Every few months, AI resets the bar.
New models show up. Costs drop. Capabilities expand. At this point, the capability argument is over – no one seriously disputes what AI can do.
What's harder to explain is what's happening inside most organizations right now. Tabs still get switched. Files still go missing. Someone is still asking who has the latest update. The tools got smarter. The friction didn't go away.
Dheeraj Pandey, co-founder and CEO of DevRev, framed the stakes plainly: almost $2 trillion in SaaS market value has been wiped out in the last six to nine months.
Referring to the research from analyst Nate Jones, Dheeraj put a number for the gap between the promise and what's actually happening inside businesses: the distance between "AI can do this" and "AI does this usefully, right here" is $4.5 trillion.
That's not a technology gap. The models exist. The gap is everything around them – the systems that don't connect, the data that doesn't flow, and the teams that never quite get AI as a participant rather than a product.
So the question isn't whether AI works. It's why it keeps stopping short of where the work actually happens.
The illusion of AI capabilities
There's a moment in every AI conversation where everything feels solved.
You see a demo. AI searches, answers, suggests actions – instantly. And you think: this is it.
But that moment doesn't last. Because in a controlled environment, with the right prompt and the right inputs, AI looks extraordinary. Real work isn't a controlled environment. It's messy, layered, and deeply contextual.
That's where the cracks begin to show – across three specific failure points that most organizations haven't named yet, let alone solved.
Failure node 1: AI has a general capability, but no specific context
AI today is trained on vast amounts of information. It knows patterns, language, and logic at scale. But it doesn’t know your business unless you make that possible.
It doesn't know why a particular customer issue matters more than another. It doesn't understand how decisions were made across teams. It doesn't carry forward the context of past interactions unless that context is made available to it.
So while it can generate answers, it often lacks what makes those answers useful: context.
That’s why this shift, again from Dheeraj Pandey, matters:
“Now the human is saying, you search, you answer, act. But overall, the fact that the machine is saying, I'm willing to do more for you, I'm not going to throw point – you point, you click, you scroll. But I'm going to search for you, I'm going to do answers for you, and I'm going to act on your behalf”.
It’s a shift from tools to action. But that shift only works if AI can understand what’s actually happening within your business ecosystem.
You can see this clearly in fraud detection at HDFC Bank.
In December, India processed 21.6 billion UPI transactions alone – more than 50% of all global digital transactions. Suchi Mahajan, SVP of Fraud and Analytics at HDFC Bank, works inside that number every day.
The scale isn't the hard part. The hard part is that fraud doesn't announce itself.
"Fraudsters have moved away from hacking the system to hacking the human," she explained. They use emotions – greed, fear, urgency – to get people to authorise transactions themselves. Which means the signal is no longer in the transaction. By the time a fraudulent payment clears, the story that enabled it has already been written.
The fraudulent transaction is the last step. AI that only sees that step misses everything.
Failure node 2: The integration gap no one wants to talk about
Even when organizations recognize the need for context, they run into a more practical problem of where that context actually lives.
Enterprise data isn’t neatly organized. It's scattered across tools, locked in legacy systems, and fragmented across teams that operate independently. Every system holds a piece of the puzzle, but nothing connects cleanly. And AI, no matter how advanced, can only work with what it can access.
This is where most AI initiatives break down. Not because the model isn't good enough, but because the system around it isn't ready.
You see this in how TCC Global approached AI. Damien Katris, Director of AI Strategy at TCC Global, started with a question most companies skip: what's actually slowing us down?
The answer was simple. His team was spending 22% of their working week – a full day, every week, per person – just looking for answers that already existed somewhere in the business.
Not a model problem. A systems problem. The answers were there. They just lived in separate places; no one had connected.
So that's where they started. Not with an AI strategy, but with integration. Once the data sources were linked and accessible, the results were hard to argue with: search time dropped from 22% to 2%. That’s a 92% reduction, and team output rose 47%.
"It feels like getting a day a week back," Damien said. "And that's not metaphorical. That's literally what happened."
That's the integration gap in concrete terms. Not a technology failure – an infrastructure one.
Failure node 3: AI never makes it to the team
Even in organizations that manage to solve for context and integration, there’s a final hurdle that often goes unnoticed.
AI remains confined to the tool layer.
It becomes something people “use” occasionally rather than something that actively participates in workflows.
Solving for context and integration still isn't enough if AI stays at the surface – something individuals query occasionally, then set aside. That's the tool layer. And most AI deployments never leave it.
The reason is structural. Real work doesn't happen in isolation. It moves between roles, across systems, through decisions that require shared context. A tool sits on a desk. A teammate moves with the work. The distinction sounds semantic. At IndiGo, it's operational.
Varun Gupta, who leads architecture and innovation at the airline, described what lands on his central payment desk: not the easy cases, since those are already fully automated. What reaches a human agent is the 0.3% that automation can't resolve. A customer on the phone, anxious, with money deducted and no ticket issued. The agent has to investigate a payment system, a booking engine, an inventory system, and production logs – all simultaneously – while keeping the customer calm, following SOPs, and making a judgment call on whether to refund or rebook at a changed price.
"There is an element of intelligence and reasoning which that human is applying," Varun said. "And I would want Computer to do exactly the same."
That's the line worth sitting with. Not: can AI handle more automation? But: can AI enter the reasoning layer and the part of work that used to stop it cold?
At IndiGo, Computer, by DevRev, now searches across those four systems, surfaces context in the right sequence, and executes the action. The agent isn't stitching anymore. They're deciding. The judgment stays human. The coordination doesn't have to.
That's what it means for AI to make it to the team – not more automation, but presence inside the moments that actually require thought.
So, where is the $4.5 trillion really stuck?
Not in the models. The models are ready.
It's stuck in a simpler problem: most organizations are asking AI to perform before they've given it anything to work with. No shared context. No connected systems. No presence inside the moments where judgment actually gets exercised.
The companies starting to close that gap aren't doing it by finding better AI. They're doing it by treating integration, context, and team-level adoption as the actual work – not the prerequisites that someone else will handle.
The $4.5 trillion isn't locked behind a capability breakthrough. It's waiting on an infrastructure decision that most leadership teams keep deferring.
