Part 57 - AI-Buddy..... How do we build digital customer advisors?
From search fields and filters to a conversation that helps the customer find the right
Imagine you walk into a physical specialty store.
You are standing in front of a wall of products.
An employee comes over and asks:
"What can I help you with?"
You might answer:
"I need a jacket for hiking."
A good employee doesn't immediately point to the most expensive jacket in the store.
She asks further.
When should you use it?
Summer or winter?
How long are your walks?
Is low weight important?
Will it withstand heavy rain?
Do you already have a midlayer?
What do you want to use?
Through the conversation, something important happens.
The employee helps you translate a need into criteria .
She then connects the criteria to the knowledge she has about
the products.
It is this basic mechanism that we are trying to recreate when we build a digital customer advisor.
The customer does not necessarily need to know the product name.
She doesn't need to understand the category structure.
She also doesn't need to know the technical terms.
She can start with her own language:
"I need a light jacket that will keep me dry in the mountains."
The digital customer advisor's task is to understand the need, ask relevant questions, use the company's product knowledge and help the customer find relevant alternatives.
Artificial intelligence makes this type of dialogue far more interesting than before.
But a language model alone is not a good customer advisor.
It needs access to correct product information.
It must know its limitations.
It must know when to ask further.
And the business must decide which tasks it actually needs to perform.
get to perform.
For small and medium-sized businesses, this can open up a particularly
interesting opportunity.
The expertise that was previously only available when
experienced employee stood behind the counter, can be made more accessible digitally – even when the employee is not present .
Magne & the Friend
Magne
Friend...
Now you're going to the store.
The buddy
Finally.
What should I sell?
Magne
Nothing.
The buddy
This is going to be a strange store.
Magne
Heh heh.
You are supposed to help the customer.
The buddy
I like that better.
Magne
The customer comes in and says:
"I need a new laptop."
What do you answer?
The buddy
"Here are 184 laptops."
Magne
You're fired.
The buddy
Already?
Magne
That's what the search field could do.
A counselor must ask questions
The buddy
Okay.
Let me try again.
What will the customer use the laptop for?
Magne
Studies.
The buddy
What types of programs?
Magne
Regular word processing, web browser and video conferencing.
The buddy
Will the customer wear it a lot?
Magne
Yes.
The buddy
How important is battery life?
Magne
Very important.
The buddy
Will the customer play demanding games?
Magne
No.
The buddy
What is the budget?
Magne
Maximum 12,000 kroner.
The buddy
Now we are starting to have a basis for decision-making.
Magne
And the customer didn't need to know how much RAM or what processor to look for.
The buddy
Exactly.
She told me what she needs .
From customer language to product data
Magne
But now comes the difficult part.
How do you know which products are suitable?
The buddy
I need product information.
Magne
Again?
The buddy
Heh heh.
We've spent several articles on this for a reason.
I must be able to connect customer needs to relevant product features.
Magne
So:
"I'll carry the PC every day"
may mean that weight becomes important.
The buddy
Yes.
"I need it for the whole study day" makes battery life relevant.
"I'm just going to use regular office applications" affects how much performance is needed.
And the budget limits the options.
Magne
Then something interesting happens.
The customer speaks needs .
The product catalog speaks attributes .
The buddy
And the digital customer advisor tries to build a bridge between them.
But the advisor must be able to explain why
Magne
Then you will find three products.
Can you just show them?
The buddy
I should do more than that.
Magne
What then?
The buddy
Explain why they are relevant.
For example:
"This model is the lightest of the three and is a good fit if you carry your PC a lot."
Or:
"This one has a longer stated battery life, but weighs more."
Magne
It's much better than:
“We recommend product B.”
The buddy
Yes.
A recommendation becomes more useful when the customer understands the rationale and trade-offs .
A good advisor should not pressure the customer
Magne
But the online store wants to sell.
Shouldn't we program you to recommend the product with the highest margin?
The buddy
Then we have a trust problem.
Magne
So the most expensive isn't automatically the best?
The buddy
Of course not.
If a less expensive model meets the customer's needs better, the advice should be able to say so.
Magne
Even if the business makes less profit on that particular purchase?
The buddy
If the goal is long-term trust and a good customer experience, the right advice can be more valuable than the maximum value of a single transaction.
Magne
So we need to decide what the advisor should optimize for.
The buddy
That is a very important question.
The advisor must be able to say "I don't know"
Magne
What happens if the customer asks about something you have no information about?
The buddy
Then I shouldn't come up with an answer.
Magne
But generative AI likes to answer.
The buddy
Therefore, the system must be built with boundaries.
If the knowledge base does not provide a certain answer, the advisor should be able to say:
"I don't have enough information to answer this with certainty."
Magne
And maybe send the customer on?
The buddy
Yes.
To an employee.
Magne
It almost sounds old-fashioned.
The buddy
It's called good customer service.
Heh heh.
People must be part of the solution
Magne
So the goal isn't to make sure the customer never talks to a human?
The buddy
No.
The goal should be to help the customer as best as possible.
Some questions are simple and repetitive.
Others require experience, judgment, or information the system does not have.
Magne
Then the digital advisor can handle anything...
The buddy
...and humans take over when human expertise is actually needed.
Magne
It can also improve customer service.
The buddy
If it is implemented well.
The employee may skip some of the easiest repetitive questions and spend more time on the difficult ones.
The digital advisor needs a knowledge base
Magne
Now I'm going to guess what you're going to say.
The buddy
Run.
Magne
We can't just install an AI and think we have a customer advisor.
The buddy
Absolutely right.
Magne
We need:
product data,
product descriptions,
professional knowledge,
buying guides,
FAQs,
delivery information,
return,
warranty...
The buddy
...and clear rules for what the system can and cannot do.
Magne
So everything we've built through this section starts to come together here.
The buddy
Yes.
The digital customer advisor is essentially an interface to the knowledge the business has already built .
What about small online stores?
Magne
But now this quickly sounds expensive and complicated.
What about the little specialty shop?
The buddy
It doesn't have to start with the world's most advanced AI advisor.
Magne
Where does it begin?
The buddy
With the most common questions.
Magne
As we have already talked about.
The buddy
Yes.
What do customers ask before they buy?
What criteria do they use?
What products are they comparing?
What does the experienced employee usually ask in return?
Magne
So we can actually design the consulting conversation before we build the technology.
The buddy
We should.
The technology should come after understanding the task .
We can start with one category
Magne
So we don't need to connect the advisor to 20,000 products on the first day?
The buddy
I wouldn't do it.
Magne
What would you do?
The buddy
Choose one product category where:
customers need guidance,
the business has good professional knowledge,
the product data is reasonably good,
and it is possible to control the quality.
Magne
Barley.
Test.
Leather.
The buddy
Improve.
Magne
And then expand.
The buddy
Yes.
It's often a better AI strategy than trying to automate the entire business on Monday.
How do we know if the advisor is good?
Magne
A chatbot can look impressive.
The buddy
It is not a good measure of quality.
Magne
What should we measure?
The buddy
For example:
Does the customer receive relevant suggestions?
Are the facts correct?
Does the customer understand why the products are recommended?
Does the conversation lead to fewer questions?
Does the customer find the product faster?
Are fewer products being chosen on the wrong basis?
When the system doesn't know, does it handle it correctly?
Magne
So we need to measure the quality of help , not how human the chatbot seems.
The buddy
Exactly.
The advisor can also teach the business something
Magne
Wait a minute.
If a thousand customers talk to the advisor...
Then we suddenly have quite a lot of insight.
The buddy
Potentially, yes – if the data is handled legally and responsibly.
Magne
We can see which questions are recurring.
The buddy
What features do customers prioritize?
Which products are often compared.
Where the advisor lacks knowledge.
Where customers become uncertain.
Magne
So the advisor cannot just use the business's knowledge.
It can also help the business discover what knowledge it lacks .
The buddy
There we get an interesting circle.
From consulting to learning
Magne
Let me try.
The business builds product knowledge.
The digital advisor uses the knowledge.
Customers ask questions.
The questions reveal new knowledge needs.
The business improves knowledge.
And the advisor gets better.
The buddy
There you have it.
Magne
Then this is not just a chatbot.
The buddy
No.
If done well, it becomes part of a learning knowledge system around the customer .
Academic specialization
A digital customer advisor is more than a chatbot
The terms chatbot , AI assistant , shopping assistant , and digital advisor are often used interchangeably.
It is useful to distinguish between them.
A simple chatbot can be built to answer a limited set of questions.
An AI-based customer advisor may have a more comprehensive task:
to understand the customer's expressed needs,
ask follow-up questions,
find relevant information,
connect needs to product features,
present alternatives,
explain differences,
and help the customer further when the system itself cannot solve the task.
The decisive factor is therefore not whether the solution has a chat window.
The crucial thing is what task it actually solves for the customer .
Consulting begins with needs – not products
Traditional product recommendations are often based on products.
Customers who viewed this also viewed this.
Customers who bought this also bought this.
Such mechanisms can be helpful, but they are not necessarily counseling.
Consulting requires a greater understanding of what the client is trying to achieve .
A digital customer advisor should therefore be able to work from:
needs → criteria → relevant characteristics → products → explanation.
Not only:
product → similar product.
The dialogue must have a purpose.
A digital advisor should not ask questions just to seem conversational.
Each question should reduce uncertainty or improve the recommendation.
If the customer wants running shoes, relevant information could be:
substrate,
run amount,
experience level,
preferences,
and any other criteria that are actually relevant to the product selection.
Which questions should be asked must be determined by expert knowledge of the relevant product category.
Thus, conversation design also becomes knowledge design .
The advisor needs reliable sources
A language model can generate plausible answers from general knowledge.
That is not sufficient for a commercial client advisor.
The business should define which information sources the advisor can use.
For example, it could be:
the product database,
The PIM system,
approved product descriptions,
buying guides,
FAQs,
delivery information,
return policies,
warranty terms,
and other quality-assured subject content.
This makes it possible to build the advice around the business's documented knowledge , rather than relying on free generation.
Retrieval-Augmented Generation
A relevant technical approach is Retrieval-Augmented
Generation , often abbreviated RAG.
The principle is that the AI system first retrieves relevant information from a defined knowledge base and then uses this
the information as the basis for the answer.
For an online store, this could mean the customer asking:
"Is this jacket suitable for winter wear?"
The system finds relevant product information and any guides before formulating the answer.
RAG does not eliminate the risk of errors.
But the principle makes it possible to link the generation more closely to the business's own and controlled information sources.
Structured and unstructured knowledge must work together
A good digital advisor may need several types of information.
Structured data can tell:
price,
weight,
size,
stock status,
material,
compatibility.
Unstructured content can explain:
who the product is suitable for,
how it is used,
what compromises it entails,
and why a particular characteristic matters.
Consulting becomes particularly interesting when the system can combine the two.
The data provides precision.
The subject content provides context.
Explainable recommendations build understanding
A recommendation should preferably not only be presented as a result.
“We recommend model B.”
The customer can benefit more from:
"Model B fits your criteria because it is lighter than A and has a longer stated battery life. Model C has higher performance, but based on the usage you described, you don't seem to need the extra capacity."
The explanation enables the customer to evaluate the recommendation.
This strengthens the customer's own decision-making ability and can at the same time make the advice more transparent.
Uncertainty must be part of the design
Generative AI can produce errors.
Therefore, a digital customer advisor should be built for situations where it doesn't know .
It may involve:
to request more information,
to highlight uncertainty,
to avoid recommendations when the knowledge base is insufficient,
or to escalate the conversation to a human.
A system that always gives a confident answer can seem impressive.
But it can be a bad advisor.
Transmission to humans must be planned
Human handoff should not be considered a defeat for the AI system.
It's part of the service design.
The business should define situations where human assistance is necessary.
It could be when:
the customer requests it,
the question is outside the knowledge base,
the consequences of mistakes are great,
the situation requires discretion,
or the customer has a problem the system cannot solve.
A good digital service therefore does not have to choose between humans or AI .
It can design a good interaction between them.
Privacy must be built in from the start
A dialogue-based solution can get customers to share more information than a traditional search field.
Therefore, the business must think carefully about what information is actually necessary.
The customer should not be encouraged to share personal information that is not needed for the task.
The business must also have control over how information is processed, stored and possibly further used.
This is not something that should be added after the advisor is built.
Privacy must be included in the design itself.
The digital advisor can become a new source of insight
The conversations can also provide the business with valuable insights at an aggregate level.
They can show:
what questions customers ask,
what words they use,
which criteria they prioritize,
where product information is weak,
and where the advisor is unable to help.
This allows the system to be part of a continuous learning loop:
customer questions → insights → improved knowledge → better advice.
This is perhaps one of the most interesting characteristics of digital customer advisors.
They cannot just deliver knowledge.
They can help the business understand what knowledge needs to be improved .
Small and medium-sized businesses should start small
For a smaller business, it may not be practical to build a digital advisor for the entire product range at once.
A better start could be one category where:
customers have clear advisory needs,
the employees have good professional skills,
the product information is of high quality,
and the business can control the results.
The solution can then be tested against actual questions.
What did the system understand?
What did it misunderstand?
What information was missing?
When should it be forwarded to the customer?
This way, the business can gradually develop both the technology and the knowledge base.
From customer needs to recommendation
The digital advisor must bridge the gap between the customer's language and the product catalog
Step | What's going on? | Example |
1. Need | The customer describes the situation in their own words | "I need a lightweight laptop for studies" |
2. Clarification | The advisor asks relevant questions | Budget, usage, weight and battery requirements |
3. Criteria | The need is translated into product features | Low weight, long battery life, moderate performance |
4. Pickup | Relevant products and knowledge exist | Products within budget and requirements |
5. Comparison | The alternatives are assessed against the criteria | A is the lightest, B has the longest battery life |
6. Explanation | The customer learns why the products are suitable | Advantages and trade-offs are explained |
7. Decision | The customer retains control over the selection | The customer chooses based on their own priorities |
8. Escalation | Humans take over when necessary | Complex or unanswered questions |
What does the digital customer advisor need?
Good advice is built from the bottom up.
Building block | Why it is important |
Customer needs | The advice must be based on the task the customer is trying to solve. |
Product data | Provides precise facts about the products |
Product knowledge | Explains what the properties mean in practice |
Taxonomy and technical terms | Make knowledge more consistent |
Buying Guides and FAQs | Provides context and answers to common questions |
Rules and restrictions | Defines what the system can and cannot do |
AI model | Makes dialogue and linguistic understanding possible |
Human quality assurance | Checks professionalism and improves the system |
Human handoff | Ensuring human help when AI is not enough |
Feedback loop | Turn customer questions into a basis for improvement |
What can a smaller online store do first?
The digital customer advisor doesn't have to start as a giant AI project
Step | Question | Practical action |
1. Select area | Where do customers need the most help? | Start with one product category |
2. Collect questions | What are customers actually asking for? | Use customer service and employee experience |
3. Map the counseling | What does a good employee ask in return? | Document the conversation |
4. Clear product data | Do we have the information the advisor needs? | Improve attributes and data quality |
5. Document knowledge | What do employees know that the systems don't? | Create guides, explanations, and rules |
6. Define boundaries | When should AI not respond? | Create rules for uncertainty and escalation |
7. Test | Does the advisor provide correct and useful answers? | Test with real customer questions |
8. Learn | Where does it fail? | Improve the knowledge base |
9. Expand | Is the quality good enough? | Gradually add more products and needs |
Technical terms
Key concepts when building digital customer advisors
English technical term | Explanation |
Digital Shopping Assistant / AI Shopping Assistant | A digital solution that helps customers formulate needs, understand alternatives and find relevant products. |
Conversational Commerce | Commerce where dialog-based interfaces are used for, among other things, search, guidance, service or purchasing processes. |
Conversation Design | Planning how a digital service conducts a meaningful dialogue with the user. |
Natural Language | Human language as used in regular communication, as opposed to predefined commands or search structures. |
Intent | What the user is trying to achieve through a question or action. |
Retrieval-Augmented Generation | An approach where relevant information is retrieved from a defined knowledge base before a generative model formulates the answer. |
Knowledge Base | A structured or documented collection of information that the system can use as a basis for responses and advice. |
Structured Data | Information organized into defined fields or formats so that it can be easily processed by machine. |
Unstructured Data | Information that is mainly found in free text, documents, images or other forms without a fixed tabular structure. |
Explainable Recommendation | A recommendation where the customer is given information about why a product or alternative is suggested. |
Human Handoff | When a digital service transfers the customer to a human because the situation requires human assistance. |
Feedback Loop | A process in which experiences from use are used to improve the knowledge, service or system. |
AI Governance | Principles, responsibilities and control mechanisms for the responsible development and use of AI systems. |
Now we are starting to see the whole picture.
Magne
Friend...
Now we have come quite far.
The buddy
We have.
Magne
We started with a pretty simple idea.
The customer enters the online store and finds a product.
The buddy
And now?
Magne
Now I see a completely different online store.
The customer comes with a need.
The product information helps the customer understand.
Product knowledge explains the differences.
AI can help employees.
And the digital customer advisor can make parts of the knowledge available through a conversation.
The buddy
Do you see what has happened to the product catalog?
Magne
Yes.
It's still there.
We need products, prices, categories, inventory status and attributes.
The buddy
But?
Magne
It is no longer the entire online store.
We have built around the products:
explanations,
comparisons,
guides,
technical terms,
customer questions,
experiences,
counseling
and connections between knowledge.
The buddy
And what is it starting to look like?
Magne
A universe of knowledge.
The buddy
Exactly.
Magne
We're almost at the end of this section, buddy.
The buddy
Yes.
And now we're going to gather the threads.
Magne
From the old online store that primarily showed us what it had in stock ...
The buddy
...to an online store that can also help the customer understand what she needs, why it's suitable and how she can choose .
Magne
Then the next question is actually already given.
The buddy
That's it.
From product catalog to knowledge universe
Magne
Then we collect all of Part 7 there.
The buddy
And after that we can start to look up.
From the online store...
to the business.
From today's use of AI...
to what happens when AI becomes a natural part of how people work, search, learn and act.
Magne
So now we have not only just begun.
The buddy
No, buddy.
Now we are beginning to see where the whole journey has taken us.
Recommended literature
Developments in artificial intelligence are moving faster than perhaps any other field of study in our time. No single book can provide all the answers, but good books can provide a solid foundation for understanding the technology, the opportunities, and the challenges.
In the KI-Kompis series, we therefore recommend a selection of books that illuminate artificial intelligence from different perspectives – technology, strategy, management, innovation, ethics, digitalization and practical application. Together, they provide a broader understanding of how artificial intelligence affects people, businesses and society.
Click on the book icon to see the full literature overview with recommended books on artificial intelligence.
Recommended books from our library
Co-Intelligence: Living and Working with AI
Author: Ethan Mollick
Short review
One of the most recommended introductory books on artificial intelligence. Ethan Mollick explains in an easy-to-understand way how humans and AI can collaborate, what opportunities the technology offers, and why understanding artificial intelligence is becoming an important skill for everyone – not just technologists.
Why we recommend the book
This is one of the best books to start with if you want a practical and understandable introduction to artificial intelligence. It is a perfect fit as the first book in our blog series about AI.
The AI Advantage: How to Put the Artificial Intelligence Revolution to Work
Author: Thomas H. Davenport
Short review
A practical and insightful book that shows how artificial intelligence can be used to create value in businesses. Thomas H. Davenport combines research and concrete examples to explain how AI can improve decisions, streamline work processes, and contribute to innovation.
Why we recommend the book
This book is well suited for managers, employees, and decision-makers who want to understand how artificial intelligence can be used in practice. It bridges the gap between technology and value creation, and therefore fits well with the philosophy behind The Invisible Capital.
Artificial Intelligence: A Guide for Thinking Humans
Author: Melanie Mitchell
Short review
Melanie Mitchell provides a balanced and easy-to-understand introduction to what artificial intelligence is, how the technology works, and what limitations it still has. The book distinguishes between myth and reality and makes complex topics accessible to a wide audience.
Why we recommend the book
This book is perfect for those who want a deeper understanding of artificial intelligence without having to be a programmer or data scientist. It complements the other recommendations by placing today's AI developments in a larger academic and historical perspective - exactly the understanding we want to build through this blog series.
© 2025 - 2026 Portico Publish | The invisible capital Privacy and use of the website | Developed and operated by Magne Bjella | Powered and secured by Wix
Comments