How Businesses Actually Decide to Add AI to Their Products (Most Get It Wrong)

Here's a number worth sitting with: 88% of companies now use AI in at least one business function. And yet, only 5.5% of those organizations achieve meaningful financial returns from their AI investments.

That gap is not an accident. It's the predictable result of a broken decision-making process that most companies run when they decide to add AI to a product.

I've watched this pattern play out across industries. In my experience managing enterprise client relationships in B2B SaaS, the conversation about AI was constant but the clarity about what AI should actually do was rare. Clients wanted AI. Vendors were selling AI. And somewhere in that space, the fundamental question, "what problem does this solve for the user?", got lost.

This is a piece about that gap. Not about the technology. About the thinking.

The Statistic That Explains Everything

An S&P Global survey from 2025 found that 42% of companies abandoned most of their AI initiatives during the year. That's up from just 17% in 2024. The average organization scrapped 46% of proofs-of-concept before they ever reached production.

Only 26% of executives believe their AI programs delivered the ROI they originally projected.

Meanwhile, 73% of AI initiatives never move beyond the pilot stage, and 53% of investors expect returns within 6 months of deployment.

The problem isn't that AI doesn't work. The problem is that companies are deploying it before they have answered three basic questions: What decision are we actually trying to improve? Who does the AI serve, and how does their workflow change? And how will we know if it worked?

Without those answers, even a technically successful AI implementation fails as a business outcome.

The Three Decisions Every Company Faces

When a business decides to add AI to a product, they're actually making one of three bets. Understanding which bet they're making changes everything downstream.

Build: You own the intelligence.

This means training or fine-tuning models, owning the data pipeline, building the infrastructure. The advantage is deep differentiation. If AI is the core of your competitive moat, and you have proprietary data that makes your model better than anything you could buy, building is the right choice.

The cost is time. Building AI from scratch takes 12 to 24 months, and your competitors are shipping things today. Build also requires internal expertise that most organizations don't have and can't hire fast enough.

Buy: You rely on external intelligence.

Off-the-shelf AI tools, vendor platforms, enterprise AI software. The advantage is speed. You can have something in front of users in weeks. The disadvantage is that you have no control over the model, the pricing, or what happens when the vendor changes their terms.

Buy works when the problem is common enough that a vendor has already solved it well, and when speed to market matters more than differentiation.

Blend (or Wrap): You direct the intelligence.

This is the approach 32% of companies are taking, and it's increasingly where the interesting product decisions are happening. You take a third-party model (GPT-4, Claude, Gemini) and build your own layer around it: retrieval systems, validation logic, business rules, workflow integrations, proprietary data access.

You're not building the model. You're building what makes the model useful for your specific user in your specific context.

The companies that win with the wrapper approach treat it as a wedge, not the finished product. They ship fast, prove the market exists, then build the data moat and workflow integrations that make the product defensible. The wrapper is not the strategy. It's the entry point.

Research suggests the break-even between building from scratch and buying externally typically lands around 33 months. For most early-stage product decisions, that math alone pushes companies toward blend or buy.

The Activity Trap

The biggest structural failure in enterprise AI isn't a technology problem. It's a governance problem called the activity trap.

Here's how it plays out: A company deploys an AI tool. They measure adoption, usage rates, and deployment numbers. 78% of employees have used it. The pilot is declared a success. Budget is renewed.

And then nobody can answer: what changed in the business because of it?

A financial firm in one widely cited case deployed AI to 40,000 employees with 78% adoption and still couldn't quantify time saved, cost reduced, or errors avoided. There was no baseline. Success had never been defined in business terms.

CIO research from 2026 puts it plainly: "The organizations that succeed will not necessarily be the ones that deploy the most tools. They will be the ones that measure outcomes early, govern with discipline, and separate real value from visible motion."

Three failure patterns show up consistently:

Pilot purgatory. The project succeeds technically but lacks a clear business value threshold for graduation. The question asked is "Can the tool do this?" instead of "Does this create measurable business value?" The result is zombie projects that consume budget indefinitely while producing nothing a CFO can point to.

The adoption illusion. High usage numbers get reported to the board. But adoption of a tool that doesn't improve a workflow is just measuring a distraction. If the AI helps someone do something faster that shouldn't be done at all, that's not an ROI win.

No stopping criteria. Most AI pilots never define what failure looks like upfront. Without explicit criteria for discontinuing a project, underperforming tools survive on inertia. McKinsey finds high-performing AI companies are 3x more likely to maintain discipline around stopping underperforming initiatives.

What Customers Actually Want (and What Companies Keep Missing)

There is a persistent expectation gap in AI product decisions, and it runs in the opposite direction most companies assume.

70% of executives believe customer expectations are evolving faster than their company can adapt. But the nature of those expectations is frequently misunderstood.

Customers want speed, relevance, and transparency. 61% expect more personalized service from AI-powered products. 90% expect immediate responses. And yet, 58% of consumers are only somewhat or not at all comfortable engaging with AI on behalf of brands.

The discomfort isn't with AI itself. It's with AI that is obviously AI without being obviously useful.

The product insight from this: customers adopt technology when it feels simple, natural, and useful. They don't want to know they're using AI. They want the outcome AI enables. A great AI feature is one users don't describe as an AI feature. They describe it as "it just works."

Companies that lead with the AI are almost always doing it for the press release, not the user. Companies that lead with the outcome, and let the AI be the mechanism, are building things people actually use.

I've seen this firsthand working with enterprise clients in financial software. They didn't want AI dashboards. They wanted to spend less time reconciling transactions manually. The AI was never the sell. The saved hours were. Any product that makes the technology the headline has already missed the point.

Five Questions to Ask Before Adding AI to Anything

Based on what the research shows and what I've seen in practice, these are the five questions that should frame every AI product decision:

1. What specific decision or workflow gets better, and by how much? If you can't name a measurable improvement, you're building a feature, not a solution.

2. What does the user have to do differently, and is that change smaller than the benefit? AI features that require users to learn a new workflow often get abandoned, even if the underlying capability is genuinely useful. The switching cost has to be worth it.

3. What data do you have that makes your AI more useful than a generic model? If the answer is "none," you're building a wrapper with no defensibility. That's not necessarily wrong at launch, but it has to be a means to an end, not the end itself.

4. How will you know in 90 days whether this worked? Define the baseline now. Pick one primary metric that reflects actual user or business outcome. Track it before you deploy. If you can't answer this question, you're not ready to build.

5. What gets worse if the AI makes a mistake? This question is almost never asked, and it's the one that matters most in regulated or high-stakes environments. AI error tolerance varies enormously by context. A hallucinated product recommendation is a nuisance. A hallucinated compliance output is a liability.

What Separates the Companies That Actually Win

McKinsey's 2025 State of AI report is clear on this: the dividing line between AI high-performers and everyone else is not budget or technical sophistication. It's how they treat AI in relation to the rest of the business.

High performers are 3x more likely to redesign workflows end-to-end rather than layering AI onto existing ones. They set outcome-based objectives tied to real business KPIs, not tool adoption metrics. And they have strong senior leadership engagement, not just a sponsored pilot program.

The companies losing money on AI are the ones that bought a tool, ran a pilot, declared success based on activation, and moved on. The companies winning are the ones that redesigned the work the tool was meant to do, measured the before and after, and kept iterating.

That's not an AI insight. That's a product management insight. The question is never "should we add AI?" The question is "what problem are we actually solving, who has it, and how will we know when it's solved?"

AI is unusually good at making busy work feel like progress. The companies that will win the next five years are the ones that don't confuse the two.