You may have seen a 2x2 map in conference slides that says businesses in the AI era come in 4 kinds. It usually comes with a number saying that companies at the center of a network grow faster than everyone else. Put the number and the map on the same slide, and the whole map sounds measured.
The source is a short MIT CISR research briefing called Business Models in the AI Era, published on October 16, 2025. Read the actual paper and every big set of numbers comes from the 2018 version of the map. The AI-era map has no numbers on which companies sit in which box, not a single one. The paper says up front that it is telling a forecast of how business models will evolve ("we are sharing our predictions"). If you never ask which boxes have data, you end up making decisions on a prediction as if it were a measured result.
This article sorts the numbers into 3 piles and labels every one: measured, prediction and not in source. How something was judged to be not in the source is explained at the end of the article.
Part 1The measured half is the 2018 map
The numbers measured on real companies use the Digital Business Models framework that CISR published as a book in 2018. It sorts companies into 4 types.
- Supplier sells through other companies, such as a manufacturer, or an insurer that sells through agents
- Omnichannel sells directly, both online and in physical stores, such as a retailer or a bank
- Modular Producer makes a product or service that others can plug in and use right away, such as a payment provider
- Ecosystem Driver is the first place customers think of when they need to handle one whole area of their life, such as buying a home, and then connects them to other providers
CISR collected data on 2,378 companies in total, in 2013, 2019, 2022 and 2025, asked the same set of questions each round, and placed each company in one model. measured Here is how 12 years played out.
Supplier fell from 46% to 15%, Omnichannel from 24% to 4%, Modular Producer rose from 18% to 23%, and Ecosystem Driver rose from 12% to 58%. By 2025, Ecosystem Driver was the only model in that round's sample whose revenue grew faster than its industry average, by 6 percentage points. measured
These numbers come from real answers, but measured does not mean the years can be compared directly. They come with 4 limits.
- The middle years are missing CISR collected data in 2019 and 2022 too, but the paper only gives the first and last figures. not in source That is why the lines in the chart are straight: they do not tell you when the shift happened.
- Each round is not the same set of companies, and it shrank a lot There were 1,311 companies in 2019 and 152 in 2025. So part of Ecosystem Driver's rise from 12% to 58% may reflect who answered each round, not only companies changing model. The paper reports neither a response rate, nor how many companies answered in more than one round, nor any weighting. Omnichannel at 4% in 2025 is roughly 6 companies (Productize's own calculation), and the claim that revenue grew 6 points above the industry average also comes from 2025 alone.
- Some claims have no numbers behind them The share of companies that lead or take part in an ecosystem rose from 30% to 81%. measured But the rest of that sentence, "with commensurate increases in revenue growth and net profit", has no number attached. not in source And the 30% to 81% figure counts every company that leads or takes part in an ecosystem, while 12% to 58% counts only companies placed in the Ecosystem Driver model, so the two are different numbers.
- One company, one label A footnote says so itself: "The methodology places each company into one business model, but companies can operate in multiple business models"
The revenue figure only tells you that in 2025, being an Ecosystem Driver and revenue growing faster than average showed up together. It does not tell you that being an Ecosystem Driver made revenue grow. This half is real data, yes, but it is data from the digital era and says nothing about what the AI era will look like.
Part 22 new questions and 4 boxes nobody has measured
The AI-era map asks a company two questions. The first is how far you act for the customer. You either help customers reach the result they want, which the paper calls assist, or you carry it out for them on your own within agreed limits, which it calls represent.
The second is how the work runs inside the company. The first way is called structured: the steps are set in advance and AI follows them, with people reviewing and approving every decision before it is carried out. The second is called adaptive: you set the result you want and let AI put the steps together itself. People own the goals and the limits, but do not approve each decision one by one.
These two questions replace the ones on the 2018 map, which asked how well you know your end customer, and whether a company passes goods along a chain or sits in an ecosystem. Cross the two questions and you get 4 boxes.
The paper's examples come from 3 lines of business: financial services, which CISR uses as a hypothetical example; retail, which is Amazon; and telecoms, which is One NZ, a New Zealand provider and the only case study in the paper.
- Existing+ The existing business with AI added. A financial services company uses AI to give personalized investment advice. One NZ uses agents to answer customer questions, fully resolving over 60% of queries from prepaid customers and 40% from enterprise customers, where it is still in beta. company self-reported, no baseline
- Customer Proxy Acts for the customer along a set process. The financial services company has AI manage a customer's portfolio on its own within set limits. One NZ's agents can already upgrade plans, raise disconnection requests and open support tickets, with a person still approving. The paper says One NZ is moving toward this box.
- Modular Curator Helps the customer by bundling services from many providers. The financial services company assembles investment, insurance and credit from several providers into one bundle. One NZ used a set of agents during a major storm to check power failures and recommend options for people to decide on. But that example is internal work; it does not bundle several providers' services for customers, as the box is defined.
- Orchestrator Delivers the result for the customer by coordinating many providers. The financial services company manages the portfolio toward long-term goals without the customer having to ask. Amazon's "Buy for Me", in beta, lets a customer tap a button and Amazon buys the item from another brand's website. The paper cites Amazon's own news page, with no usage numbers. For One NZ, this box is still a plan for the future. prediction
So where did this map come from? A footnote explains that in 2025 CISR brought together 4 researchers, one of them a former CIO, the head of IT at a company, to think through how the way business models are sorted would change as AI advances. It says: "We then tested the framework by presenting and discussing it in a series of senior executive meetings." The paper does not say how many meetings or how many executives. not in source And there was no survey of which companies sit in which box.
One more thing slides tend to add on their own: arrows from the old models to the new boxes, such as Supplier turning into Existing+. The paper never maps old to new. not in source If you see arrows like that on a slide, that is the slide maker's interpretation, not a research finding.
Part 3Where the real 2025 money went
CISR did not measure this, so we have to look elsewhere. A December 2025 report by Menlo Ventures, a US venture capital firm, surveyed 495 AI decision-makers at US enterprises between November 7 and 25, 2025, then combined the results with its own market-sizing model. survey-based estimate
Menlo estimates that enterprises spent $37 billion on generative AI in 2025, 3.2 times the $11.5 billion of 2024. In the category of AI that helps work in every department, worth $8.4 billion, there are copilots, assistants that work alongside people such as ChatGPT Enterprise or Microsoft Copilot; agent platforms such as Salesforce Agentforce or Glean; and personal productivity tools such as Granola. Copilots took 86%, or $7.2 billion, while agent platforms got 10%. That 86% is a share of that category only, not of the total spend.
On what is actually running, Menlo writes that "Only 16% of enterprise and 27% of startup deployments qualify as true agents", by Menlo's definition of AI that plans, acts, observes the result and adapts.
Productize reads this money as landing almost entirely in the Existing+ box, because a copilot helps people do their existing work the existing way. Acting for customers, or letting AI set its own steps, does not show up in spending figures yet. But that conclusion is Productize's reading, placing Menlo's numbers on CISR's map. Menlo and CISR do not cite each other, and CISR has not endorsed this reading. Menlo itself predicts: "General-purpose copilots dominate today, but as agents become more powerful, we can expect a shift from assistance to automation."
This source has 3 limits. The sample is US enterprises only, the figures leave out AI built into existing software, and Menlo is an investor in Anthropic and in several companies ranked in the report. The newest data does not disprove the prediction. It only shows that in 2025 most of the money had not yet reached the boxes the map predicts.
Part 4How to use this map without getting lost
Start with the measured half. Ask where most of your revenue comes from: selling through others, selling directly across channels, making something others build on, or being the first place customers think of and then sending them on to other providers. Picture an accounting firm that gets almost all its clients through referrals from a software company: that is close to Supplier. If the same firm is where clients walk in on their own, and it passes them on to a lawyer or a bank, that is close to Ecosystem Driver (a Productize hypothetical). That answer has numbers measured on real companies behind it.
Use the AI-era map as a set of questions, not a target. It is a question about where you position yourself, not an instruction to invest more. Even if the 4-box map turns out to be wrong, 2 pieces of work are still worth doing.
- Charge for finished work, not for the software The diagram in the paper says a business that acts for customers earns from the results it delivers for them, not from selling products and services. For the accounting firm in the example above, that means charging for each set of books it closes for a client, instead of selling the client software to do the books themselves (a Productize hypothetical). The starting point is selling the work, not the tool, which is worth doing whether or not CISR's prediction comes true.
- Be able to say who decides the work is done Once you charge for results you have to answer who decides the work is done. That question does not wait for the map to be right.
Letting AI set the steps of the work itself, without a person approving each one, is still an open question. Whether that work actually runs shows up on-site at the customer, not on a slide.
So how do you know the business has really changed? A claim and a measured result are different things. If you say you now act for your customers, you should be able to point to which part of your revenue is charged on results. If you say AI now sets the steps itself, you should be able to point to which kinds of decisions AI makes without a person approving each one. If you cannot point to it, you are running the same business under a new name.
Next time you see this 4-box map on a slide, ask what evidence backs the box someone says they have moved into. And if one day you are the one saying it, what will you put on the table?
Sources and references
- Peter Weill, Ina M. Sebastian, Stephanie L. Woerner, Gayan Benedict, Business Models in the AI Era, MIT CISR Research Briefing, Oct 16, 2025 source
- Peter Weill, Stephanie L. Woerner, What's Your Digital Business Model?, Harvard Business Review Press, 2018 source
- Amazon News, Amazon's new 'Buy for Me' feature helps customers find and buy products from other brands' sites source
- Menlo Ventures, 2025: The State of Generative AI in the Enterprise, Dec 9, 2025 source
The not in source label means Productize searched the paper's text, its diagram and the transcript of its audio edition, and did not find it there. CISR's members-only companion files were not read, so not finding something is not proof that it does not exist.
Read next: All articles · Y Combinator's playbook is your competitor's plan · Outcome pricing sounds fair, until you ask who decides done · Before you buy AI, pick the problem and how you will measure it