AI-Era Business Models: MIT Studied 2,378 Companies Over 12 Years — Here’s What They Found
Writen by
nguyen hoang khai
In 2013, 46% of companies worldwide operated as traditional Suppliers — producing and selling products through linear, one-directional value chains. By 2025, that figure had fallen to 15%. Over the same period, the Ecosystem Driver model — companies that build platforms connecting customers with multiple providers — grew from 12% to 58%.
This isn’t a tech media forecast or a consulting firm trying to sell a report. This is data from MIT CISR (Center for Information Systems Research), analyzing 2,378 companies over 12 years, published in 2025 by researchers Peter Weill, Ina M. Sebastian, Stephanie L. Woerner, and Gayan Benedict. And the more unsettling part isn’t the shift itself — it’s what MIT CISR says comes next: AI is forcing every business model to evolve again, right now.
The question is no longer “Is your company using AI?” It’s “To what extent can AI act on behalf of your customers — and are you ready for that?”
The Conventional Wisdom About AI in Business — And Why It’s No Longer Enough
When most executives talk about AI, they’re thinking inside a familiar framework: use AI to do existing work faster and cheaper. Chatbots replacing call centers. AI summarizing documents instead of analysts. Predictive models instead of manual forecasting.
This framing isn’t wrong — it’s just addressing the most visible and lowest-value layer of the transformation. According to McKinsey’s Global Survey 2025, 88% of enterprises are using AI in some capacity. Yet only a fraction of them are achieving financial performance that clearly outpaces their industry.
What MIT CISR’s research reveals — and what most organizations are missing — is that AI isn’t just an automation tool. It’s the driver of an entirely new business model logic. Companies that deploy AI only to “speed up existing work” are competing in a race where the rules have fundamentally changed around them.
12 Years of Structural Shift: How the Four Legacy Models Are Disappearing
MIT CISR classified companies along two axes: depth of knowledge about the end customer and ecosystem scope (standalone versus multi-partner). This produced four models — and their trajectory over 12 years tells a clear story:
- Supplier: Sells products/services through intermediaries, limited visibility into the end customer. 2013: 46% → 2025: 15%
- Omnichannel Business: Serves customers directly across multiple channels with deep customer knowledge. 2013: 24% → 2025: 4%
- Modular Producer: Provides services that plug into other companies’ ecosystems (think Stripe, AWS, Twilio). 2013: 18% → 2025: 23%
- Ecosystem Driver: Builds a platform connecting customers with multiple providers within a customer domain. 2013: 12% → 2025: 58%
The collapse of the Omnichannel model from 24% to 4% is the most instructive signal. It’s not that multi-channel selling became less important — it’s that companies who were purely omnichannel, without an ecosystem layer, are being absorbed by broader platforms. Grab isn’t just “omnichannel” — it’s an Ecosystem Driver that owns the entire customer domain of mobility, food, and payments.
The Agentic AI Era: Four New Business Models MIT CISR Proposes
This is where MIT CISR moves from historical data into what they call “speculative but evidence-informed” territory — built from the 12-year dataset plus case studies conducted in senior executive sessions. In the age of Agentic AI, both classification axes shift entirely:
The new vertical axis is no longer “how well do you know your customer” — it’s “how much can AI act on behalf of your customer.” The horizontal axis shifts from “standalone versus ecosystem” to “predefined process versus no predetermined process.”
Model 1 — Existing+ (Augment Your Current Model)
Keep the existing business model intact; integrate AI into key touchpoints to improve speed, cost, and experience. An insurance company still sells life insurance — but AI now auto-approves policies in three minutes instead of three days, risk models price more accurately, and a chatbot handles 70% of claims inquiries without human escalation.
This is the most immediately practical starting point — and also where most enterprises are stuck, mistaking it for a destination rather than a first step.
Model 2 — Customer Proxy (AI Acts on Behalf of Customers)
AI is delegated to execute specific actions on behalf of customers — within a predefined process. An investment app that automatically rebalances a portfolio monthly based on parameters the customer set. A travel platform that rebooks a cheaper flight when one becomes available, matching the customer’s saved preferences.
The key variable here isn’t technology — it’s trust. Customers must trust the company enough to say “do this for me.” That trust is built over time through track record and transparency, not through AI capability alone.
Model 3 — Modular Creator (AI Assembles the Solution)
AI assembles modules of services — including from third parties — to achieve customer outcomes, without a fixed predefined process. Instead of executing a set of predetermined steps, AI decides how to combine tools to reach the goal. An AI assistant that knows you want to “prepare for an investor pitch” — it researches the firm, analyzes market data, drafts the deck, schedules the room, and sends calendar invites, all without a predefined workflow.
Model 4 — Orchestrator (AI Coordinates an Ecosystem)
The most advanced model. AI coordinates an entire external ecosystem of partners in real time to deliver the best outcome for the customer — unconstrained by any predetermined process. This is the trajectory of today’s leading Ecosystem Drivers: instead of customers navigating multiple providers, an AI Orchestrator analyzes needs, selects optimal providers dynamically, negotiates terms, executes, and monitors — all in the background.
According to Gartner, by 2028, 33% of enterprise software will incorporate agentic AI — up from less than 1% in 2024. The Orchestrator model is where that statistic is pointing.
Enterprise AI Maturity: Why 70% of Companies Are Burning Money on AI
Alongside the business model research, MIT CISR also surveyed 721 enterprises to measure AI maturity and its direct financial impact. The results are unambiguous:
- Stage 1–2 (isolated experiments, selective scaling): Financial performance below industry average
- Stage 3 (industrializing AI enterprise-wide): Financial performance clearly above industry average
- Stage 4 (AI embedded in organizational DNA): Financial performance significantly above average
Most critically: the jump from Stage 2 to Stage 3 produces the largest single financial impact. Stage 3 requires four things happening simultaneously: a scalable AI architecture, data and outcomes made transparent via business dashboards, a pervasive test-and-learn culture, and deep business process automation.
As of 2025, only 31% of companies in the study had reached Stage 3. That means roughly 70% are still at Stage 1 or 2 — investing in AI but not seeing meaningful financial returns, and often drawing the wrong conclusion that “AI doesn’t work for our industry.”
The problem isn’t that AI is ineffective. The problem is that most organizations are deploying AI in a way that practically guarantees it will be.
Common Objections — And Honest Answers
“These models only apply to Big Tech. An SME can’t become an Orchestrator.”
Partially true — Orchestrator requires scale and resources that most SMEs don’t have. But Modular Creator and Customer Proxy don’t. A 50-person SaaS company can absolutely build a Customer Proxy capability within a narrow customer domain. The MIT CISR framework doesn’t require you to be Amazon — it asks you to decide which direction you’re moving in.
“Our customers aren’t ready to let AI act on their behalf.”
This is the most legitimate objection, and MIT CISR acknowledges it directly. Trust cannot be forced — it’s built through demonstrated track record. That’s exactly why Existing+ is the right first step: prove AI works well in a supporting role, then gradually expand the level of autonomy customers are willing to grant.
A note on the framework’s limits: The Agentic AI portion of this research is explicitly built from executive conversations, not longitudinal survey data like the 2013–2025 analysis. The pace of AI change could make these projections wrong in either direction — faster or slower than anticipated.
Three Questions Every Engineering Leader Should Answer Now
If you’re a CTO or Head of Engineering reading this, the MIT CISR framework isn’t an academic exercise — it’s a set of strategic forcing questions:
- Which of the four legacy models does your company currently operate? If the honest answer is “pure Supplier,” you’re part of a cohort that shrank from 46% to 15% in 12 years — and that trend shows no sign of reversing. This is structural risk, not market risk.
- Which of the four AI-era models do you want to become? This answer determines what data infrastructure you need, what AI capabilities to build, and what kind of trust you need to start earning from customers today. Business model choice must precede AI strategy — not the other way around.
- What Stage of AI Maturity are you at? If the honest answer is Stage 1 or 2, the constraint isn’t lack of AI use cases — it’s lack of the foundation to make AI generate real value. Invest in data architecture and test-and-learn culture before buying more AI tooling.
Our Perspective
We work with engineering teams at various stages of this transition. The pattern we observe consistently is: the teams that move fastest aren’t the ones with the best AI tools — they’re the ones with a quality foundation solid enough to iterate quickly without everything breaking.
As you move from Existing+ toward Customer Proxy, deployment velocity increases. As AI begins acting on behalf of customers, system reliability requirements shift from “acceptable” to “non-negotiable.” This is why quality engineering isn’t a cost you trim during AI transformation — it’s the prerequisite that makes the transformation possible.
MIT CISR frames the question of which model to pursue. The follow-on question is: does your quality infrastructure have the readiness to execute on that model?
