AI Without the Hype · Part 6 of 10
An AI assistant that helps your employee prepare a reply and an AI assistant that sends the reply itself are two different purchases.
In the first arrangement, someone in your business checks the answer before the customer receives it. In the second, the system is speaking on your behalf. That difference deserves more attention than the quality of the demo conversation.
For a first customer-service project, I would usually test employee-reviewed drafts. It gives you a chance to learn what the tool gets wrong while an informed person still controls the response.
Pick a narrow group of questions
Suppose you run an equipment rental business. Customers ask about opening hours, collection instructions, availability, pricing, and whether a piece of equipment is suitable for a particular job.
Those questions do not all need the same treatment. Opening hours can come from an approved document. Current availability needs reliable current inventory information. Suitability may require a qualified employee to understand the task and explain operating limits.
Do not turn all three into “the chatbot handles questions.” Decide which questions are in scope and which go directly to a person.
If most inquiries are about a handful of stable facts, clearer website information and standard replies may solve enough of the problem without a conversational system.
Use information the business actually maintains
An AI-generated answer about your refund policy is only useful if it reflects the policy you currently apply. Give the system approved source material and name someone to maintain it.
For the rental business, this might mean a controlled set of FAQs and instructions. Dynamic facts, such as whether a machine is available tomorrow, need a trustworthy live source or a handoff to staff. A general answer should never be presented as a confirmed reservation.
NIST’s Generative AI Profile describes risks involving false information and the way people interact with AI systems. In customer service, a practical concern is a plausible answer that the customer reasonably treats as a promise.
Test specifically for invented prices, availability, deadlines, and exceptions.
Separate answers from actions
A tool might be able to explain the booking process without being allowed to create a booking. It might draft a refund response without issuing money.
List the permitted actions individually. Can it send emails, change an address, cancel an order, issue credit, or modify a record? What approval does each action require? What identity checks protect customer-specific information?
Giving a system more access because “it might be useful later” adds responsibilities before you have established the need. Start with the actions the chosen task actually requires.
For a first trial, a read-only drafting arrangement can be easier to evaluate than a system that changes customer records while you are still learning its failure modes.
Make the handoff useful
A customer should be able to reach a person without arguing with the tool. Make the route clear. Carry the relevant conversation forward so the employee does not ask the customer to start over.
Define what happens outside business hours. If someone requests a person at 9 p.m., tell them how the request is recorded and when your business actually expects to respond. Do not promise immediate assistance unless you provide it.
Set escalation triggers for complaints, disputes, uncertainty, requests outside scope, and anything requiring qualified advice. Tell customers when they are interacting with an automated assistant as a straightforward operating practice.
Try to make it fail before customers do
Use fictional cases to test a changed policy, an unavailable item, conflicting information, and a customer asking for an exception. Ask what happens when the source system is down.
Include a customer message that tells the assistant to ignore its instructions or reveal private information. Treat customer-provided text as material to handle, not authority to change the system’s permissions. Have the provider explain how that boundary is implemented and tested.
Time the employee’s review and correction work. Track whether customers get an accurate resolution, how often they need to repeat themselves, and how often staff must undo a response. A low handoff count can mean the assistant is useful, or that customers cannot reach anyone. Look at the conversations.
When direct replies might make sense
Once a narrow workflow has been tested, you may decide that certain routine answers can be sent directly. Make that a specific decision with approved sources, limited actions, monitoring, a responsible owner, and a way to switch it off.
The question is whether your business can stand behind the answers it gives. Set the scope accordingly.
Sources
NIST: Generative Artificial Intelligence Profile. NCSC: AI and cyber security. The rental-business scenario is hypothetical.