Look at the job you want done. If the steps are predictable, a conventional workflow may be enough. If the work involves interpreting messages or deciding what to investigate next, AI may help—with limits that match the risk.
A customer sends a long email about a maintenance problem. Someone reads it, works out which service is needed and passes it to the right team. Could you automate that? Probably. Does the whole process need an AI agent? Not necessarily.
That distinction matters when you're paying for AI workflow automation. A system that follows five sensible rules can be more useful than an autonomous assistant that needs constant supervision. The right choice depends on where people currently spend their time and why the work gets stuck.
Some processes need clear rules, not more AI
Take a purchase request. If approval depends on department and order value, the route is already known. Software can assign the manager, send a reminder and keep a record of the decision. There's little benefit in asking a model to work out a route you can write down.
Before discussing technology, check whether those rules actually exist. Who can approve a higher amount? Who covers an absent manager? What happens to an incomplete request? If the business hasn't agreed on the answers, adding AI won't resolve the disagreement.
Let AI handle the untidy part
The useful AI step might be much smaller: reading that maintenance email and suggesting a service category, location and short description. A person can check uncertain details, while the rest of the process follows normal scheduling rules.
Keep the original message alongside the summary. Staff need to see what the customer said, especially when an extracted detail looks wrong. The scheduling system should still check availability and service entitlement before confirming a visit. A good summary isn't permission to make a booking.
So when does an agent make sense?
Consider a support enquiry where the next step depends on what you find. You might need to look up an order, check a delivery record and ask for missing information. Here, choosing the next action is part of the work.
Anthropic's guidance distinguishes fixed workflows from agents that choose their own steps. It's a helpful distinction regardless of which provider you use. But start small: one investigation task, a limited set of records and a clear point at which the assistant stops and hands over.
| The job | A sensible starting point | What still needs checking |
|---|---|---|
| Route a known approval | Rules and assigned owners | Who has authority |
| Sort differently worded enquiries | AI suggestions inside a workflow | Uncertain classifications |
| Investigate an incomplete case | A tightly scoped agent | Tool access and stopping rules |
Decide what the system cannot do on its own
Reading an account and changing it are different permissions. So are drafting a quotation and sending it. Make those distinctions explicit before connecting an assistant to live business systems.
For actions that need approval, show the proposed recipient, amount or record change before execution. Keep a record of the decision. Also ask what happens after a timeout: did the action fail, or did it happen without the confirmation coming back? Retrying blindly could create a duplicate.
A pilot should answer a business question
Choose one process and measure how it works today. Then test the proposed automation with straightforward cases, incomplete requests and actions it must refuse. Look at correct completions, corrections and review time—not just how many runs finished.
Include usage fees, maintenance and supervision in the cost. If staff spend longer checking the output than doing the original job, the pilot has told you something useful. You can narrow the task or choose a simpler approach before making a larger commitment.
What to bring to the first discussion
You don't need a technical specification to begin. A few anonymised examples, your existing rules and a description of the desired handover are a useful start. If the work involves sales leads, our CRM requirements guide can help clarify ownership first.
At a Gligx consultation, the aim is to work out which steps are worth automating and how you'll judge the result. Whether the eventual solution needs an agent should follow from that discussion.
Sources and further reading
References checked on 23 October 2026.
Frequently Asked Questions
Does my business need an AI agent or a workflow?
Start with the job. If the steps and decisions are predictable, a rule-based workflow may be enough. An agent is worth exploring when choosing the next step is part of the work.
Can AI help without running the whole process?
Yes. It could suggest a category or summarise a message while ordinary software handles routing and a person checks uncertain details.
What should still need a person's approval?
Base that decision on the consequences. Payments, binding quotations, access changes and external commitments usually need explicit review rather than automatic execution.
How will we know whether the pilot is useful?
Compare correct completions, corrections, review time and total running cost with the current process. Counting completed runs won't show how much work staff had to redo.