What is an AI agent for a small business?

The plain-English definition

An AI agent is software that does work, not just conversation. Four traits separate it from every "AI chat" product you have seen so far:

• It uses tools. An agent connects to the systems you already run — email, calendars, a CRM like HubSpot, spreadsheets, your inventory — and acts through them: it looks things up, drafts replies, prepares updates.
• It keeps memory. Persistent memory carries across sessions, and knowledge-base grounding over your uploaded documents — product specs, policies, price lists, SOPs — means answers come from your business, not from the open internet.
• It follows procedures. Teach a procedure once as a skill and it is repeated the same way every time. "How we quote", "how we confirm a booking", "how we escalate a complaint" become reliable routines instead of tribal knowledge.
• It escalates to humans. When an action matters — sending a message, spending money, changing a record — the agent queues it for a person to approve instead of acting alone.

The closest analogy is a capable new hire: give them your handbook, show them the tools, let them draft before they send, and widen their responsibility as trust builds. The difference is that this teammate reads every message instantly, never loses a thread, and works nights and weekends.

Not a chatbot. A private team of agents.

The easiest way to place agents on the map is a three-step ladder.

Step 1 — the chatbot

A chatbot answers questions in a conversation. That is the whole job. It lives inside the chat window, follows a script or a language model, and forgets you the moment the window closes. Useful for FAQs; stuck the moment a customer asks "can you check if it's in stock?"

Step 2 — the AI agent

An agent uses tools, keeps memory, and carries out multi-step work. The same enquiry now runs end to end: read the message, check live stock in the connected system, draft a reply with the price, queue it for approval, and follow up if the customer goes quiet. The conversation is just the surface; the work happens behind it.

Step 3 — the agent team

Real businesses are not one job, so the mature setup is not one agent. It is a team of specialists that delegate to each other: a front-desk agent answering WhatsApp hands a stock question to an inventory specialist; a research agent passes findings to the agent that drafts your weekly report. This is the shape Olano is built around — the positioning is literally "Not a chatbot. A private team of agents." — with 180+ specialist agents and 19 starter teams to compose from.

For a deeper side-by-side, read AI agent vs chatbot.

What keeps it safe

The reasonable worry about giving software the keys to customer conversations is exactly the right one, and it is why serious agent platforms are built approval-first rather than autonomy-first. Olano's model has five parts:

• Trust levels 0–4. Every agent runs at an explicit level — Observer, Assistant, Collaborator, Autonomous, Developer — so autonomy is a dial you turn, not a switch you flip.
• Approval gates by category. Financial actions, external sends, file writes, secrets, and delegation each have their own gate. Anything outbound waits for approval by default; read-only research and drafting can run without gates because they change nothing outside the workspace.
• An immutable audit trail. Every action is logged. When you want to know what was said, sent, or changed — and when — the full history is there.
• Hard spend caps. AI usage runs on a transparent meter with a hard spend cap agreed before launch, so cost is a decision you made, never a surprise.
• A private, isolated workspace. Each customer's agents run in their own workspace, with encrypted connector tokens and API keys, role-based access controls, and data encrypted at rest and in transit.

In practice this means an agent can do a full morning's drafting and research on its own, while the ten actions that actually touch a customer or a bank balance sit in a queue waiting for a human tap.

What an agent needs from you

Agents are not magic; they are trained teammates. Three inputs make the difference between a gimmick and a genuinely useful hire:

1. Documents for the knowledge base. Product specs, price lists, policies, SOPs — whatever you would hand a new employee on day one. The agent grounds its answers in these instead of guessing.
2. Procedures to teach as skills. Write down (or talk through) how you handle a quote, a booking, a refund request. Taught once as a skill, the procedure is repeated the same way every time.
3. An owner for approvals. Someone on your team reviews the approval queue, edits drafts, and gives feedback. Ten minutes a day at the start; less as trust levels rise.

Where to start

Not with a grand automation programme. Start with one workflow — usually the one with the highest volume of repetitive messages, which for most SMEs is customer enquiries or the inbox. One workflow is cheap to try, easy to judge, and quick to deploy: with Olano, every build is quoted and approved before we build, a single workflow is typically live within 1–2 business days of onboarding, and bigger rollouts take 1–3 weeks. Live in days, not months.

The simplest first step is a free 30-minute AI assessment: we map where your team's hours actually go, pick the one workflow with the best payoff, and tell you honestly if an agent is not the right fit yet.