“AI agent” has meant several things in the last eighteen months: an assistant that answers, a copilot that suggests, an automated task on a schedule, or software that plans, acts, and recovers from errors. If you are deciding whether to use one, ask what it actually does.
Here is the definition we use after ten years of delivering automation in customers’ systems: an agent reads incoming work, plans the steps, carries them out using connected tools, checks what happened, and retries or asks for help when something fails.
Those steps are not magic. They depend on working connections, checks, and a clear plan for failures.
Start with four things people often mix up.
An assistant helps through conversation. Used this way, tools such as ChatGPT, Claude, or Gemini answer questions while the person decides what to do next. A product may also offer actions and integrations; check which features are actually set up for your task.
A copilot helps inside a tool you already use. It may suggest an email draft, code, or a design. You decide which suggestions or actions to accept.
A rule-based automation follows steps you set in advance. For example: when a ticket arrives with a certain tag, create a Slack thread and assign it to a team. If the rules do not cover the case, a person takes over.
An agent can choose its next steps from the information in the task. It reads the work, makes a plan, uses tools, checks the result, and tries again or asks for help. You do not have to write a rule for every possible sequence, but you still need limits on what it may do.
That sounds simple until it has to work reliably every day.
The five steps.
- Read: gather the information from an email, ticket, contract, or customer record. Real inputs may be noisy, incomplete, or contradictory. Decide what information is essential before the task can proceed.
- Plan: decide what needs to happen, which tools are needed, and in what order. Unfamiliar information or too many steps can make the model lose track. Keep the task narrow enough to check.
- Act: use an authorised tool to look up a record, draft an email, or move a file. Actions that affect a customer or make a commitment wait for human approval. Each extra tool adds another connection that can fail.
- Check: confirm what happened. Did the draft appear? Did the record save? Did the connected system reject the request? An attempted action is not the same as a completed one.
- Retry or ask for help: if the action worked, move to the next item. If it failed, try again where safe or pass it to a person. Set limits on retries, and include enough detail for that person to understand the problem.
These are the steps a useful demonstration should show, including what happens when one fails.
A supplier’s demo may use clean data, working connections, and a task rehearsed until it succeeds. It does not show how the system handles a long queue or incomplete information unless you test those conditions too.
Everyday use is different. Tools limit the number of requests they accept, access credentials expire, and unfamiliar cases confuse models. Human review is there to catch the mistakes that could otherwise reach customers.
A working agent needs four safeguards. First, a saved record of what it tried, what worked, and what failed, owned by the customer and kept when the model changes. Second, human review before anything affects a customer, contract, or commitment. Third, authorised connections to the customer’s tools. Fourth, a work log and alerts: when did the agent pause, retry, or ask for help, and who was told?
Leaving out one of those safeguards makes the system harder to trust and failures harder to fix.
Agents can fail because information is missing, a tool behaves unexpectedly, or the review queue is so long that people start approving without checking. A failure on Friday can go unnoticed until Monday if the alerts are missing. Corrections also need to be recorded so the same problem does not keep returning.
An agent we ran inside an insurance brokerage last year misread a claim because it missed a policy rider, an addition that changes the policy’s terms. The reviewer caught it. Three cases later, the team had taught the agent what that rider meant. Saving the correction gave later tasks the information they needed. The model alone was not the fix; the retained guidance mattered.
Another agent, inside a law firm, produced drafts that were technically correct but sounded wrong for the client relationship. They were polite and thorough, but missed the partner’s voice. The fix was to give it two years of the partner’s correspondence as a reference for tone, emphasis, and phrasing. The drafts then matched the person more closely.
The pattern is that useful information has to be available when the task runs. The surrounding system needs to keep it and retrieve it.
Here is what those safeguards look like in daily work.
An operations team at a financial services firm has an agent called Morgan. Each morning it reads the overnight email and relevant history, then works out which messages need a reply, a file lookup, or a supervisor. It drafts straightforward replies, checks customer records where needed, and uses the relevant templates. Then it checks whether each draft or lookup worked. If a connection fails or a tool rejects the request, Morgan retries or flags it.
Then Morgan waits. A person reviews the drafts before sending. Most need little attention. When a draft gets the tone wrong or should have gone to a supervisor, the reviewer fixes it and records the correction for future work.
That is the daily process: read, plan, act, check, review, and save what the team learns. After three months, Morgan handles preparation for 80% of the overnight queue before human review. After six months, the person on the team spends 20% of their time on the queue and gets 90% of the credit. The team still owns the customer-facing send.
Not every task needs an agent. A weekly report that always runs the same query may only need a scheduled job. An email to the same three people may need a template. A submitted form may need a simple automatic trigger to pass its details to another system.
An agent is more useful when the work varies and needs background information. Each item may require a different sequence of steps. It still takes time to gather the right guidance, prepare reviewers, and check whether the process is improving.
Ask three questions: what information does this agent need, what must a person approve, and how will we know something has gone wrong? Clear answers are more useful than an impressive demonstration.
Related See the legal intake employee example →