Agentic AI is AI that acts rather than only answers. An AI agent is given a goal, reads the information available to it, decides the next step, uses tools such as email, a CRM or a database to carry it out, and repeats until the task is done or a person needs to decide. A chatbot stops at the reply. An agent stops when the work is done. A first agent on one business workflow is typically live in three to six weeks.
This post explains how an agent works, what it can take on in a business, four worked examples, what it costs and how long it takes, and the controls that keep it safe.
How does an AI agent work?
Every agent runs the same loop, however simple or elaborate the task.
- Read. It takes in the situation: the new email, the purchase request, the overdue invoice, the support ticket, plus whatever context it has been given access to.
- Decide. Using a language model, it works out what the next step should be and whether it is allowed to take it.
- Act. It uses a tool: sends an email, looks up a record in the ERP, updates the CRM, drafts a document, posts on WhatsApp.
- Check. It looks at the result and decides whether the task is done, whether another step is needed, or whether a person should look.
- Escalate or finish. Anything above its limits goes to a named person with the context attached. Everything else is logged and closed.
The model provides the judgement. The tools provide the hands. The limits and the logs are what make it fit for a business.
What is the difference between an agent and a chatbot?
A chatbot answers questions inside a conversation. Ask it which purchase orders are waiting and it will tell you, if it has access. An agent completes the task: it finds the orders, compares the three vendor quotes, drafts the approval and routes it to the person who owns the budget, then waits for the tap. The chatbot's job ends at the reply. The agent's job ends when the work is done.
A second difference is memory of the process. An agent follows a written workflow with defined steps and gates. A chatbot has a conversation.
Four worked examples
These are patterns we have built, with the details changed.
1. Enquiry follow-up for a services firm. Enquiries came in from the website, email and WhatsApp, and about a third were never answered within a day. The agent reads each enquiry, looks up whether the sender is an existing client, drafts a reply in the firm's tone with the right next step, and books a call slot if asked. Replies to new contacts go out automatically; replies that quote a price wait for a partner's approval. Median first response fell from a day to eleven minutes.
2. Purchase approvals for a construction company. Site engineers raised requests on the phone. The agent checks the request against the material list, gathers the vendor quotes that arrive by email and WhatsApp, reads them from PDFs and photographs, builds the comparison with tax and freight, and sends it to the approver. Nothing is ordered until the approver taps. Days from request to order went from nine to two.
3. Support triage for a software product. Every incoming ticket is read, classified, matched against known issues and the documentation, and given a draft reply. Tickets about billing or anything that mentions cancellation go straight to a person. Routine answers go out after a one-click review by the support lead, who now handles three times the volume.
4. Weekly reporting for a distributor. Every Monday the agent pulls sales, stock and receivables from three systems, writes the two-page summary the owner used to compile by hand, flags the five numbers that moved most, and sends it. Nothing here needs approval because nothing leaves the company; the owner reads it over coffee instead of building it.
Each of these is one agent on one workflow. The second agent in a company is cheaper than the first because the integrations already exist.
What can an agent take on in a small business?
The pattern is the same regardless of industry. Good first candidates are tasks that are repetitive, cross two or more systems, and have a clear rule for when a person must decide:
- Following up enquiries and quotes
- Chasing overdue invoices, politely and on schedule
- Reading documents into structured data: quotes, invoices, forms, claims
- Sorting an inbox and drafting replies
- Preparing recurring reports
- Onboarding steps: creating accounts, sending welcome packs, scheduling training
Poor first candidates are tasks with no written process, tasks where the data lives only in someone's head, and tasks where a wrong action cannot be reversed and cannot be checked first.
What does it cost, and how long does it take?
| Step | Typical duration |
|---|---|
| Pick the workflow and map the steps and gates | 1 week |
| Build and connect to your systems | 1 to 2 weeks |
| Run beside a person, measure, tune | 2 to 4 weeks |
| Hand over with dashboard, logs and a named owner | included |
Three to six weeks to a first agent is the normal range. Cost is scoped per workflow and quoted in writing after a call. The first agent carries the cost of the integrations; later agents reuse them. Running costs are model usage and the orchestration platform, reported monthly, and for most small and mid-sized companies they are modest next to the hours involved.
Is agentic AI safe?
An agent is as safe as its limits, so the limits are designed first.
- Its own login, minimum access. The agent sees only what its job needs.
- Spending and sending limits. It cannot send more than a set number of messages, approve above a set amount, or touch records outside its scope.
- Human approval gates. Sending, paying, approving and publishing wait for a person wherever you set the threshold, and each approval is logged.
- Full logging. What the agent read, decided and did, step by step, so any case can be replayed.
- A kill switch that two people know how to use.
- Private deployment when the data cannot leave your environment.
The checklist for this is in safe AI for business.
Where Quadrivium fits
We build AI agents for business workflows on Claude, OpenAI models and n8n, connected to the systems you already run, with the gates, logs and kill switch in place from the first run. Our team holds n8n professional certification and Quadrivium is a Microsoft, Anthropic and OpenAI partner. The service is on the agentic AI page. For the wider programme an agent usually sits inside, read what is AI transformation; for how a build is scoped and priced, custom AI development.
