Why Most Companies Are Getting AI ROI Wrong

A marketing leader told me last month that her team had quietly become an AI shop. Every writer using a copilot. Campaign briefs drafted by a model. A chatbot on the site, an agent triaging the inbox, summaries of every call landing in Slack before the rep had left the room. By any adoption metric, they were ahead of the field.

Then her CFO asked the only question that matters. What did all of it return.

She did not have an answer. Not because the team was failing, but because everything they measured was about speed. Drafts produced faster. Tickets closed faster. Hours saved. None of it connected to a number the CFO recognized. She had a productivity story and a value vacuum, and she is very far from alone.

This is the real state of AI in mid-2026. Adoption is close to universal and value is close to invisible. G2 reports that 96% of marketers now use AI in their roles, but only 41% can demonstrate ROI, down from roughly half a year earlier. Zoom out to the whole enterprise and it gets starker. MIT's NANDA initiative found that after $30 to $40 billion in enterprise spend, roughly 95% of generative AI pilots are producing no measurable P&L impact. BCG puts only 5% of companies in the group capturing significant value, with 60% generating none at all.

The reflex when you see numbers like that is to blame the technology. Wrong model. Wrong tool. Wrong vendor. That reflex is exactly why the gap keeps widening.

The ROI Problem Is a Measurement Problem First

Most teams are measuring the wrong thing, so of course the number looks bad.

When you measure AI by time saved on tasks, you cap its value at the cost of the task. A writer who drafts a blog post in twenty minutes instead of two hours has saved you an hour and forty minutes of writing. That is real, but it is small, and it is the entire ceiling of a task-level metric. You will never find transformative ROI by counting minutes, because minutes were never where the money was.

The companies seeing actual returns are measuring something else. They ask whether AI changed a decision, not whether it accelerated a task. Did the campaign get pointed at a better segment. Did the deal get prioritized correctly. Did the team stop doing three things that were not working. Those are outcomes, and outcomes are where the EBIT line actually moves. McKinsey's research is blunt about it: most AI applications today are tools that accelerate existing work and largely preserve the underlying workflow. The payoff only shows up when the work itself is redesigned.

So the first ROI error is measuring acceleration when you should be measuring outcomes. But there is a deeper one underneath it.

The Deeper Error Is Structural

You cannot redesign work around AI if the AI cannot see the work.

This is the part the value-gap reports keep circling without quite naming. MIT's researchers concluded that the core barrier was not infrastructure, regulation, or talent. It was learning. Most deployed systems, they wrote, do not retain feedback, adapt to context, or improve over time. They produce a fast answer, forget the moment it is given, and start cold the next day. A system with no memory of your business cannot make a better decision about your business. It can only make a faster generic one.

I made this argument about output in Context Is the Whole Game, and it applies just as cleanly to returns. The model is commoditizing. What determines whether AI produces value is the context wrapped around it: the customer data, the product detail, the history of what worked, the rationale behind past decisions, all reachable and current. A team with thin context gets thin returns no matter how good the model is, because the model has nothing specific to reason about.

This is why the ROI gap is really an operating-model gap. In The AI-Native Company I argued that the companies pulling ahead are not the ones with the biggest model bills, but the ones whose structure lets intelligence flow upstream of decisions. The same divide explains the ROI numbers exactly. The 5% capturing value did the unglamorous work of making their context usable. The 60% bolted AI onto a structure built for a different era and are now wondering why the pilots stalled.

What to Measure Instead This Quarter

If your AI ROI looks disappointing, change the question before you change the tool.

Stop reporting hours saved. Start reporting decisions improved. Pick three decisions your team makes regularly, where you can actually point at a campaign, a segment, a budget allocation, and ask whether AI made that decision sharper. That is a number a CFO will recognize, and it is the only kind that compounds. As I argued back in What Is Operational Intelligence, speed without context produces noise, not signal. Speed metrics measure the noise. Decision metrics measure the signal.

Then look honestly at your context layer, because that is the real lever. If your AI cannot reach your customer history, your past campaign results, and your product detail in a usable form, no model upgrade will fix the returns. The work that closes the ROI gap is not buying a better tool. It is making your own context legible to the tools you already have.

The Takeaway

The ROI gap is not evidence that AI does not work. It is evidence that most companies are measuring the wrong thing and feeding their AI too little context to do anything worth measuring.

Time saved is a vanity metric. Decisions improved is the real one, and you only get there when your AI runs on a context layer rich enough to reason about your actual business. The companies that figure this out in 2026 will not be the ones that spent the most. They will be the ones that stopped counting minutes and started counting outcomes, and built the context to make those outcomes possible.

AI does not have an ROI problem. Most companies have a measurement problem and a context problem wearing an ROI problem as a costume.

Tracy Thayne* is the founder of Expona, an AI-powered operational intelligence platform for B2B marketing. Read the Expona founder story or subscribe to the blog (below) for weekly insights on context, AI, and the operating model of the next decade.*