Amigdala Partners
EN
Contact
AI-generated image, labelled AI Generated: in the stock room of a spare parts supplier, a warehouse lead hands a small sealed parcel to an older courier driver standing by a hand truck, with tall shelves of boxes and an open loading door behind them.AI
AMIGDALA / INSIGHTS
← All Insights
INSIGHTS / Perspective

Before an AI assistant answers your buyers, decide which answers you can keep

LISTEN TO THE ARTICLEAI-generated voice
Explore the perspective
THE PERSPECTIVE

When an AI assistant answers a buyer, the answer becomes a promise someone else in the company has to keep. Before launch, sort the questions buyers really ask by whether you can stand behind the answer, and give every firm answer an owner.

← InsightsThe perspective01 / 08
In this article +

Which buyer questions should a customer-facing AI assistant answer firmly, conditionally, or hand over, and who keeps each fact true?

The perspective

An AI assistant that answers buyers' questions can look finished in a demo. It can look up the catalogue and it writes well, at any hour. The real test comes later, when a buyer acts on an answer and someone else in the company has to make it true.

A buyer who reads "yes, that part fits your machine" or "it can be with you on Friday" is likely to read it as a commitment. The warehouse, product management and customer service inherit that commitment, whether or not anyone checked that they could keep it.

So before launch we'd change the question. Instead of asking whether the assistant answers well, ask of each buyer question: can the company stand behind the answer, and who keeps the fact behind it true?

A hypothetical parts supplier

Consider a hypothetical company that sells spare parts for commercial kitchen equipment to restaurants and service technicians. It wants a chat assistant on its website that answers buyers' questions about parts, delivery and returns, before and after they order, so they get answers without waiting for a person.

Four questions come up again and again. Does this gasket fit my oven? Can it arrive by Friday? Can I send it back if it doesn't fit? Will it stop the leak?

The assistant can write a fluent answer to each of them. What the company can stand behind changes from one question to the next.

Fit depends on a compatibility table. If product management keeps it complete and current, the company can stand behind a yes or a no. For a model the table doesn't list, the same question can't get a firm answer. Friday delivery depends on stock, the order cut-off time and the courier, so the warehouse can commit to it only under conditions that change during the day.

Returns follow a written policy that customer service owns: the answer is firm as long as the assistant quotes the current version. Whether the part will stop the leak is a diagnosis. Nobody in the company can promise that from a chat message, and a confident answer would be a guess the company then has to honour or take back.

Sort the questions before you build

We'd start with the questions rather than the assistant. Collect the ones buyers really ask, from customer service emails, call notes and returns, and put each one in one of three groups.

Answer firmly. The company owns the fact, the fact is current and a named role keeps it that way. The fit question belongs here only for the models the compatibility table actually covers.

Answer with a condition and a next step. The fact is real but it changes, or it depends on the order. The assistant says what's true at the time of the answer, what it can't confirm, and when the buyer will get a confirmed answer: for delivery, in the order confirmation that the warehouse sends.

Hand over. The answer would be a judgment nobody has made. The leak question goes to the company's technical support, with the conversation passed on so the buyer doesn't have to explain it again.

An illustrative example, not a measured result. The groups, questions and owners will differ in every business.

Sorting is a decision about the service, and it belongs with the business rather than with the team that builds the assistant. The commercial lead decides each question's group, after the owner of the fact has agreed they can keep it.

Give every firm answer an owner

This isn't about who maintains the assistant. It's about who keeps the fact true for the buyer. For every firm answer, name the role that keeps the fact true and say what happens when it changes. When product management adds a new oven model, it updates the compatibility table before the assistant gives firm answers about that model. When the return policy changes, customer service replaces the text the assistant quotes on the same day.

If nobody will take that on, the question moves down a group. A conditional answer or a handover is better than a firm answer that nobody maintains. And when the assistant can't tell which group a buyer's question belongs to, it hands over.

Check the answers against what happened next

Testing the assistant's replies shows whether they're accurate today. It doesn't show whether the company can keep them. So we'd also take a sample of firm and conditional answers and check each one against what actually happened afterwards: the part that was shipped, the date it arrived, the return that was accepted or refused.

Every mismatch points to a fix. The owner of the fact fixes the source or the process, and the commercial lead decides whether the question changes group. Do this before widening what the assistant covers.

The strongest objection

This can sound like ordinary product-data housekeeping with an AI label, and partly it is. Some of what the exercise finds will be content and process problems: a missing compatibility table or a cut-off time nobody wrote down. Fixing them may make an assistant unnecessary for a while. That's a useful result too.

What the assistant changes is reach and tone. A form or a static page leaves its gaps visible. An assistant fills them with a plausible sentence, in the same confident voice it uses for the facts the company owns. From the text alone, the buyer has no way to tell the two apart, so the company has to draw the line before launch.

Where it isn't worth the effort

Not every question needs this treatment. Opening hours and contact details can go into the first group with little effort, as long as someone is named to update them. A pilot with a few trusted customers can start with a short list of firm answers and hand over everything else. Prices, discounts and contract terms always go to the handover group: they need commercial authority, and we'd keep them with people until that authority is defined.

What the platform doesn't decide

An assistant platform may ground answers in your documents, show the page it used and pass a conversation to a person. Those features help. They don't know whether the document is still true, which role keeps it true, or whether the warehouse can honour the date the assistant has just given. Those are choices for the business, not the platform.

Before launch, take the questions your buyers ask most, put each one in a group, and write a role next to every firm answer.

FURTHER PERSPECTIVES

Keep exploring.

AI-generated image, labelled AI Generated: before a meeting, a man in a light blue shirt places a printed page on a wooden table next to a printed dashboard, while a woman in a navy jacket, glasses in hand, stops to read it.AI
Perspective

When an AI number improves, check what changed underneath it

An AI workflow runs on services that suppliers update on their own schedule, sometimes including how use is counted. We propose a company change log covering your changes and theirs, and fixed test requests rerun after every change, before anyone reads the trend.

Amigdala Team
Read more
AI-generated image, labelled AI Generated: in an office overlooking a factory hall, three colleagues holding a laptop, a folder and a binder stand by the desk of a senior engineer, who looks up from her drawings.AI
Perspective

Before you trust an AI pilot, ask how much time the expert had

A pilot result is only as good as the review behind it, and the people who can review it are few and can be wanted by several projects at once. We propose budgeting their time before funding and comparing proposals person by person. Every pilot result then comes with the review time the expert actually had.

Amigdala Team
Read more
AMIGDALA / CHALLENGE EXPLORER