If you are adding AI to your app because the board asked for AI, stop. AI earns its place in a product when it does a specific job better than the alternative, and users can feel the difference when it does not. Most AI features fail that test, and the numbers on this are now hard to argue with.
A 2025 MIT study of 300 public AI deployments found that 95 percent of enterprise generative-AI pilots delivered no measurable impact on profit and loss (MIT, 2025). Gartner expected at least 30 percent of generative-AI projects to be abandoned after proof of concept by the end of 2025 (Gartner, 2024). A RAND study put the failure rate of AI projects at around 80 percent, roughly double the rate for ordinary IT projects (RAND, 2024). Adding AI is easy. Adding AI that is worth having is not.
Should you add AI to your app?
Only if it does a specific, valuable job that a simpler approach cannot. The test is embarrassingly basic: name the task, name what it replaces, and say why the AI version is better for the user. If you cannot do that in a sentence, you do not have an AI feature, you have a line for a pitch deck. The failure rates above are not really about the technology. RAND found the main causes of AI project failure were organisational and strategic, not technical: unclear problems, the wrong data, and no business case (RAND, 2024).
Why do most AI features fail?
Because they are added to the product rather than designed into it. The pattern is a decision to "add AI" followed by a search for somewhere to put it, which is backwards. When McKinsey looked at where generative AI actually creates value, it found roughly 75 percent of the potential concentrated in just four areas: customer operations, marketing and sales, software engineering, and research and development (McKinsey, 2023). Value is narrow and specific. Sprinkling a chatbot across a product that needed none of it lands nowhere near those pockets.
The market is starting to punish it, too. In 2024 the US Securities and Exchange Commission fined two firms 225,000 and 175,000 dollars for "AI washing", marketing AI capabilities they did not actually have (SEC, 2024). And even among companies genuinely using generative AI, only around 6 percent attribute more than 10 percent of their profit to it (McKinsey, 2024). Broad adoption, narrow payoff.
When does adding AI to your app earn its place?
When the job is high-volume, repetitive and pattern-heavy, and getting it slightly wrong is survivable. McKinsey’s own examples are the giveaway: generative AI can cut human-handled customer contacts by up to half in sectors like banking and telecoms, and lift marketing and sales productivity by a few percent of spend (McKinsey, 2023). Those are real jobs with real volume behind them.
AI is worth building in when the job looks like this:
- The task is high-volume and repetitive.
- It is pattern-heavy, the kind of work models are genuinely good at.
- A small error rate is survivable rather than dangerous.
- You can name what it replaces and why the AI version is better.
If your product has a task like that, where AI removes genuine, repeated effort, it can be excellent. That is a different thing from bolting a generative feature onto a workflow that was working fine.
What does bolt-on AI cost you?
More than the licence fee, in three ways. There is the direct cost: early generative-AI deployments ran between 5 and 20 million dollars at enterprise scale (Gartner, 2024), and running inference in production is not trivial even at smaller scale. There is the maintenance cost, because an AI feature is not a fixed thing you ship once, it needs monitoring, evaluation and correction as models and data drift. And there is the trust cost, which is the one people miss. Only 26 percent of consumers trust organisations to use AI responsibly (Qualtrics, 2025), and 71 percent say they will abandon an AI experience that is irrelevant to them (reported, 2025). A bad AI feature does not sit there harmlessly. It actively costs you users.
How do you tell the difference?
You start from the job, not the technology. Work out what your users are actually trying to do, find the parts that are repetitive and high-volume, and ask whether AI does those parts materially better. If it does, build it properly and measure it against a real target. If it does not, leave it out and spend the money on something your users will notice. "Add AI" is not a strategy. Knowing which specific problem AI solves in your product, and being honest when the answer is none, is the whole job.