AI Without the Hype · Part 2 of 10
A customer fills out a form. Someone copies the details into a spreadsheet. Someone else sends a standard acknowledgment. Later, the owner asks why this still takes so much time.
An AI tool might be able to help. A straightforward connection between the form and the spreadsheet might solve the problem too.
Before buying anything, separate three questions: does the process need to change, can the steps follow fixed rules, and does any step need help interpreting variable information?
Those questions point you toward different kinds of solutions.
When a clearer process is enough
Consider estimates that never get followed up. If no one has decided who owns follow-up or when it should happen, the first job is to make those decisions.
A shared queue and a daily review might be enough for a business sending a handful of estimates each week. At a higher volume, a system that assigns tasks and sends reminders may be worth the investment.
AI does not relieve you of deciding who is responsible. A tool can propose a workflow, but someone in the business still has to agree to it and run it.
If you are unclear about the underlying problem, start with these five questions before buying business software.
When fixed rules do the job
If you can describe the step as “when this happens, do that,” ordinary automation deserves a look.
For example, when an approved inquiry form arrives, create a record, assign the territory using the ZIP code, and send the acknowledgment already approved by the business.
No language model needs to compose a new acknowledgment every time. The message can be a template. The territory can come from a maintained lookup table. Missing or unrecognized ZIP codes can go to a person.
This approach still needs testing, error handling, and someone to maintain it. Its advantage is that the intended behavior is explicit. You can inspect the rule when something goes wrong.
When AI has a useful role
Now change the example. Inquiries arrive as free-form emails. One customer writes a paragraph about a broken gate. Another sends a short message describing a repair and asks about availability.
A language model could help extract the requested service, identify missing details, and draft a response for review. The wording varies, so a rigid form or keyword rule may not capture everything you need.
That is a potential use case, not a guarantee that a particular tool will handle your messages well. Test it against the actual variation in your work. Include messages with two requests, uncertain locations, and something your business does not offer.
Give the tool a way to flag uncertainty. An empty field marked “needs review” is more useful than a location it guessed.
You can combine the approaches
For our inquiry process, an AI step might propose a service category. A person approves it. Fixed rules then create the appropriate task and send the correct template.
That division lets you choose where judgment is needed and where a clear rule should control what happens next.
Anthropic distinguishes workflows that follow predefined paths from agents that choose their own steps. That distinction is useful when a vendor uses “AI agent” to describe everything from a drafting assistant to a system allowed to take action across your accounts.
Ask what it actually does. Does it suggest a category? Update a record? Send a message? Decide what to do next without a person? Each is a different level of responsibility.
Compare the whole job
Ask a vendor to demonstrate one ordinary case, one incomplete case, and one exception. Watch how the system handles each. Find out who sees failures and how they recover the work.
Compare the total effort of the options: setup, subscriptions, employee review, corrections, and maintenance. A cheaper rule-based solution can become cumbersome if the inputs vary too much. An AI solution can be unnecessary if the work is already structured and predictable.
The answer should follow the requirements of your process. There is no prize for using a more complicated tool.
A useful question to take into the demo
“What part of this workflow needs AI, and how would the solution work without it?”
A capable provider should be able to explain the role AI plays, where fixed rules take over, and how uncertain results are handled. If that explanation stays vague, you are not ready to judge the purchase.
Source
Anthropic: Building effective agents. Workflow examples and buying recommendations are illustrative.