Reflections on MIT’s “State of AI in Business 2025” Report
One statistic that reverberated loudly this year: 95% of organizations are getting zero return on their GenAI investments. Not low returns. Not “early returns.” Zero.
MIT’s State of AI in Business 2025 report makes it clear that this failure is not about models, regulation, or budgets. It is about approach, and more importantly, a profound misunderstanding among business leaders of what AI actually is and how it transforms organizations.
After reading the report and reflecting on my own experience building digital businesses, leading shared service transformations, and running a fintech that processed billions in annual transaction volume, the diagnosis is clear:
GenAI is not failing. Leaders are failing to understand the paradigm shift.
Below are my key insights and takeaways.
Productivity ≠ P&L Impact. This Is Where Most Strategies Die#
MIT notes that tools like ChatGPT and Copilot are achieving widespread adoption, but primarily as individual productivity tools, not drivers of enterprise value. Enterprises confuse activity with impact. Just because your workforce is generating better emails doesn’t mean your company is transforming.
The report highlights a critical disconnect: ChatGPT improves individual output. But has almost no influence on corporate workflows, cost structures, or revenue lines. This aligns with what I see in real-world transformations: organizations are excited to “use AI” but rarely redesign the core processes where real value is created.
AI doesn’t create P&L impact unless it replaces something expensive. Usually external. The MIT report reinforces this. For one best in class organization covered in the study:
- BPO elimination = $2M–$10M annual savings
- Agency replacement = 30% reduction
- Vendor consolidation = millions saved
That’s real money. And it explains why countries like the Philippines, home to massive BPO sectors, should be paying attention. The structural implications are significant.
The Real Divide Isn’t Technical. It’s Organizational#
The report’s central argument is brilliantly summarized: The dividing line is not intelligence; it’s memory, adaptability, and learning capability. AI that doesn’t learn might be more akin to autocomplete with better grammar. This matters because enterprises still operate under a pre-AI architecture. They’re built around static tools, rigid processes, and departments that don’t talk to each other.
But GenAI isn’t a tool. It’s a learning system. And learning systems only deliver value when:
- They integrate into workflows
- They remember context
- They adapt faster than humans can patch them
Most enterprise AI fails precisely because:
- It has no memory
- It can’t evolve
- It breaks in edge cases
- It doesn’t match how the business actually works
Executives evaluate AI like they evaluate SaaS. But AI is not SaaS. It’s a fundamentally different class of technology.
Shadow AI: The Most Important Signal Boards Are Ignoring#
One of the most fascinating insights from the report: 90% of employees use AI for work. Only 40% of companies have purchased an AI subscription.. This is the “shadow AI economy.” Employees have already crossed the divide. Enterprises have not.
This should keep executives awake at night, not because of risk, but because of what it signals: Your workforce knows what “good AI” looks like. And they will quietly reject anything worse.
I’ve seen this firsthand in digital transformations:
- If the official tool is slow → people circumvent it
- If the tool is rigid → spreadsheets reappear
- If the workflow doesn’t match reality → shadow systems emerge
Shadow AI is the new “spreadsheet problem,” but at 100x the speed and complexity.
More importantly: shadow AI reveals the real use cases. The ones your employees already value. Smart leaders will learn from this rather than resist it.
The Successful Startup Playbook Lives On#
MIT’s findings match what I’ve observed mentoring founders and investing in startups, regardless of whether it involves AI or not: winners go narrow and deep. Losers go broad and shallow.
According to the report, startups that cross the divide:
- Start with a single painful workflow
- Achieve visible value in days or weeks
- Learn from feedback immediately
- Embed deeply into existing systems
- Expand outward only after winning a foothold
And importantly: Trust and distribution matter more than the underlying model. MIT notes that successful startups close deals by:
- Leveraging system integrators
- Partnering with BPOs
- Using board and executive referrals
- Riding existing enterprise procurement channels
This is classic enterprise sales. GenAI hasn’t changed it. What has changed is that the strongest moats are being built around:
- Workflow integration
- Domain-specific memory
- Organizational learning
Not around models.
The Enterprise Model Is Already Shifting: From Software to Services#
In a way, enterprises have really always purchased services (albeit packaged as software). The report offers a subtle insight that many will miss:
“Successful buyers treat AI vendors more like BPO providers than SaaS companies.”
This is huge as it makes what was often implied explicit: enterprises don’t want tools. They want outcomes. And they want vendors accountable for those outcomes.
This shift mirrors what I saw at UBX as we moved toward embedded finance and managed services:
- Clients no longer wanted the plumbing
- They wanted the result: instant payments, automated lending, reconciled settlements
- And they wanted commercial models aligned to outcomes, not licenses
GenAI is following the same trajectory. Enterprises will cross the divide when they stop buying software and start buying capability. This aligns perfectly with the concept of AI being packaged as agents. Some solutions even give their AI agents names!
The Next Architectural Shift: The Agentic Web#
MIT closes with a view of the near future that aligns strongly with what I see happening: the rise of the Agentic Web. Systems of interoperable, learning, memory-rich agents coordinating across the enterprise. This is not science fiction. The infrastructure is already here in the form of evolving agent frameworks.
This evolution is exactly what may allow enterprises to:
- Replace BPOs
- Automate back-office operations
- Achieve zero-touch processes
- Generate exponential learning loops
When people talk about “AI replacing jobs,” they often misunderstand. The more relevant shift is that AI may replace entire categories of outsourced workflows before it replaces internal roles.
The implications for countries like the Philippines are profound.
What Leaders Should Do Now#
Based on MIT’s data and my own experience, here are the strategic imperatives:
- Stop Piloting. Start Integrating. Pilots are where AI goes to die. Integration is where value lives.
- Buy, Don’t Build, unless the process is uniquely mission-critical. MIT’s data: external builds succeed twice as often as internal ones.
- Focus on back-office automation. This is where the hidden ROI lives, and where BPO-heavy organizations can gain the most.
- Empower line managers, not central AI labs. Your power users already know where AI creates value.
- Choose tools that learn. Anything static is obsolete before deployment.
AI Isn’t a Technology Problem. It’s a Leadership Problem#
MIT’s research echoes what I’ve seen across three decades of digital transformation in banking, insurance, government, and fintech: AI rewards organizations that are willing to redesign themselves. It punishes those that try to bolt innovation onto old structures.
The GenAI Divide is not a gap in capability. It’s a gap in imagination.
The winners won’t be the companies that try AI. They’ll be the ones that let AI change how they work.
And that window, as MIT warns, is closing fast.