4 Lessons from failed AI adoption ideas

Failure has a way of teaching what success never quite manages to. After 30+ agentic AI projects across industries, from healthcare to automotive, we’ve picked up a fair share of stories where the path to value wasn’t a straight line. Some of these stumbles were ours. Others belonged to clients who walked through our door after trying to forge ahead on their own, with mixed results.

Either way, the lessons stuck.

Each of the stories we share today is one that we lived through. And since we’re elaborating on ineffective patterns of thinking, we refrain from using any names or brands.

Lesson 1: The sky is the limit for AI

The utopian promise of all-encompassing AI is unhinged. The companies that made a big bet on the overarching capabilities of AI burned their fingers (Apple Intelligence, anyone?). What we’re betting on at Vstorm, with a lot of success, is more modest.

Vstorm is in the business of ‘defining parameters for AI to work’, not in the business of ‘solving any problem with potential intelligence explosion’.

Remember Maverick’s (played by Tom Cruise) saying when confronted with a vision of the future in which drones rule the sky instead of human pilots. He said, “Maybe, but not today.”

The few years in the LLM adoption journey taught us that each of our 30+ successful projects had one thing in common: we agreed with our customer exactly what the isolated playground for AI to excel is, so we could make the model play the rules of the game well, and achieve a success rate above human actors. We repeated the same across industries, from healthcare to automotive.

Maybe someday, the sky will be the limit. For now, we advise limiting your ambition to a local playground and winning there consistently.

Lesson 2: Skewed perspective of ‘AI taking on the human roles’

“I need AI to replace my team” — it is something we too often hear from top management. And it is the first thing that autocorrects itself during a successful project. Your management might feel like it’s a good idea to put AI into human roles, but that’s a temporary thing. This changes the moment we begin to ground the discussion in pragmatics and current state-of-the-art technology.

One key revelation here is that LLMs are great pattern spotters that can leverage the set of patterns your business operates on. These patterns consist of records of data: how fulfilment happened, how a purchase order was registered, how the customer was answered. But even the largest store of such records does not account for what was never verbalised, and that is experience. LLMs cannot draw from that resource and are limited to recorded patterns. This is already a significant space, but if plotted as a bell curve, these patterns occupy the middle.

The place where the experience of your team counts is on the edges. It is the unexpected, non-trivial, unforeseen situations where LLMs are prone to fail, but your trusted team is not. And for the foreseeable future, management needs both: AI to handle the commonalities, and people to look after it and solve for the extremities.

Lesson 3: Ambitious expeditions do require a guide

Building production-grade AI can be imagined as an expedition. One climbs a mountain to reach the business value at the top. In this analogy, Vstorm is a guide. However, it is surprising how many projects we open are actually failed expeditions, ones where the customer has tried alone, or with some underqualified IT vendor, to forge a path forward that simply did not work.

At Vstorm, we like those projects over others, because they create circumstances in which our customer actually expects and hopes for us to lay out the path forward.

Such a path is built in the early stage of the project, and we like building it by kicking off with an on-site workshop that joins both business and technical perspectives. Vstorm can be your sherpa if you take your preparation seriously.

Lesson 4: Slow down with your success, turbo

We have seen situations where our customers, after getting a reliable proof-of-concept working at 80–90% accuracy, felt that the solution was good enough for productive use and simply ran with it. While we were happy with their enthusiasm, such an approach usually backfires.

Any AI agent is a stochastic element in a business process, very unlike rule-oriented, deterministic systems. And because the workings of AI are probability-based, they require a different approach to testing. The same result passing a test twice does not mean it would not change on the third run. So an agent working ‘good enough’ is not actually good enough just yet.

Even with a success, a working proof-of-concept, around the corner, we take a well-crafted path to ensure that the AI can be deployed with good oversight, protective measures, and full observability. Only then can we expect to gain the benefits of using AI instead of exposing ourselves to unforeseen problems.