Every company seems to have approved the AI budget by now. Far fewer can point to what actually changed because of it.
McKinsey's research puts the number in stark terms: more than three quarters of companies are using generative AI, and more than 80 percent of those same companies report no material impact on their bottom line. Goldman Sachs' chief economist went further earlier this year, describing the return on the current wave of AI investment as basically zero.
That gap is usually blamed on the model, but it's rarely the model that's actually missing. What shows up instead is in how people behave around the AI they've been given. Some teams hand a task over completely, trusting an agent to close it out with no real way to check what happened along the way. Other teams go the opposite direction, keeping AI boxed into narrow, pre-approved steps with a person signing off at every stage, which is safer but slow enough that most of the time savings disappear. Both are really the same problem wearing different clothes: nobody built a way for a person to see what the AI actually did, so trust ends up calibrated by guesswork instead of evidence.
Chatbots made this worse in a quiet way. They were a genuinely useful first step, but they answer every question the same way regardless of who's asking or what's at stake, and they offer no window into the reasoning behind the answer. That's tolerable for a low-stakes question. It becomes a different situation entirely once an agent can close a security incident, approve a request, or touch a record that other systems depend on.
A few things separate the organizations where AI is actually earning trust from the ones still hoping it will.
The first is whether actions are visible and reversible. If an agent does something on someone's behalf, whoever's accountable needs to see what happened and be able to undo it before it becomes permanent. Without that, approval is theater. People sign off on outputs they can't actually interrogate.
The second is whether the system's reliability is legible, not just real. Even in carefully controlled academic studies, peer reviewed research has found leading language models fabricate between 28 and 39 percent of their outputs, with clean source material and no adversarial conditions involved. In a live setting, one wrong answer rarely stays contained. It becomes the input for whatever comes next, and errors compound quietly across a chain of agents while confidence in the whole system erodes. An interface that shows its reasoning gives someone a chance to catch that before it spreads.
The third, and the one most often skipped, is whether using the tool actually feels good. Seventy three percent of office workers expect AI and automation to make their jobs more satisfying. At the same time, sixty percent say their current tools frustrate them regularly, and more than half say that frustration affects their morale. Those two numbers sit right next to each other for a reason. Interface quality decides which one wins.
The stakes go beyond convenience. Gartner projects that by 2028, a quarter of enterprise security breaches will trace back to AI agents acting outside boundaries nobody enforced, and separately estimates that only one in five companies today has a governance model mature enough to handle autonomous agents at all. Closing that gap has less to do with writing a new policy and more to do with designing how the AI shows its work in the first place, rather than bolting explanations on afterward once someone asks for them.
There's a competitive dimension too. Research from BCG on companies built for the AI era found them putting up nearly double the revenue growth and more than triple the shareholder return of competitors who haven't made the same investment. That gap compounds every quarter it's left unaddressed, and it has very little to do with which model anyone happened to license.
The organizations getting real value from AI right now aren't the ones with the most advanced model, they're the ones where a person can look at what an agent did, understand why, and trust the answer enough to act on it without re-checking everything by hand. That's a design problem before it's a technology one, and most roadmaps still treat it as an afterthought.
