Does this task need AI, or would simple automation do the job?
Use simple automation when you can write down the rule that decides the next step. Use AI when a step requires someone to interpret words, pictures, or documents that can vary from one request to the next. A single job can use both. For example, AI can read an email and suggest which team should receive it. A fixed rule can then deliver it to the right queue. If the email asks for an exception that only a manager can grant, the manager still makes that decision.
What decision are you asking the software to make?
Take the last request your team finished. Open the original message and the record of what happened. Somewhere between receiving the request and closing it, a person made a choice. What did they look at? Could someone else reach the same answer by following a written instruction? Ask the person who did the work to show you the detail that settled the question. That detail will tell you more about your options than a list of AI features.
Suppose someone submitted a form, selected “Change my appointment,” and supplied a booking number. The destination was already decided when the person selected the category. Software can check that the number belongs to an open booking and create a task for scheduling. You can write down the expected result before running the rule. If the booking is closed, a different written rule can send the request to the person who handles closed bookings.
Now suppose the same person emails, “Thursday is going to be tricky. Is there anything later in the week?” A colleague understands that someone probably wants a change. The word “appointment” never appears. A system that interprets language can suggest a category and quote the part of the message that supports its choice. It still needs a booking number and a clear instruction about what it may do next.
Could a better form solve the problem?
Before paying for interpretation, ask whether the missing detail is easy for the sender to supply. If staff already sign in to submit internal requests, a field with four familiar choices may remove an entire sorting job. Put the choices in words the sender recognizes. “I need access to a file” is easier to select than the name of the department that grants access. The department can change later without forcing people to learn a new form.
A form can also create work. Imagine a returning customer replying to a long email exchange. They have already given you the booking reference and explained the change. Sending them to a form may produce a second request while the original email still sits in the queue. Look at both sides of the transaction. A change that saves the office two minutes by asking every customer to retype five fields deserves a second look.
Try this with ten recent requests. Hide the destination each request eventually reached. Ask a colleague to sort them using only the fields your software already captures. When they need to open the message, record what they were looking for. You might find that adding one field would solve eight requests and leave two for a person. You might instead find that customers describe the same need in dozens of ways. Those findings point to different investments.
What would a mixed workflow look like?
Consider a shared inbox for a business that takes bookings. One morning it receives three messages. The first comes from the booking form and includes a category, a reference, and a requested date. The second is a short email asking to move a visit. The third asks to move a visit and waive the usual charge. All three may land on the same screen today. They need different treatment when you hand off the work.
For the form submission, a rule checks the reference, looks up the booking, and puts the request in the scheduling queue. It can attach the requested date without promising that the date is available. A scheduler sees the request in the same place they use today. If the reference fails the check, the rule sends the item to a named person who can correct it. There is no reason to ask a language model to read a date that the form already supplies.
For the short email, ask AI for a limited result. It can propose “change booking,” extract any reference the sender supplied, and show the phrase that led to that category. Software can check the proposed reference against the booking system. It can also check that the proposed category is one your team accepts. If two bookings match, the item needs a person or a question for the customer. A plausible choice between two people with the same surname is still a guess.
For the fee request, preserve both parts of the message. Someone can start looking for dates while the person authorized to waive a charge reviews the exception. Give that person the original request and the applicable terms. A customer who says the fee was waived last time has given you a fact to check. They have not granted the system permission to waive it again. Closing the item as soon as it reaches scheduling would lose half the job.
Anthropic describes routing as one use for a model inside a predefined workflow. The practical question for the inbox is whether interpreting the message improves the handoff enough to justify the extra step. You can test that without giving a model the ability to reschedule a customer or approve a charge. Read Anthropic’s guide to workflow design.
Where do simple rules work best?
Look for steps that rely on an exact value or an agreed calculation. A due date comes from a signed agreement. A total comes from quantities and approved prices. A reminder goes out when a record reaches a specified age and still has a specified status. The same inputs should produce the same result, and the owner should be able to explain why. Existing spreadsheet formulas, form settings, and software integrations may already do the job.
Write the rule with an example on each side of its boundary. If a reminder is due after five business days, what should happen on the fourth day? Who supplies the holiday calendar? If a customer replies moments before the reminder runs, which record tells the system to stop? These questions can reveal unfinished business decisions. Introducing AI will not settle whether your office counts a holiday as a working day. Someone with authority has to decide that first.
One more test is whether you can remove the step altogether. If a team copies a status from one system to another solely to prepare a weekly list, check whether the first system already offers that list. The lowest-maintenance answer may be a saved report. Find out who uses the second copy before retiring it. A field that looks redundant to one person may start another team’s work.
When does AI earn its place?
AI is worth a trial when people spend time reading variable material to produce a result you can check. Think of summarizing a long handoff for the next shift, finding the requested change in a message, or drafting a reply from a current policy. For each candidate, name the evidence a reviewer can see. A summary linked to the original notes is easier to check than a confident sentence with no indication of where it came from.
Give the trial a fair comparison. For inbox sorting, compare AI with the best rule you could reasonably maintain. Include an ordinary busy day and some awkward messages, such as “Thanks for sorting the booking. I still need a receipt.” Count how many items reach the right person, how many need another handoff, and how long someone spends checking them. A tool that recognizes the word “booking” but misses that the booking is settled has failed that example.
Consider the consequence of a wrong answer. Sending an internal note to the wrong colleague may cost a few minutes. Sending private account details to the wrong customer is a different decision. Even if both actions begin with the same classification, they need different permissions. Start by letting the software prepare the internal item. Keep delivery to the customer behind a separate check until you have evidence that the whole process works.
What should you take into a first automation discussion?
Bring one completed request with the customer details removed, the rule your team applies, and one example where that rule runs out. For each step, identify who supplies the input, what a correct result looks like, and who can resolve an exception. Mark the first place a colleague has to read something and interpret its meaning. That is a candidate for AI. The exact lookups and transfers around it are candidates for ordinary automation.
Then ask for a demonstration using your examples. Change one important detail halfway through. A switch from “send me a copy” to “send my colleague a copy” should change the checks before anything leaves the business. Watch where the new detail appears, what the software does with it, and who can stop the job. By the end you should be able to point to the rule, the interpretation, and the decision you still own. To price the option you have chosen, use the worked example for a recurring task.