May 19, 2026
8 min read
The AI conversation has matured in the last 18 months. Most executives no longer ask whether AI will deliver value. They ask why so many of their pilots have stalled, why the deployments that did make it to production aren’t generating the ROI the vendors promised, and what separates the companies actually winning with AI from the companies still running experiments.
New research from MIT Technology Review Insights, conducted in partnership with Celigo and based on a survey of 500 senior IT and engineering leaders at mid-to-large US enterprises, gives the clearest answer I’ve seen.
It’s not the models. It’s not the use cases. It’s not even the teams.
It’s the integration architecture underneath.
The MIT report contains a lot of useful data, but one statistic stands out as the single most important finding for any leader thinking about AI investment:
Companies with enterprise-wide integration platforms are five times more likely to use diverse data sources in their AI workflows. Fifty-nine percent of them are running AI on five or more data sources. Among companies without an integration platform, that figure is zero percent.
Read that again. Not 5%. Not 1%. Zero.
This is what separates the companies producing real business intelligence from the companies generating sophisticated-looking noise. AI in isolation, running inside a single application and drawing on a single data source, is a productivity tool. AI with access to multiple, integrated, harmonized data sources is something else entirely. It’s the difference between a chatbot that summarizes one system and a system that can actually answer business questions.
Most companies don’t have the second thing. They have the first thing and are wondering why it isn’t transformative like they were promised during the procurement cycle. In this case the integration platform isn’t just another expense item on the budget sheet, it’s the connective tissue between your entire GTM stack.
A company deploys AI inside their CRM, whether that’s Salesforce Agentforce, HubSpot Breeze, or whatever the platform calls it. The AI is genuinely capable. It summarizes calls, scores leads, drafts follow-ups. The reps like it. The leadership team sees activity in dashboards. It feels like progress.
Then someone asks a real business question. “Which renewals are at risk this quarter, and why?”
The CRM AI can answer part of that. It can see deal stage, engagement frequency, and recent activity. What it can’t see: the support ticket trend in Zendesk that shows escalating frustration. The payment delays in NetSuite that signal procurement-side issues. The product usage drop captured in your data warehouse. The conversation sentiment in Gong recordings. The CSAT scores in Gainsight.
The signal is real. But it lives in five systems, and THAT AI can only see one of them.
This is the gap the MIT research quantifies. AI deployed inside individual applications produces locally optimized output. AI deployed across an integrated data layer produces actual business intelligence. The difference isn’t the model. It’s the architecture feeding it.
A subtle but important point that doesn’t always come through in coverage of research like this: integration platforms are necessary, but they aren’t sufficient.
Connecting systems is plumbing. It moves data from one place to another. What integration platforms can’t do on their own is define what the data should mean once it gets there. They can’t decide which system is authoritative when Salesforce says a customer is “Active” and Gainsight says they’re “Churning.” They can’t enforce a single definition of “qualified” across Marketing, Sales, and CS. They can’t design how a renewal-risk signal should flow into an action.
That work is architecture. It’s the layer above the plumbing. And it’s where most companies fall short, even those that have invested heavily in iPaaS tools. This is work that AI tools cannot do at this point and still requires human guidance.
The MIT findings make this distinction visible in the data. Companies using integration platforms for specific workflows see real lift over those without any platform at all, but they don’t approach the results of companies using integration enterprise-wide. The pattern suggests something I’ve been arguing for years: the value of integration scales with how cohesively it’s designed, not how broadly it’s deployed. Bolting on point integrations creates connectivity. Designing an integration architecture creates leverage.
This is also where the operational shortcuts catch up to you. Every undocumented Zapier flow, every native connector, every API integration that someone built two years ago and never revisited, these accumulate. They become the GTM debt that makes AI deployment painful, because the AI has to reason over the inconsistencies you’ve been quietly accruing.
The other major finding in the MIT report that deserves attention: 95% of surveyed companies say their AI workflows already have at least some level of autonomy. 92% expect that autonomy to increase in the next 12 to 18 months.
We’re not talking about AI suggesting actions for humans to take anymore. We’re talking about AI systems that make decisions, take actions, and create downstream consequences, all on their own.
This is where bad data foundations stop being an efficiency problem and start being a real risk to the business.
When AI suggests a wrong answer, a human reviews it and overrides. When AI takes a wrong action, that action has to be unwound. When AI takes a thousand wrong actions because the underlying data was inconsistent, you have a remediation project that can take quarters to clean up.
The MIT data here is again unambiguous. Companies using enterprise-wide integration platforms are 24% likely to assign significantly more autonomy to their AI in the next 12-18 months. Companies without integration platforms? Zero percent. The leaders aren’t pulling back from autonomy. They’re accelerating into it, but only because they have the foundation to do so safely.
This is the part of the AI race that the vendor pitches don’t talk about. Autonomous AI without architectural readiness isn’t an advantage. It’s a liability that compounds and creates risk across your entire organization.
The MIT findings validate something I’ve been arguing for years: the companies that win aren’t the ones with the best tools or the largest budgets. They’re the ones who took the time to build the foundation. The unglamorous, definitional, boring work of getting your data, your integration architecture, and your governance in shape.
What’s different now is that AI is going to expose those gaps faster than they’ve ever been exposed before. The era of papering over bad foundations with spreadsheet massaging, manual reconciliation, and analyst heroics is ending. AI doesn’t tolerate the workarounds your team has been running for years to make broken systems look functional.
That foundation has three components, and most companies have invested unevenly across them.
Clean, governed data. AI amplifies what it’s trained on. If your data has definitional drift across systems, AI doesn’t fix it. AI exposes it at scale.
Integration architecture, not just integration connectivity. Tools that connect systems are common. Architectures that govern how those systems should interact are rare. The MIT data suggests the second is what actually drives outcomes.
Clear ownership of the systems and definitions. AI deployed without clear governance of the underlying data and workflows will drift. The drift won’t be visible until the AI produces an answer no one trusts.
If you’re evaluating AI investments and you’re not confident in those three foundations, you’re not really evaluating AI. You’re evaluating a tool that’s going to surface every gap in your operations at machine speed.
If you’re evaluating AI investments and you’re not confident in those three foundations, you’re not really evaluating AI. You’re evaluating a tool that’s going to surface every gap in your operations at machine speed.
The basics are still the differentiator. They were before AI. They just stopped being optional.
The MIT report goes deeper into each of these areas, covering team structure, autonomy levels by function, and the specific patterns that separate successful deployments from stalled ones. If you’re trying to make the case internally for foundational investment before AI investment, it’s the strongest external data I’ve seen.
We’ve made the full report available, in partnership with Celigo, for anyone serious about closing their own architecture gap.