AI agents that do the repetitive work inside your systems, with a human in the loop
- One workflow first, measured against today
- Read-only first, approvals for every action
- Audit logs, spend limits and a kill switch
- Connected through APIs, events or MCP
Agents vs automation: when you need one
Anthropic's engineering guidance draws a useful line between workflows, where language models and tools are "orchestrated through predefined code paths", and agents, where the model decides its own steps. Most business processes sit on a ladder:
- A rule or a scheduled job, when the steps never change.
- Robotic process automation, when a screen has no API.
- A fixed workflow with a language model as one step: reading an e-mail, extracting fields from a document, drafting a reply.
- An agent, when the steps genuinely vary from case to case.
Each step up adds flexibility and adds risk and cost. We recommend the lowest rung that does the job, and say so when a cron job is enough.
Where agents fit
Order exceptions
Triage failed payments, address problems and stock conflicts, propose the fix, and queue it for a person to approve.Lead qualification and CRM hygiene
Enrich new enquiries, flag duplicates, route leads to the right person and keep records consistent.Document intake
Read invoices, forms and applications, extract the fields, check them against your rules and send exceptions to a person.Reporting
Pull the numbers from several systems on a schedule, write the summary, and point out what changed and why it might matter.Back-office reconciliation
Match payments, orders and invoices across systems and prepare the mismatches for review.Your workflow
If a task is repetitive, rule-heavy and checked by someone anyway, it is a candidate. The AI readiness audit ranks the candidates.How we connect agents to your systems
An agent is only as safe as the doors it is given. We expose the smallest set of actions it needs, through one of three routes:
- APIs: direct calls to your store platform, CRM or custom application, with credentials scoped to the agent.
- MCP servers: the Model Context Protocol lets several AI applications use the same tools and data through one server. See MCP explained for business owners.
- Events and queues: the agent reacts to events (a new order, a failed payment) instead of polling, which keeps it predictable and cheap.
This is integration work our team has done for years on Laravel, Node.js, Magento, Shopify and AWS. Our article from ESB to APIs, events and MCP explains how the layers fit together.
Guardrails
Permissions
Least privilege: read-only access first, and each write action added on purpose.Approval steps
A person approves actions that cost money, reach a customer or are hard to undo, until the logs show they can be trusted.Audit logs
Every input, decision and action is recorded, so you can see what the agent did and why.Spend limits
Caps on model usage and on the value of actions per day, with an alert before they are reached.Kill switch
One switch stops the agent and returns its work to the queue your team already uses.Delivery model: one workflow first, then scale
- Pick one workflow and measure it as it runs today: volume, time per case, error rate.
- Build a short pilot, sized to the workflow, with the guardrails above.
- Run it in the shadow of your team: the agent proposes, people decide, and we compare.
- Decide with the numbers: stop, adjust, or let it act on its own where the logs support it.
- Scale to the next workflow only after the first one is owned by someone on your side.
What we have built
We have built a retrieval-grounded (RAG) assistant, an MCP server that lets AI applications use a business system through defined tools, embeddings-based semantic search, and an AI workflow agent whose actions a person approves. We describe them in general terms here, without client names or numbers; ask us for the details on a call.
The rest of our background is the systems AI has to plug into: we have built websites, online stores and mobile apps since 2005, today mostly on Laravel, Node.js, Magento, Shopify, WooCommerce and AWS, and in October 2026 we moved our own site from WordPress to Astro on AWS.
Further reading from our team:
What we will not do
The same rules apply to every AI project we take on. They are the reason some projects end with a smaller scope, and a few with a recommendation to wait.
- Put an agent into production that acts without a person who can approve, stop and undo what it does.
- Train or fine-tune a model on your data, or send it to a provider that does, without your written consent.
- Publish AI-generated text on your site or send it to your customers without a person reviewing it first.
Frequently asked questions
What is the difference between an AI agent and automation?
Automation follows steps you defined in advance. An agent is given a goal and tools and decides which steps to take. Most business workflows need automation with a language model as one step; an agent is worth its extra risk only when the steps genuinely vary from case to case.
Can an agent make changes in our systems without approval?
Only the actions you choose to allow. We start agents read-only, add write actions one at a time with a person approving each, and only remove an approval when the logs show it is consistently right and the action is easy to undo.
What is MCP and do we need it?
The Model Context Protocol is an open standard for exposing a system's tools and data to AI applications. It is useful when several assistants or agents need the same system; for a single workflow a direct API integration is often simpler.
How long does a pilot take?
We scope a short pilot on one workflow, then run it alongside your team before it handles real cases on its own. The pilot ends with measured results and a decision to stop, adjust or scale.
Which systems can you connect to?
Anything with an API or a database we are allowed to read: store platforms such as Shopify, Magento and WooCommerce, CRMs, helpdesks, accounting tools, spreadsheets and custom Laravel or Node.js applications, usually through AWS services.
Pick one workflow for a pilot
Tell us which task your team repeats most often. We will tell you which rung of the ladder it needs.
