What an AI agent is, and when you need one
An AI agent is a language model connected to a set of tools, such as your CRM, inbox, database or calendar, that decides which tools to use and in what order to finish a task. A regular automated workflow follows a fixed path. An agent picks its path based on what it finds along the way.
That flexibility is useful and risky in equal measure. If a task always follows the same steps, a normal workflow is cheaper, faster and easier to audit, and we’ll recommend that instead. Agents make sense when the steps vary from case to case, like researching a prospect, resolving a support request that could touch three systems, or reconciling records that don’t line up.
AI agents for business work best on narrow jobs with clear success criteria. We build agents that do one job well, with a short list of tools they’re allowed to use, and slot them into the wider AI workflow automation we run for you. Customer-facing assistants are covered on our AI chatbots page.
Agents we build
Prospect research agent
Given a new lead, it reads the company website and public listings, checks your CRM for past contact, and writes a short brief with talking points before your salesperson replies.
Support resolution agent
It reads a ticket, looks up the order in Shopify or your database, checks the policy document, and drafts a resolution. Refunds and credits wait for a person to approve.
Operations reconciliation agent
It compares records across two systems, such as bookings against invoices, works out why they differ, and either fixes simple mismatches or lists the rest for review.
Internal knowledge agent
Staff ask questions in Slack or Teams and the agent answers from your SOPs, price lists and past project files, citing the source document each time.
Scheduling agent
It reads booking requests by email, checks staff calendars and job locations, proposes times and books the slot once the customer confirms.
Guardrails built into every agent
An agent is only as safe as the limits around it, so these come standard.
Scoped tool access
Each agent gets its own credentials with read or write access only to what its job requires.
Approval gates
Actions that spend money, contact customers or change important records pause for a person to approve.
Step and cost limits
Agents stop after a set number of steps or a usage cap, so a confused run can’t loop forever.
Full run logs
Every tool call, input and decision is recorded so you can see exactly why the agent did what it did.
Test sets
We keep a library of real past cases and re-run it whenever the prompt, model or tools change, to catch regressions before they reach you.
Fallback to a person
When the agent isn’t confident or hits something outside its job, it hands the case to your team with notes.
How we build a custom agent
- Define the job We write down the goal, the inputs, what a good result looks like and which actions need approval. This becomes part of the proposal, with a fixed monthly scope and price.
- Connect the tools We set up the tool connections, often through APIs or MCP servers, with the narrowest permissions that work.
- Test on past cases The agent runs against real historical examples until its answers match what your team would have done.
- Shadow mode It runs alongside your staff and suggests actions without taking them, so you can compare before trusting it.
- Supervised launch and monitoring Once you sign off, it goes live with approvals in place. We review its logs and send a report every two weeks.
Questions we get about custom AI agents
What’s the difference between an AI agent and a chatbot?
A chatbot mainly holds a conversation and answers questions. An agent takes actions in your systems, like updating a record or booking a slot, to finish a task. Some projects combine both, and our AI chatbots page covers the customer-facing side.
Which AI model do you use for agents?
Usually Claude or OpenAI’s GPT models, chosen by which one handles your task and tools best in testing. We design agents so the model can be swapped later without rebuilding the whole thing.
Can an agent act without anyone approving it?
Only for actions you’ve agreed are low risk, like adding a note to a CRM record. Anything involving money, customers or deleting data goes through an approval step by default.
What happens when the agent gets something wrong?
The run log shows exactly which step went wrong and why. We add that case to the test set, fix the prompt or tool, and re-run the tests before updating the live agent.
Where does the agent run?
Agents can run inside n8n, as a small cloud service on your account, or on your own server. We pick based on your data rules and what the agent needs to reach.