Skip to content

What to expect from AI

AI doesn't know everything already. It knows what you teach it. This page tells you what to really expect before you start.

What it does well

  • Answers the same question 400 times a day, always with the same patience.
  • Answers at 3 AM, on Sundays, or during holidays.
  • Answers in 5 languages without extra costs.
  • Finds the right answer in your documents in a few seconds.
  • Hands over the conversation to a person when needed.

What it doesn't do

  • It doesn't decide for you. It doesn't give discounts, accept returns, or promise delivery dates.
  • It doesn't negotiate. If a customer insists, the bot sticks to the rule you gave it.
  • It doesn't invent offers. If it doesn't find the answer in your content, it says it doesn't know.
  • It doesn't know things that aren't written anywhere. If the August opening hours are only in the owner's head, the bot doesn't know them.

The first month is training

On day one, the bot only knows what we uploaded. Real questions from your customers are always slightly different from the expected ones. Gaps emerge in the first few weeks: this is normal, not a defect. That's why training exists.

What we ask of you

Very little, and at the right moments. Before go-live, we cover the most important topics together: the questions your customers actually ask. Then we go online, and most of the improvement happens there, in real use: the system automatically detects unanswered questions, and we fix them. We ask you for two things: report when you notice a wrong answer, and let us know when something changes in the company (a new product, a promo, a closure).

Conversation quality

A conversational AI doesn't follow a script: the same question can receive answers phrased differently, and that's normal. What must remain the same is the result: the same intent and the same information lead to the same conclusion, using different words.

A few too many confirmations or an occasional rephrasing are part of the game, just like with a new employee. However, if a piece of data already provided is requested again, or if the same request leads to the right result once and not the next, that is an error to report: it can be fixed, see Errors and corrections, where you can also find what is and isn't an error.

What we do

glacom® experts monitor the system, fix technical errors, and supervise important changes. You are not alone: see Human control.

Limits

  • A newly activated bot makes more mistakes than a bot trained for three months. The right comparison is not "the bot versus an experienced employee", it is "the bot versus no answer".
  • If your content is old or wrong, the answers will be old or wrong.

From the app

From the glacom® app, you can read conversations and report answers to be corrected, wherever you are.