Two companies buy the same robot; one gets praise from customers and the other gets complaints. The difference is almost always the content loaded into it, not the hardware or the model inside.
Where the robot's answers come from
A customer-facing AI robot typically draws on two sources.
The model's general knowledge. Broad, but it knows nothing about your company and may be out of date.
The documents you supply. Price lists, service catalogues, procedures, contact details, frequently asked questions. This is the decisive source for anything about your business.
The usual mechanism is that the robot searches your documents for the relevant passage and answers from it. The direct consequence: if information is not in the documents, the robot either says it does not know or, worse, guesses.
This makes content preparation not a side task delegated to whoever is free. It is the part that determines the product customers actually experience.
Preparing documents a robot can use
Organise by question, not by company structure. Documents arranged by department are hard to search. Arranged by the questions customers actually ask, retrieval is far more accurate.
One idea per short passage. Long paragraphs mixing several topics cause the robot to retrieve the wrong part. A passage should answer one question completely.
Avoid internal abbreviations. The robot does not know what an internal product code means unless you explain it.
Mark time-limited information. Prices, promotions, schedules — give them an effective date so it is clear when they need review.
Write what you do not do. Services not offered, areas not covered, work outside scope. This matters as much as what you do offer, because it lets the robot decline correctly instead of guessing.
Start with the thirty most common questions. This yields the largest and fastest return. Everything else can be added over time.
The update process
Stale content is the leading cause of wrong answers, and it is a process problem rather than a technical one.
Assign one named person. Not a department. This person is responsible for the robot's content being correct and current.
Attach it to existing processes. When prices change, when a service is added, when hours change — updating the robot must be a required step within that same process, not a separate task that is easily forgotten.
Schedule a periodic review. Read the whole content set monthly, even when no changes have been reported. Something is always out of date without anyone mentioning it.
Test by asking. The fastest way to find stale content is to ask the robot ten common questions and listen. Ten minutes, and it should be weekly.
Keep a change history. So that when a complaint arrives, you know what the content said at the time.
Handling wrong answers
This will happen. What matters is how the system responds.
Teach the robot to say it does not know. Sounds simple but must be configured deliberately. I do not have that information — let me connect you with a colleague is far better than a fluent guess.
Provide a handover path. A staff-call button, or the robot alerting the desk when it detects an out-of-scope question.
Log every unanswered question. This is the best source for what content is missing, and it accumulates without effort.
Read conversation transcripts. A random sample each week. You will find wrong answers nobody complained about — customers simply left.
Limit scope deliberately. For sensitive topics — firm pricing, professional advice, complaints — configure the robot to hand over rather than answer.
That last point deserves emphasis: limiting scope makes a robot more useful, not less, because customers trust a robot that knows what it knows over one that answers everything.
Voice and manner
This affects customer perception more than most people expect.
Pick a consistent register. Suited to your customers. What works in a café does not work in a clinic.
Shorter than you think. A listening customer has far less patience than a reading one. Three sentences beats ten, with detail reserved for follow-up questions.
Avoid document language. Text written to be read differs from text spoken aloud. Read drafts out loud while writing.
Prepare the opening and closing. How the robot greets and ends determines the impression, and these are the only two lines every visitor certainly hears.
Test with real people. Have a few staff and regular customers talk to it and describe how it felt. Feedback in the first ten minutes usually identifies exactly what to fix.
Frequently asked questions
What does the robot base its answers on?
The model's general knowledge plus the documents you supply. For anything about your company, your documents are decisive — if the information is not there, the robot either says it does not know or guesses.
How should documents be organised?
By the questions customers actually ask rather than by department, with one idea per short self-contained passage, no internal abbreviations, and explicit statements of what the company does not offer.
What is the leading cause of wrong answers?
Stale content. It is a process problem, and the fix is making the robot update a required step inside the existing process for changing prices, services or hours rather than a separate task.
Does limiting scope make a robot less useful?
The opposite. A robot that knows what it knows and hands over on sensitive topics earns more trust than one that answers everything, including some things fluently and wrongly.
More in Reception and event robots and What AI robots are.
