Why traditional applications fail to get customers to use them (and what changes with artificial intelligence)

There is a question that comes up in almost every first meeting with a company that has already tried to digitise its relationship with customers: “we tried it, we built an application so that they could order directly, and nobody used it. Why would it be different this time?” It is the best question anyone can ask, and it deserves an answer that is not “because now it has artificial intelligence”.

This article is that answer. It is written by a company that builds applications with AI, so read it knowing that; that is also why it ends with the cases where artificial intelligence fixes nothing.

Why the previous application was not used

When an application fails, it almost never fails because of a technical fault. It fails because it asks the customer to adapt to it, and the customer has ten other suppliers asking exactly the same.

Think of what a conventional ordering application demands: install it or remember a web address, remember a password, learn a navigation, find among thousands of references the product they call “the five-litre one”, respect a quantity format, and do all of it on a screen while standing in a kitchen, on a building site or behind a counter. Each step is small. The sum is the reason why, on the second order, they pick up the phone again.

The underlying figure comes from Spain’s statistics office, INE: in 2024 only 26.6% of Spanish companies with ten or more employees sold through e-commerce, and in that survey e-commerce means orders received via a website, an app or electronic data interchange — not an order typed by hand into an email. Put the other way round: roughly three out of four companies received no sales through a channel designed to collect orders. There, orders arrive by phone, by message or through the sales rep, and somebody types them in. It is not that companies have not built channels; it is that the customer chooses, every time, the one that asks the least effort.

And the channel that asks the least effort is already installed. According to the CNMC Household Panel, WhatsApp is the preferred messaging app for 93.9% of Spanish internet users. The panel measures people, not companies, and that is worth saying; but the person placing the order in a bar, on a building site or in a shop is one of those people.

Internal adoption fails for the same reason. The employee who knows the process has a shortcut — a sheet of paper, a spreadsheet, a phone call — and the new application takes it away without giving anything back. By the third day, the shortcut is back.

What a person does well, and why customers prefer one

Look at what happens when that same customer calls their usual sales rep. They say “give me the usual, plus two more boxes of last week’s thing”. The rep understands what “the usual” is, knows what went out last week, makes a suggestion — “shall I add the other one too, you must be running low” —, resolves the exception without anyone filling in a form, and the customer has learned nothing new. There is zero friction because the intelligence is on the other side: the rep adapts to the customer, not the other way round.

The problem is that this person does not scale. They cannot know thousands of references with their formats and equivalences, they are not there at seven in the evening or in August, they cannot serve every customer at once, and every order they take by phone is an order somebody later types into the management system. For decades, the only way to grow without more sales reps was to force the customer to use an application. And the customer, rightly, said no.

What changes with an application that integrates artificial intelligence

Four things, and they should be concrete, because “it has AI” explains nothing.

1. The customer talks the way they talk. They write a message, send a voice note, take a photo of the order sheet they fill in by hand, or send an email; from the application or directly on WhatsApp. The system interprets what they asked for using that customer’s own vocabulary — “the five-litre one”, “the big sack”, “the usual” — and turns it into real references from your catalogue. There is no navigation to learn and no format to respect. They interact the way they would with a person.

2. The application has all the context. This is the biggest difference and the least explained. A well-built assistant is the world’s greatest expert on your company: it knows the entire catalogue with its formats, the equivalences each customer uses, their order history, their negotiated price list, their terms, your service rules and your exceptions. No person can hold all of that in their head; your best sales rep knows their own customers well, not all of them. And that context is not invented: price, stock and terms are looked up in your management system at the moment of answering — the model neither remembers nor calculates them. How it is built so that it cannot invent them has its own page (in Spanish).

3. It proposes instead of waiting. It opens the conversation with the last order at today’s prices, flags what that customer used to order regularly and has not ordered for a while, suggests what similar customers buy, offers a substitute when something is out of stock. It is exactly what the good sales rep did, but for every customer at once and with all the data in front of it. What it can propose and where it gets it from, without selling for the sake of selling, is explained separately (in Spanish).

4. It is always available and never gets tired. At seven in the evening, in August, with twenty customers writing at the same time. No queue and nobody typing afterwards: the confirmed order enters the management system through the same circuit as if the rep had entered it.

And the same mechanism works internally. The employee who filled in a report, typed a delivery note or looked for a figure across three screens stops doing so: they dictate, photograph or ask, and the application adapts to the way they work. Internal adoption stops being a battle because the application no longer takes away the shortcut; it is the shortcut.

What artificial intelligence does not fix

This is where an honest supplier has to slow down, because the promise of “this time it will work” breaks at exactly these points.

• If the catalogue and the price lists are not in the system, there is no context to look up. If the price for part of your customers lives in someone’s memory or in a spreadsheet, that is the first job, it is yours, and AI does not do it.
• If the process has no written criteria, the assistant has nothing to apply. What to do with an order above a certain amount, with a customer with an outstanding balance or with a discontinued reference is decided by your company, in writing, and the system applies it.
• An assistant built to “sell on its own” breaks trust. A system optimised to push volume produces forced recommendations, the customer notices by the second conversation and stops trusting everything else. The EU AI Act, moreover, expressly prohibits manipulative techniques (Article 5). What drives adoption is relevance, not insistence.
• You have to say it is an artificial intelligence. Article 50 of the same Regulation requires informing people that they are interacting with an AI system, and it has applied since 2 August 2026. It is not an obstacle to adoption: what the customer values is being understood first time, not having a person on the other side. What is an obstacle is feeling deceived.
• Exceptions still go through a person. No system understands one hundred per cent of what reaches it. Doubtful items are asked about in the same conversation or passed to someone on your team; what changes is who types, not whether anyone reviews.
• Adoption is not guaranteed; it is measured. That is why the first phase of a project like this is a proof of concept on real cases from your company, not a rollout to your whole customer base: you check what percentage of orders would go through without intervention and how many customers come back to use it, and those numbers decide whether to continue. If the answer is that it does not pay off, that has to be said too.

How to tell whether your case is one of these

Five signs which, together, indicate that an application with artificial intelligence can achieve the adoption the previous one did not:

1. Your customers already write to you on WhatsApp or send you voice notes, and someone on your team types them in.
2. You have a wide catalogue in which each customer uses their own vocabulary for the same products.
3. Prices and terms vary by customer and live in your management system.
4. You already tried an application or a portal and usage faded within a few months.
5. The bottleneck is the people who type, not the people who decide.

If three or more apply, a conversation is worth having. If your management system is one of the usual ones on the market and your terms are standard, a specialised product may serve you sooner and cheaper; how to decide between a product and custom development, with our bias declared, is on another page (in Spanish). And what an application with AI costs, with what is included and what is not, too. Our English overview is at custom applications with AI.

And does the previous application get thrown away?

No. The portal or application you already have stays for the customer who does use it, and the assistant provides an entry point through the channel the rest use. Both end up in the same place, which is your management system. What gets retired is not the previous application: it is the obligation for the customer to adapt to it.

Sources

• INE, “Survey on the use of ICT and e-commerce in companies. Year 2024 – First quarter 2025”, final data, 22 October 2025: 26.6% of companies with 10 or more employees sold through e-commerce in 2024. ine.es
• CNMC, Household Panel, press release “Internet uses and OTT services”, 31 October 2025 (Q2 2025 data; 5,176 households and 8,709 individuals): WhatsApp, preferred messaging app for 93.9% of internet users. cnmc.es
• Regulation (EU) 2024/1689 (EU AI Act), Article 5 (prohibited practices) and Article 50 (transparency), applicable since 2 August 2026. eur-lex.europa.eu