"A Good Assistant Will Tell You No" - The Complaint That Made Us Rebuild Everything
Most companies open their origin story at the moment something worked.
Ours is easier to tell from the moment something failed, and the failure belonged to us.
Before HighSky AI existed, Kaloyan Hristov and his partners ran an online shop selling bags. Like every store at that scale, they had a discovery problem: too many products, not enough of anyone's attention to find the right one. So they did the reasonable thing and installed a chatbot.
Then a customer came in to order five or six safari bags.
The system recommended the wrong ones. At a higher price than it needed to be. The customer bought them, took them on the trip they were bought for, and came back angry. Not annoyed about a delivery date. Angry that his trip had been damaged by a recommendation he had trusted.
That complaint, as Hristov told Fakti.bg in June, is the reason the company exists.
The Sentence That Came Out of It
Six months of testing standard chatbots followed. None of them sold anything in a way that felt natural, and several of them actively pushed customers away.
What emerged from that period was not a feature list. It was a definition, and it is the sharpest thing anyone at this company has said in public:
"A good assistant will tell you 'no, this isn't right for you, but this is'."
Read that again as a commercial statement, because it sounds like the opposite of one.
Every instinct in retail technology pushes the other way. Recommendation engines optimise for what is likely to be bought. Upsell widgets optimise for basket value. A system built to maximise the current transaction will never volunteer that the thing in your cart is wrong for you.
But the safari bag customer did buy. The transaction succeeded. The relationship is what broke, along with the repeat purchase, the referral, and the trust that would have carried the next five orders.
An assistant that can say no is not leaving money on the table. It is the only kind that survives contact with a customer who finds out later.
Why This Is Harder Than It Sounds
Saying no requires knowing enough to be right about it, and that is where most implementations fall down.
To tell a customer that a product is wrong for them, the system needs to understand what they are actually trying to do, understand the product in enough depth to judge the fit, and be structured so it does not simply invent a confident answer when it lacks the information.
The third point is the one Hristov spent most time on. The architecture uses several specialised agents rather than one general model: one that knows the products, one that handles delivery, one for business questions. Splitting the work is what keeps hallucinations out, because no single component is being asked to be an authority on everything. We covered how that architecture works in more detail separately.
The second point comes with an uncomfortable dependency, which he stated plainly: the technology works best when product descriptions are detailed.
That is worth sitting with if you run a large catalogue. An AI agent does not compensate for thin product data. It exposes it. If three of your twenty thousand SKUs have a proper description and the rest have a manufacturer's title and a photo, the agent will be excellent on three products and vague on the rest, and the vagueness will be visible in a way it never was when customers were quietly failing to find things on their own.
The good news is that most large retailers have already done this work. As we heard at Balkan eCommerce Summit, they have invested years in exhaustive descriptions, quality photography and careful comparisons. That investment has been sitting unreachable behind navigation. This is the thing that reaches it.
Preventing a Purchase Is a Business Model
The Fakti.bg piece framed the customer benefit as avoiding unnecessary spending, and that framing deserves defending rather than softening, because a merchant reading it might reasonably ask why they would pay for something that talks customers out of buying.
Two answers.
The first is arithmetic. A wrong purchase does not stay bought. It comes back as a return, which costs the shipping twice, the handling, the restocking and the refund. In categories where returns run high, the difference between a right recommendation and a plausible one is the difference between a sale and a negative-margin round trip.
The second is what the safari bag customer demonstrated. The purchase that is never regretted is the one that brings someone back. A store that has told you honestly that something is not for you has earned a level of trust that no discount code buys.
None of this is charity. It is just a longer time horizon than a single checkout.
What the System Knows, and What It Does Not Keep
Hristov also addressed the question that follows any mention of personalisation. The agent adapts based on customer signals, while the information stays encrypted and is deleted from their systems.
This is consistent with how the company has described the model elsewhere: what makes a recommendation good is a picture of preferences, not a picture of a person. What you searched for, what you bought before, what you said you needed. Not your name, your age or where you live.
For a merchant this matters in a practical way rather than a philosophical one. Personalisation that runs on preferences travels well across the GDPR conversation, the security review and the procurement questionnaire. Personalisation that runs on accumulated identity does not.
Where He Thinks This Goes
Asked about the next five years, Hristov described a shift away from one shop looking the same to everyone. Either sites become heavily personalised per visitor, or they start to resemble a conversation interface, closer to what people are already used to from chat assistants.
Both directions have the same consequence for a large retailer. The grid of products and the filter sidebar stop being the primary way anyone reaches your catalogue. What replaces them is a conversation that has to know your inventory well enough to be useful in it.
The stores that will handle that transition comfortably are the ones whose product data is already good and whose systems can already answer a question in real time. That is not a five-year project. But it is not a weekend one either.
The Part Worth Taking Away
There is a version of this story where the safari bags are a funny anecdote about early days.
The more useful reading is that a bad recommendation is not a neutral event. It is not a missed opportunity that leaves things where they were. It actively costs you the customer, and the customer finds out at the worst possible moment, which in this case was somewhere on a trip with the wrong bags.
Most stores never hear about it. The customer does not write in. They just do not come back, and the number that records this does not exist on any dashboard.
If you want to see what an assistant that can say no looks like on your own catalogue, that is a twenty-minute conversation rather than a deck.
FAQ
Why would a merchant want an AI that talks customers out of a purchase? Because a wrong purchase does not stay bought. It returns, costing shipping twice plus handling and restocking, and it costs the relationship. A recommendation the customer does not regret is what produces the second and third order.
What does an AI agent need in order to give an accurate recommendation? Detailed product data, above everything else. The agent judges fit against what your catalogue actually says. Thin descriptions produce vague answers, because there is nothing specific to reason from.
How are hallucinations prevented in a shopping context? By splitting the work across specialised agents - one for products, one for delivery, one for business questions - so no single component is asked to be an authority on everything, and by pulling product answers from live catalogue data rather than model memory.
Does personalisation require collecting personal data? No. What improves a recommendation is a preference profile: prior searches, prior purchases, stated needs. Identity data adds legal and security exposure without making the recommendation better.
Will AI agents replace the product page? Not immediately, but the primary route into a catalogue is shifting. Expect either heavy per-visitor personalisation or conversational interfaces, and in both cases the grid and the filter sidebar stop being the main way people find things.
Kaloyan Hristov is CEO and co-founder of HighSky AI. The original interview ran in the business section of Fakti.bg on 4 June 2026. HighSky AI builds the SkyCommerce Suite for large online stores: AI Sales Agent, AI Search Engine, AI Voice Agent and Process Automation.