Ideas Made to Matter

Platform Strategy

How to shape platform strategy for humans and AI agents

Beth Stackpole
4 minute read

What you’ll learn: 

  • Surging adoption of AI agents has significant ramifications for how platforms are designed.
  • “Data governance in the AI era is really orders of magnitude harder than it was,” according to MIT Sloan School of Management professor Pierre Azoulay.
  • Platform design now needs to accommodate key decisions around agent access, agent identity, and liability.

Many of the past decade’s great business ventures were platform plays — think Amazon, Uber, or Airbnb. As entrepreneurs and businesses consider this strategy for their next endeavor, they need to understand the platform’s role and how AI agents might shake up this field.

That’s the focus of Platform Strategy: Designing for Humans and AI Agents, an MIT Sloan Executive Education course taught by professors Pierre Azoulay and John Horton. The four-day live online course is designed to help attendees understand the difference between platforms and traditional products and services, including when and where a platform strategy makes sense. 

Surging adoption of AI agents has significant ramifications for how platforms are designed. The course tackles these issues through an exploration of agents’ impact on search costs and transaction costs. It also offers guidance on coring — designing a platform before any participants join it — and governance decisions for hybrid platforms, where humans and AI agents interact.

“Two-sided markets, network effects, and governance — none of these things are particularly new, but there are things happening to make the platforms of today both more interesting and more complex,” said Azoulay, an MIT Sloan professor of technological innovation, entrepreneurship, and strategic management, in a webinar previewing the course.

“The [basic principles] don’t stop working in an AI world, but the questions you ask become more pointed, sharper, maybe more difficult to answer,” he said.

Getting platform basics right

Despite AI’s promise of transformation, it doesn’t change the basic role of a platform. Platforms serve two purposes, Azoulay said: 

  • Serving as a shared technology foundation on which others can build using standard interfaces.
  • Playing a market intermediation role, delivering mechanisms to reduce frictions that prevent participant interactions. 

Platforms typically reduce search costs by helping participants more easily find and match to a product or service (think eBay for shopping) or reduce transaction costs by handling payments, contracts, or logistics more cost-effectively (think Airbnb or Uber) than individual businesses doing that work on their own.

Nearly every platform aims to build network effects, Azoulay said: As more people join, the platform becomes more valuable. Those effects can be same-side, as with WhatsApp or Dropbox, which become more useful as more contacts join, or cross-side, as when having more Airbnb hosts makes the platform more valuable to guests, and vice versa.

The impact of AI agents

Introducing AI agents can have both positive and negative impacts, depending on the situation, Azoulay said. While initial search results can go faster with AI agents narrowing the field, there is potentially more work involved in evaluating recommendations that are optimized for agent ranking, not human needs.

At the same time, settlement and escrow can become cheaper with automated verification, dispute triage can be automated, and cross-language transactions are easier to set up. But other costs, such as proving that a participant is human, may rise.

Some transaction costs are completely new, Azoulay said, such as coordination failures between agents, multiagent error cascades, and synthetic dispute generation — “frictions that basically didn’t exist when every participant was human.”

AI changes the data equation

Nearly every AI pitch deck these days claims to have a data moat — unique or proprietary data that rivals can’t easily replicate — but most moats are illusions, Azoulay said. 

Anyone, including foundation model providers, can now scrape public data, and competitors can use synthetic data to approximate a company’s specialized data. 

AI agent and mobile device

Platform Strategy: Designing for Humans and AI Agents

Attend Online

And the moment that an employee pastes confidential customer data into a public chatbot, “that data is somewhere you did not authorize,” Azoulay said. “Data governance in the AI era is really orders of magnitude harder than it was.”

Data can still be a real moat in three cases, he said: 

  • When a platform has continuous interaction data exclusive to it, such as Uber’s ride matching or Stripe’s transaction logs.
  • When it can legally collect data that competitors can’t, as in regulated industries like healthcare and finance.
  • When specialized know-how is needed to find, clean, and structure the data.

Coring decisions matter more than ever

Coring is everything decided about the platform’s architecture — the rules, access points, and control mechanisms — in the period before a single user has joined. A successful core provides direct value and sets the scene for network effects. It also ensures that companies can extend their platform with complementary capabilities while still maintaining control.

With the advent of agentic AI, it’s important to revisit coring decisions. Platform providers need to work through strategic coring issues like:

  • Agent access, including which AI agents are allowed on the platform — those developed or sanctioned by the platform, or any preferred AI assistant.
  • Agent identity, to authenticate and build trust into machine-to-machine interactions.
  • Liability — determining who is responsible when a transaction involving an agent goes wrong.

Because coring decisions are sticky — that is, difficult to change once the product is live — platform developers must consider every nuance in advance. “You can make some peripheral changes, but by and large, once you’ve launched, it becomes very costly and risky to change your core,” Azoulay said.

Looking ahead, Azoulay doesn’t see a particular platform model holding court. He believes that there will be a variety of models, including human-agent hybrid platforms, those run by agents only (agent-to-agent), and platforms where agents aren’t allowed. 

His advice is to bake decisions into initial platform design and then stay on course. “You can’t [use] a ‘lean startup playbook,’ where you come up with a minimum viable platform and iterate from that point,” Azoulay said. “Changing the deal years in breaks the expectation that brought users onto your platform.”

Shape platform strategy for humans and AI agents


Pierre Azoulay is a professor of management at the MIT Sloan School of Management and a research associate at the National Bureau of Economic Research. His research focuses on the impact of different funding regimes on the rate and direction of scientific progress. His latest projects examine the complex relationship between risk and return in scientific research.

John Horton is a professor of management and an associate professor of information technologies at MIT Sloan. He is also an adviser to Anthropic and the co-founder of Expected Parrot, an open-source platform for designing, running, and analyzing research with digital personas and people. His research focuses on the intersection of labor economics, market design, and information systems. He is particularly interested in improving the efficiency and equity of matching markets.

For more info Tracy Mayor Senior Associate Director, Editorial (617) 253-0065