AI agents do best with repeating work that starts from messy input: sorting a shared inbox, qualifying leads, keeping a CRM up to date, processing documents and preparing quotes. What these jobs share is free text, several systems and a limited set of possible actions. A person usually approves what reaches customers or money.
What kinds of work do agents handle?
These are types of work, not customer stories. Your version will depend on your systems and your data.
| Work | What the agent does | Where a person stays |
|---|---|---|
| Shared inbox | Reads the mail, finds the customer, sorts it, drafts a reply | Approves replies, at least at first |
| Lead qualification | Reads the enquiry, looks up the company, scores it, routes it | Sales makes the call on who to pursue |
| CRM upkeep | Turns mails and call notes into updates and tasks | Checks changes to deal value and stage |
| Invoices and documents | Pulls out the data, matches it, flags differences | Approves exceptions and payments |
| Quotes | Reads a job description, gathers details, prepares a draft | Checks price and terms before sending |
| Reports | Gathers figures from several systems, writes the summary | Reads before it's sent on |
| Customer questions | Answers from approved sources, escalates the rest | Handles complaints and anything unclear |
What does one of these look like step by step?
Take the shared inbox. A mail arrives from a customer asking to move a delivery.
- The agent reads the mail and works out what's being asked.
- It looks up the customer and the order in your systems.
- It checks the delivery rules and which dates are possible.
- It drafts a reply and prepares the change.
- A colleague sees both, with the source data next to them, and approves.
- The agent sends the mail, makes the change and logs what it did.
Steps 1 and 4 need AI. Steps 2, 3 and 6 could largely be plain rules and connections. That's typical. A good agent is mostly ordinary software with a little judgement in the right places.
What makes a job a good candidate?
- The input is free text: mails, PDFs, notes, forms with open fields
- The next step depends on the content
- The information sits in more than one system
- The possible actions are limited and can be listed
- You can tell a good result from a bad one
- A mistake can be caught before it does damage
If most of these are true, it's worth a closer look. Then score it using how to choose a first process.
What looks like an agent job but isn't?
- Fixed data transfers. New form entry goes to the CRM. Rules do this better.
- Standard confirmations. The same mail every time needs no AI.
- Decisions nobody can explain. If your team can't say what a good decision is, an agent can't either.
- Rare work. A few times a year won't repay the setup.
How do you judge examples you see elsewhere?
Be sceptical of impressive stories, including vendor ones. Ask what the agent does alone and what a person still checks. Ask what happens with the cases it can't handle. Ask how the result was measured and against what. A demo shows the happy path. Your real work is mostly the other paths, so test on your own messy input.
At 5DAYS, Business Agents are aimed at sales processes, CRM automation, document processing and operational workflows. Simple lead handling, data processing and reports can often be done with a smaller Automation Sprint.
