
Model Context Protocol (MCP) Server Guide: Build Tool-Connected AI
Connecting agents to business systems, with permissions, approvals and evaluations.
Articles and guides about this work. They explain methods and trade-offs; they aren’t case studies.
Which work would you hand over first?
Describe the process, the systems it touches, and the points where a person has to decide. I’ll tell you which parts suit automation, which suit an agent, and which should stay with your team.