AI agents and assistants

Why it matters

Assistants save minutes per task; agents can remove whole tasks. For a growing business, that is the difference between hiring another administrator and not. But an agent that sends the wrong email to a hundred customers, or reconciles payments against the wrong invoices, does damage at the same speed. The businesses getting value from agents are the ones treating them as new team members with a probation period: narrow duties, clear limits, and a supervisor who checks the work until trust is earned.

Assistants, copilots and agents

An assistant is conversational: you ask, it answers or drafts, you decide what to do with the result. It has no access to your systems beyond what you paste in. A copilot is an assistant embedded inside a specific tool, such as your email, spreadsheet, CRM or code editor. It can see the data in that tool and propose actions, but a person accepts each one.

An agent goes further. It is given a goal and a set of tools it may use (read a calendar, search a database, send an email, update a record), and it plans and executes a sequence of steps to reach the goal, deciding as it goes. The most useful distinction for a business owner is who presses the button. With an assistant or copilot, you do. With an agent, the software does, within whatever boundaries you set. That is why the boundaries matter so much.

What agents can do for a small business today

The reliable use cases share three features: the task is repetitive, the steps are well understood, and a mistake is recoverable. Triaging and routing inbound enquiries to the right person. Preparing a draft quote from a standard price list and a customer’s request, for a person to approve. Chasing overdue invoices through a defined sequence of polite reminders. Reconciling supplier invoices against purchase orders and flagging exceptions. Scheduling appointments within set rules. Gathering information from several systems into a morning briefing.

The less reliable use cases involve novelty, judgement or irreversible actions: negotiating with a customer, deciding whether a complaint is valid, paying anything, deleting anything, or sending unsupervised communications about sensitive matters. These are the tasks that keep a human at the controls, with the agent preparing the groundwork.

• Good agent tasks: triage, routing, drafting from templates, chasing, reconciling, scheduling, summarising across systems.
• Keep a human on: payments, deletions, complaints, legal or health matters, anything with a named individual’s sensitive data.
• Rule of thumb: if you would not let a new hire do it unsupervised in week one, do not let an agent either.

Guardrails: how to delegate without losing control

Guardrails are the limits you place on what an agent can see, do and spend. The most important is scope: give the agent the minimum access needed for its job, in the same way you would set permissions for a staff member. An invoice-chasing agent needs to read the debtor ledger and send emails from one mailbox; it does not need write access to your accounting system or your customer database.

The second is approval points. Decide which actions the agent may take alone and which must pause for a person: often anything that leaves the business, anything above a dollar threshold, and anything that cannot be undone. The third is logging: every action the agent takes should be recorded so you can audit what happened. And the fourth is a kill switch: a simple way to stop the agent immediately if it misbehaves. Most agent platforms support all four; use them from day one rather than after an incident.

Running agents as part of the team

The businesses that succeed with agents treat them operationally, not as a one-off experiment. Someone owns each agent, reviews its work on a schedule, and keeps its instructions current as the business changes. Price lists move, staff leave, policies shift; an agent working from last year’s rules will confidently do the wrong thing. Reviews should get lighter as trust builds, from checking every action to sampling, but should never reach zero for anything customer-facing.

Expect to adjust. Early agents misread edge cases, over-chase a customer with a payment plan, or route an enquiry to the wrong person. Each incident is a chance to tighten the instructions or add an approval point. This ongoing care is the real cost of agents, and it is where a managed service or a well-run internal routine earns its keep.

How to introduce your first AI agent

1. Choose a repetitive, recoverable task Enquiry triage, invoice reminders or quote preparation are typical starting points.
2. Write the job description Goal, steps, what good looks like, and what the agent must never do.
3. Grant minimum access Its own login, read-only where possible, one sending mailbox.
4. Set approval points External communications, money and deletions pause for a person.
5. Test on realistic data with no live effects Run through a week of real scenarios and review every decision.
6. Go live with full review Check every action for the first fortnight, then sample.
7. Review and update monthly Refresh instructions as prices, people and policies change.

Real-world example

An Adelaide HVAC business hands over its invoice chasing

Torrens Climate Control installs and services air conditioning for homes and small commercial sites. With sixty-odd invoices a month, the office manager, Dev, spends a day a week chasing payments and often misses follow-ups during busy periods. The owner sets up an agent inside their accounting platform’s automation tools: it reads the debtor ledger each morning, identifies invoices past due, and sends a sequence of three reminder emails at set intervals, in a tone Dev wrote.

The guardrails are specific. The agent only emails from the accounts mailbox, cannot change any invoice, skips any customer flagged as being on a payment plan or in dispute, and pauses for Dev’s approval before the third, firmer reminder. Every email is logged. In the first week the agent chases a strata manager who always pays on the 25th; Dev adds a rule for customers with agreed terms. By month three, debtor days have fallen noticeably, Dev spends an hour a week reviewing the log instead of a day chasing, and the owner is cautiously planning a second agent to triage the service inbox.

Common mistakes

• Giving an agent broad access Full admin rights turn a small misunderstanding into a large incident. Grant the minimum and use a separate login.
• No approval points on irreversible actions Payments, deletions and mass communications should always pause for a person.
• Set and forget Agents work from the instructions they were given. When prices, staff or policies change, they must be updated or they will confidently do the wrong thing.
• Starting with the hardest process A failed first agent on a sensitive task poisons the team on the whole idea. Earn trust on something recoverable.
• Treating agent actions as someone else’s responsibility Legally, the business did it. Review outputs as you would a junior staff member’s work.

KPIs to track

• Tasks completed without intervention Share of agent runs finished without a human correction. Should rise over time; if it falls, instructions need work.
• Exceptions escalated How often the agent correctly pauses for a person. A healthy agent escalates when it should, not never and not always.
• Errors reaching customers Any incorrect external action. Target zero; investigate every one.
• Supervisor hours per week Time spent reviewing logs and adjusting. Should fall as trust builds but never reach zero for customer-facing agents.
• Business outcome moved The metric the agent exists to improve, such as debtor days or response time. If it is not moving, revisit the design.

Frequently asked questions

Are AI agents ready for a small business to use?

For narrow, repetitive, recoverable tasks with guardrails, yes, and many accounting, CRM and helpdesk platforms now include agent-style automation. For open-ended or high-stakes work, not yet without close supervision.

How is an agent different from the workflow automation I already have?

Workflow automation follows fixed rules you set in advance. An agent uses a model to interpret situations, choose steps and handle variation. Agents are more flexible and less predictable; many good designs use automation for the fixed steps and an agent only where judgement is needed.

What if an agent does something wrong to a customer?

Treat it as you would a staff error: fix it with the customer promptly, review the log to see why, and tighten the instructions or add an approval point. Legally the business is responsible, so apologise and remedy as you normally would.

Do I need a developer to set up an agent?

For agents built into tools you already use, usually not. For agents that span several systems or handle a core process, a developer or managed service helps with integration, access control and ongoing maintenance.

Key takeaways

• Assistants answer and copilots suggest; agents act. The question is who presses the button.
• Good first agent tasks are repetitive, well understood and recoverable: triage, chasing, drafting, reconciling.
• Guardrails are scope, approval points, logging and a kill switch. Set them before going live.
• Treat an agent like a new hire on probation: narrow duties, full review at first, then sampling.
• The business is legally responsible for what its agents do; design and review them accordingly.