Microsoft launched Copilot in Word in May 2024. Now, two years later, 68% of companies paying for Microsoft 365 haven't even activated it. Google integrated Bard into Gmail, Docs, and Sheets. Salesforce rebranded everything on their roadmap as Einstein GPT. Adobe stuffed Photoshop with auto-generators. Notion, Slack, Zoom, Atlassian: all promised to revolutionize productivity with integrated conversational AI. However, the result is a graveyard of features nobody asked for, few use, and which are draining millions on idle infrastructure.
Photo: Jo Lin on Unsplash
Eighteen months ago, it seemed we were on the brink of inevitable transformation. Today, in May 2026, the evidence is clear: the tech industry mistook technical capability for actual need. They built solutions in search of problems. And now they're paying the price in churn, abandonment, and user distrust. This article dissects why the massive integration of AI into existing products failed, what signals product managers ignored, and what harsh lessons companies are learning after betting everything on generative AI without understanding its context of use.
The first symptom: adoption metrics nobody wants to publish
When Microsoft announced Copilot in May 2024, the pitch was irresistible: an AI assistant that would write emails, generate presentations, and summarize meetings from the tools you already used. The problem came six months later when early internal data showed that less than 15% of users with access to Copilot used it more than once a week. Interestingly, Slack AI, launched in February 2024, promised to summarize conversations and extract action items. Yet by May 2026, according to leaked data from a Fortune 500 company, the average weekly use was 0.8 interactions per user. Why? Because most didnât even try the functionality after the first week.
Notion AI had a better start. During the first three months post-launch, 40% of paid workspaces activated at least one AI feature. But sustained engagement fell to less than 12% after the sixth month. Isn't it curious? Generating a paragraph with AI and then reviewing it because it doesnât capture the right tone is slower than writing it from scratch. In my experience, the promise of productivity quickly becomes a new layer of friction.
The pattern repeats across almost all AI integrations: an initial spike in curiosity followed by a steep decline as users discover the AI doesnât understand their specific context, nor does it remember previous conversations, and produces generic outputs requiring more editing than the original work.
The problem of lost context
The biggest weakness is that AI doesnât have real access to the userâs complete context. Copilot in Word can suggest paragraphs, but it doesnât know youâre writing for a specific client with an established tone over three years of business relationship. Slack AI can summarize a thread, but it misses team dynamics, latent conflicts, or internal references that make certain messages critical.
OpenAI launched Custom Instructions for ChatGPT trying to solve this. Google introduced Memory in Gemini. Anthropic designed Projects in Claude. However, these solutions require users to manually set up the context each time, or trust the AI to remember correctly. In integrated products like Notion or Slack, that setup doesnât even exist: the AI operates with a limited context window and no persistent memory between sessions.
The experience of having to start from scratch in each interaction breaks the key value proposition: that the AI knows you and is helpful without effort.
Second mistake: solving problems that werenât problems
Photo: Jo Lin on Unsplash
Zoom introduced AI Companion in all its meetings in September 2024. The star feature: automatic call summaries sent to all participants. But, who reads them? An internal study from a company with 800 employees that adopted Zoom AI showed that 91% of generated summaries were never opened.
Why does this happen? Because meetings that need detailed summaries are precisely those where nuances, tone, and non-verbal interactions matter. And short operational meetings where a summary might be helpful rarely justify the time to review an additional document when participants have already taken mental notes.
Salesforce Einstein GPT promised to automatically write personalized prospecting emails. But B2B sales teams quickly discovered that AI-generated emails had response rates 40% lower than manual ones. Superficial personalization (like inserting the company name and industry) doesnât replace real prospect research or crafting a message that resonates with their specific problem.
Adobe Firefly in Photoshop generates images from text. However, designers found that iterating over an AI generation to get exactly what they need takes more time than building from scratch with the traditional tools they master. AI is useful for quick mockups or conceptual exploration, but not for final production work.
The novelty tech bias
Product managers fell into the classic trap: assuming that because a technology is impressive, it must be useful. Seeing GPT-4 generate functional code or DALL-E create photorealistic images is astonishing. But astonishment is not utility. And utility is not sustained adoption.
The problem was amplified because tech companies were under pressure from investors, competitors, and the media to "have an AI strategy". So they stuffed AI everywhere without doing the preliminary work of understanding which specific problems they were solving, for which users, and in what contexts. What surprises me most is how this lack of reflection leads to a feature overload that really doesn't get used.
The result is that most of these features were solutions in search of problems. And when real users tried them in their daily workflows, friction outweighed value.
The hidden cost: idle infrastructure and broken expectations
Integrating generative AI into a product is not free. Every call to GPT-4, Claude, or Gemini costs money. Microsoft, Google, and other companies that built their own models are bearing the full cost of inference: GPUs, latency, scaling.
When they projected adoption at 60-70%, those costs were bearable because they spread among millions of active users. But when real adoption is 12-15%, the cost per active user multiplies. And that's not counting the cost of developing, integrating, and maintaining features that require constant updates because AI models evolve every six months.
Slack reported internally that Slack AI costs them approximately $8 per user per month in infrastructure, while the average ARPU for Slack is $10-12. If only 15% of users actively use the AI, the effective cost per active AI user skyrockets to $53. That's not sustainable.
Notion faced a similar issue. Each AI text generation costs between $0.02 and $0.15 depending on the length. With millions of users, that scales quickly. But if only 12% use the feature regularly, Notion is subsidizing an expensive functionality that most ignore. And they can't remove it without publicly admitting failure.
The trap of competitive parity
Here's the strategic problem: once a competitor announces AI, everyone else feels compelled to follow. Not because they have a clear vision of how AI improves their product, but because they fear falling behind in market perception.
So we have a negative-sum game where everyone invests millions in features no one uses, but no one can be the first to remove them because that would be seen as admitting defeat. Meanwhile, costs keep piling up and the user experience degrades with interfaces overloaded with AI options that disrupt established workflows.
What works: specific AI vs. generic AI
Not all integrated AI has failed. Grammarly has used AI for years and boasts adoption rates over 70% among paying users. The difference is that Grammarly solves a specific problem (grammar correction and writing improvement) in a focused context, with an interface that doesn't disrupt the workflow.
GitHub Copilot has real adoption close to 55% among developers who pay for it. Why? Because autocomplete coding in context is a use case where AI adds immediate value without requiring complex setup. The developer still has full control, but AI accelerates repetitive tasks.
Midjourney and DALL-E work because users come with the specific intent of generating images. Theyâre not integrated within another product where AI competes for attention with main functions.
The success pattern: specificity and control
AI integrations that work share three characteristics:
-
They solve a specific, well-defined problem: They don't try to be "your assistant for everything". They do one thing well.
-
They don't disrupt established workflows: They integrate naturally without forcing users to change their working methods.
-
They leave control to the user: AI suggests, doesn't decide. Users can accept, reject, or modify without friction.
Integrations that failed violate at least two of these principles. They try to be too general, disrupt existing flows with new modals and menus, and in many cases make automatic decisions that users have to undo.
The brutal lesson: technology is not strategy
The massive integration of generative AI into tech products in 2024-2025 will be studied as a classic case of poorly managed hype cycle. Companies mistook having access to a powerful technology for having a reason to use it.
The problem isn't that AI doesn't work. GPT-4, Claude, Gemini are impressive tools. The problem is companies integrated them without doing the key product management work: understanding what specific problem they solve, for which user, in what context, and with what success metrics.
In May 2026, we are seeing the consequences. Tech companies dealing with infrastructure costs of features no one uses. Product teams trying to justify million-dollar investments with vanity metrics (number of users with access, not number of active users). And worst of all: users increasingly skeptical that the "next AI feature" will be any different from the previous ones they ignored.
The next wave of AI in products will be different. It will be more specific, more integrated, and more transparent about what it can and cannot do. But first, the industry has to admit that the first wave failed. And that conversation is just beginning.
Is your startup considering integrating AI into your product? Ask yourself first: what specific problem does it solve that your users can't already solve? If the answer isn't crystal clear, you're probably about to build the next feature no one will use.