Guide
AI Agents for Business: What They Actually Run
By Ryan Lanyon, Founder and AI Systems Architect, Nexus Digital
Published 16 November 2026
What makes an agent an agent
An agent, four pieces
Not a prompt, not a workflow
- Goala job, not a prompt
- Toolspicks and uses them
- Memorystate across steps
- Oversighta person on the loop
- Accessscoped to the job
An AI agent inside a business is a program that holds its own state, picks between tools on its own, and writes to a real system of record. It is not a chat window and it is not a one step automation. Most of the value shows up when an agent owns a small, repeated job end to end and reports back.
What is an AI agent?
A careful definition first. An agent has a goal written down, a set of tools it may use, memory it keeps between steps, and a human who owns what it does. Strip any one of those four and you are back to a workflow or a chatbot, both useful, neither an agent.
We only call something an agent if it holds state across steps, picks between tools on its own, and writes to a real system of record.
The useful shape is narrow. One agent, one job, one named owner. Agents built to do everything are the ones that fail loudly inside three months, because nobody can explain what they did on a Tuesday afternoon.
How is an agent different from an automation?
An automation runs the same steps every time, in the same order, without judgment. It is a recipe. If the input shape changes, the automation breaks, which is loud and easy to fix.
An agent reads the situation, decides which tool to use next, and can handle inputs it has not seen before. It is a worker, not a recipe. When it goes wrong, it goes wrong in a different way: the same job runs five different ways in five days, which is why the oversight line matters later in this guide.
| Property | Automation | Agent |
|---|---|---|
| Steps | Fixed | Chosen at runtime |
| Memory between runs | Usually none | Yes, by design |
| Handles new inputs | No, breaks | Yes, usually |
| Fails how | Loudly, on a bad input | Quietly, five ways in a week |
| Fit | Repeated work, same shape | Repeated work, variable shape |
Most teams will want both. The automation does the clerical half, the agent does the judgment half, and they talk to each other.
What jobs can an agent run today?
Four shapes work well enough to put into production inside a small or mid business. We do not hand anything else to an agent yet.
- Inbox triageLabel, draft, file, never send on its own
- Daily briefPull the numbers, write the morning message
- Repeated writingReplies in your voice, always draft only
- DispatchRoute an incoming job to the right person
What these share: a job the business already does by hand, a clear right answer most of the time, and a cheap wrong answer. We do not hand an agent anything where a wrong answer is expensive.
When should a business hire its first agent?
In our experience, the first useful agent inside a small business is almost never customer facing: it is the one that cleans up inbox triage or handles a weekly report nobody owns.
Three conditions have to be true before an agent earns its keep:
- 1The job repeats. Daily, weekly, or every few days. One-off work pays better as a prompt in a chat window than as an agent.
- 2The right answer is checkable. Somebody can read the output in under a minute and say whether it is correct. If verifying takes longer than doing, the agent is not yet worth it.
- 3A wrong answer is survivable. A draft, not a send. A proposed status, not a locked one. A file in a folder, not a withdrawal from an account.
If one of the three is missing, the honest answer is not yet.
What an agent needs around it
An agent is only as good as its guardrails. We insist on four of them before we run one in production.
What sits around every agent
One agent, one job
with four controls around it
- Goalone job, written down
- Toolsscoped to the job
- Humanon every write
- Logsevery call, kept
Goal. One sentence that says what the agent is for. Not a list.
Tools. The agent only gets what it needs, not what the account can reach. The service account for a daily brief does not need write access to accounting.
Human on every write. Every change the agent makes to the outside world runs through a draft a human approves, or the agent is not allowed to make it. Reads are cheaper than writes.
Logs. Every prompt, every output and every tool call is kept long enough to answer a question somebody will ask six months later. Without the log, you are hoping, not operating.
The Australian paper trail covers agents the same way it covers the rest of a business's technology. The Australian Privacy Principles govern the data an agent reads and writes, especially anything personal. The AI Risk Management Framework is a reasonable neutral reference for the control categories a review will look for, used alongside the OAIC rules rather than instead of them.
Who owns an agent inside the business
One named person on the leadership team per agent. Their job is not to run the thing every day, it is to notice when it is going sideways and to be the one who stops it. An agent without a named owner is a volunteer, and volunteers quit.
The owner answers four questions in any review: what is this for, who approves its writes, where are its logs, and what is the kill switch. If all four are not written down, the agent is not ready to run.
The two mistakes we see most
Mistake one, agents built for prestige rather than for repeated work. Teams build the agent that sounds exciting in a meeting, often one that talks to customers or drafts something strategic. Both are a bad first agent. The first useful one is boring, internal, and shaves an hour off a job a specific person does every day.
Mistake two, letting an agent write without a draft step. The quickest route to a bad week is an agent that sends, posts or pays on its own. A draft step costs almost nothing and catches the odd confident wrong answer. Keep the draft step until you have six months of logs that say the agent earned the right to skip it, and only for the lowest stakes writes even then.
How to run your first agent this quarter
Pick a job that one person does every day, usually in under thirty minutes, and that produces an output somebody else reads quickly. Write the job down. Build the agent around the job, not the other way round. Run it in draft mode for two weeks, read every output, then decide whether to keep it, tune it, or put it back as a workflow.
- Pick the job
- Write it down
- Build in draft
- Read outputs
- Decide
A proper review of which agent to run first sits inside the paid Strategic Assessment, which scores the business against five layers and names the one or two workflows worth building first. If you already know the job and want it built, the AI implementation guide covers what the Implementation includes, agents among them.
Book the Strategic Assessment when you want the whole picture written down.
Sources
FAQ
We call a system an agent when it holds state across steps, chooses its own tools, and writes to a real system of record. A chat window is not an agent. A one step automation is not an agent. The useful version is narrow: one agent, one job, one named owner. Agents built to do everything tend to fail inside three months, because nobody can explain what they did on a Tuesday.
When a job repeats daily or weekly, the right answer is checkable in under a minute, and a wrong answer is survivable. If all three are true, an agent earns its keep. If one is missing, the honest answer is not yet. Most first agents are internal, not customer facing: inbox triage, a daily brief that pulls numbers together, or a weekly report nobody currently owns. Customer agents work, but they are not the first one to build.
An automation runs the same steps every time, in the same order. An agent reads the situation, picks tools, and can handle inputs it has not seen before. An automation breaks loudly when the input shape changes, which is useful. An agent goes wrong quietly, in five different ways in a week, which is why human review on every write matters. Most businesses want both: the automation for clerical repetition, the agent for the judgment half.
Four controls around every agent we run. A single written goal. Tools scoped to that goal rather than to the account the agent runs under. A human approval on every write to the outside world, as a draft step. Logs of every prompt, output and tool call kept long enough to answer a question somebody will ask six months later. Strip any one of those and you are hoping, not operating.
Want this against your own numbers? The faster version is a call. We map where the hours are going in your business, then show you what an implementation would return.
Want the next one in your inbox? Join the weekly AI newsletter and get the free AI security guide.