A chatbot answers questions. An agent does a recurring job inside your tools, and asks before anything that matters.
We are an engineering firm. We build AI agents built for your operations and, beside them, custom software. In short: an operations agent reconciles, drafts, briefs and watches thresholds, on a trigger or a schedule. It reads freely, drafts for a person’s approval, and acts alone only where the action can be undone. Start with one frequent job, with rules you can write down, that you can reverse.
Chatbot, assistant, agent: three different things
The three words get mixed up in sales pitches. They are different tools, bought for different reasons.
A chatbot answers questions
A chatbot sits on your website or help centre and answers from what it was given: opening hours, delivery terms, the returns policy. It waits for a question, replies, and the conversation ends there. It serves the front door.
An assistant helps a person who’s working
Claude or ChatGPT, connected to your business data, is an assistant. Someone asks “summarise this client’s open orders” or “draft a reply to this supplier”, reads the answer and decides what to do with it. The person starts every exchange and checks every step. To set one up for your team, see how to connect Claude or ChatGPT to your business data.
An agent does the job on a trigger or a schedule, and asks when it should
An starts by itself: a delivery note arrives, or it is six in the morning. It works through the steps in your systems, reads the order, compares, drafts the reminder, and stops where a person should decide. Anthropic, the company behind Claude, draws the line in Building effective agents (December 2024): workflows follow “predefined code paths”, while agents are “systems where LLMs dynamically direct their own processes and tool usage”. In plain words, the agent picks its next step; a fixed workflow follows steps written in advance.
Term AI agent Software that uses an AI model to carry out a task in steps: it reads, decides what to do next, and uses your tools to do it.
| Chatbot | Assistant (connected) | Agent | |
|---|---|---|---|
| Who starts it | A person asks | A person works with it | A trigger or a schedule |
| What it touches | Answers from what it was given | Reads and drafts in your data | Reads, drafts and acts, within its permissions |
| Who decides | The person | The person | The agent within its rules; a person for anything that matters |
| Typical example | For example, the FAQ on your website | For example, "summarise this client's open orders" | For example, every morning, list late deliveries and draft supplier reminders for approval |
| Main risk | A wrong answer | A wrong draft | A wrong action, hence permissions and logs |
Good first jobs for an operations agent
The best first agents take the checking and the first drafts that fill the start of everyone’s day. Each example below is generic, for a firm that buys, stocks and delivers.
Reconciling: the shipment against the order, the invoice against the delivery
For example, a shipment arrives: the agent compares the delivery note with the purchase order, line by line, and flags the lines that differ. Or a supplier invoice comes in: it checks it against what was delivered and the agreed price, and passes the clean ones to the person who approves payment.
Drafting: reminders, supplier replies, quotes, waiting for approval
For example, a reminder for an overdue invoice, a reply to a supplier asking for a new delivery date, or a quote from a request that came in by email. The agent writes the draft with the right figures from your systems; a person reads it, edits it if needed, and sends it.
Briefing: the morning summary
For example, a summary waiting in the team’s inbox or chat before anyone arrives: today’s deliveries, the exceptions from overnight, the orders running late, the approvals waiting. Each line links to its record, so nobody opens five tools to build the picture.
Watching: thresholds, expiring documents, stuck approvals
For example, stock falling under its reorder level, an insurance certificate expiring next month, an approval sitting with the same person for three days. The agent notices, tells the right person and drafts the next step.
What to keep human
Money decisions, negotiations, a hard conversation with a client, and any action with no undo. An agent can prepare each of them: the figures, the history, a first draft. The decision and the final word stay with a person.
How to choose the first one
Pick the job by how it behaves. Five questions, answered yes or no:
- Does it happen every week, or more often?
- Can you write its rules down, the way you would brief a new colleague?
- Can a mistake be undone?
- Is the data it needs reachable today, in a tool, an export or an inbox?
- Is one person named to own it?
If any answer is no, start with another job. A frequent job pays back sooner. Written rules become the agent’s instructions. An undo keeps a mistake small. Reachable data means you start with the tools you have. And an owner means someone notices when the agent drifts.
Read, draft, act: give each action its permission
Approval is often pictured as one gate at the end. It works better set action by action: the same agent can read everything in its scope, draft some things, and do a few alone.
Read only
The agent looks and reports. It reads orders, stock and the inbox, and tells you what it found. For example, the morning brief, or the list of deliveries that don’t match their order. Your systems stay as they are. Start your first agent here.
Draft, then a person approves
The agent prepares the change and waits: the reminder sits in a queue, the corrected order waits for a yes. For example, payment reminders drafted at six, approved by the finance lead at eight. Anything sent outside the company, and anything touching money, stays at this level. Approval steps are standard in agent tools: OpenAI’s Agents SDK documents a flow to “pause agent execution until a person approves or rejects sensitive tool calls”, and n8n’s documentation says “You can require human approval before an AI Agent executes a specific tool.”
Act alone, only where it can be undone
Some actions are safe to take alone because they stay inside the company and can be reversed: tagging an email, updating a status, filing a document, adding a note to an order. For example, marking a delivery “checked” once every line matches. If the agent is wrong, a person changes it back. An action moves up to this level only after the agent has done it right, under approval, over a period you set.
What a production agent needs
A demo agent works once, on a good day. An agent your team relies on every morning needs eight things around it. Without them, an agent stays a pilot; we wrote about why enterprise AI pilots stall.
- A trigger. What starts it: a new email, a new delivery note, a time of day.
- Records from their owner. It reads each record from the system that owns it, so it works from this morning’s figures.
- Named tools only. It can use only the actions its job needs, each one named: read orders, draft an email, update a status. Connectors (MCP) give an agent its tools one at a time; our MCP explainer shows how.
- Approval levels. Which actions wait for a person’s yes, written down action by action, at the read, draft and act levels above.
- A log. A record of what it read, what it decided and what it did, kept where you can search it. When someone asks why a reminder went out, the answer is in the log.
- A test set of past cases. Real cases from your own history, each with its right answer. Before any change to its instructions or its AI model, the agent runs through them, and the change ships only if it still gets them right.
- A named owner. One person on your team who reads its exceptions, adjusts its rules and answers for it.
- An off switch. One step that stops it, known to the owner, and the work goes back to the team as it was done before.
Agents work on the tools you already have
An agent works on your accounting software, your inbox, your spreadsheets and your business app, through accounts and access you control. When a tool opens a connection to other software, its , the agent uses it; when a tool only has an inbox or an export, the agent works from that. You can start with the stack you run today.
Term API An application programming interface: the way software talks to a system, to ask it for data or tell it to do something.
Better software is a separate decision. When the data is scattered across spreadsheets nobody keeps in step, custom business software can fix that as its own project, run in parallel with the first agent. Each is worth doing on its own.
Who owns the agent
Ask any builder three questions before you start:
- Whose accounts? The AI provider, the hosting and the connections are in your company’s name. The builder is a guest in them.
- Whose instructions? The agent’s instructions, its rules and its test set are files you hold, readable by the next engineer or by your own assistant.
- Whose logs? The record of what it read and did is stored in your accounts, for as long as you decide.
With those three, you can leave and it keeps working. Maintenance becomes a choice: hand it over to your team with training, or keep us running it.
People stay in charge
An agent takes the checking and the first drafts off people’s mornings, so they spend that time where judgment counts: the supplier who needs a call, the client with an unusual request, the decision on an exception. People decide what each agent may read, draft or do, and they answer for the results. The team you have takes on more, and each new hire can do more.
People do their best work with the right tools. That was true before AI, and it is what we build.
Frequently asked questions
Is ChatGPT an AI agent?
In a chat, ChatGPT is an assistant: it answers when a person asks, and the person decides what happens next. An operations agent runs on its own trigger or schedule, inside your systems, with a permission set for each action. To give your team an assistant on your own data, see how to connect Claude or ChatGPT to your business data.
Are AI agents safe to give access to company systems?
They are as safe as the access you give them. Give the agent its own account with only the permissions its job needs, a person’s approval before anything that touches money or leaves the company, and a log of everything it does. Then test it on past cases before it runs on live ones.
What is human-in-the-loop for AI agents?
The agent stops at set points and waits for a person’s yes before it carries on: before sending an email, changing an order or approving a payment. The person sees the draft, edits or rejects it, and the agent continues from there.
Do AI agents work with our existing ERP, accounting software or spreadsheets?
Yes, through what each tool opens to other software: its API, a shared inbox or a file export, with an account you control. We check each tool before work starts. If the back office itself is what holds you back, an operations back office built around your processes is a separate project.
Who owns the agent if we stop working with the builder?
You do, when it is set up that way from day one. The accounts, the instructions, the test cases and the logs sit in your company’s name, so the agent keeps running and the next engineer can read how it works. Maintenance by the builder stays optional.
Talk it through with us
Bring the job you’d hand to an agent first. By the end you’ll know what an agent could change for you, and if we can’t help, we won’t follow up.
Book a free 30-minute call. No commitment. You pick the time.
