AI Value Creation: Five Provocative Propositions

AI adoption is now widespread—but adoption is not the same as value creation. As with earlier waves of technology—think social, mobile, analytics, cloud, and the Internet of Things[foot]A decade ago, MIT CISR’s Jeanne Ross and her research collaborators published insights from studying these and other powerful, readily accessible technologies in J. W. Ross, I. M. Sebastian, and Cynthia M. Beath, “How to Create a Great Digital Strategy,” MIT CISR Research Briefing, Vol. XVI, No. 3, March 2016, https://cisr.mit.edu/publication/2016_0301_GreatDigitalStrategy_RossSebastianBeath.[/foot]— AI foundation models and agents are capabilities that every one of your competitors can also leverage. The winners, then, may not be the fastest adopters, but the companies that most effectively turn AI into differentiated value: systematically identifying where it can address customer and internal needs; innovating and adapting the enterprise, including roles, workflows and metrics, to create and capture value; and continually course-correcting based on feedback.

For this briefing our team[foot]This briefing was written by MIT CISR researchers Stephanie L. Woerner, Peter Weill, Nick van der Meulen, and Ina M. Sebastian; academic research fellow Martin Mocker; and industry research fellows Gayan Benedict and Peter Reynolds.[/foot] crafted five propositions about effective management practices to drive AI business value, based on current MIT CISR research focused on trust, skills, data, AI agent proliferation, and value.

Trust Cannot Be Managed—It Must Be Built into Infrastructure

Traditionally, companies have built trust through relationships, relying on reputation, track record, shared incentives, social contracts, and human accountability. These foundations still matter. But when companies introduce AI into workflows, they cannot rely on trust alone to ensure reliable outcomes.

Autonomous AI agents—increasingly developed by third-party providers and operating within your systems and workflows—arrive with no relational history, no social stakes in upholding your interests, and no human accountability to you. They operate at machine speed and a scale that no existing relationship can govern. Yet you must work with them at that speed and scale.

When agents transact too quickly for human relationships and judgment to provide effective oversight, companies must replace reliance on earned goodwill with auditable, contestable technical controls for fairness, compliance, and trustworthiness. You cannot engineer an external agent’s good intentions, and the interests of the principal it serves may diverge from yours at any time. But you can engineer controls around the agent’s behavior: whether it performs its stated function reliably, follows defined rules, stays within established boundaries, and leaves an auditable record of its actions—and whether you can reverse, dispute, or appeal its decisions after the fact.

We studied a financial services company that deployed an AI lending agent from a third-party vendor inside its credit decision process. Six months in, the agent began exhibiting subtle but systematic bias in its weightings—not from malice, but from a drift in the training data. The company had to isolate the agent midstream. While the vendor and the company had a good relationship, it alone could not surface problems quickly. To do so, the company needed to build trust into the infrastructure with audit trails, contestation protocols, and a dashboard continuously monitoring performance.

Companies that translate relational trust into auditable infrastructure will be able to work with third-party agents at speed. Those that don’t will face one of two consequences: bottlenecks from keeping humans in the loop to maintain “trust,” or liabilities from trusting agents they cannot verify or contest.

The Skills Crisis Isn’t About AI Training; It’s About Losing Expertise

AI upskilling providers are flooding talent markets with programs to train existing staff. The logic seems obvious: teach people to use AI tools and they’ll be more productive. But this approach is solving for the wrong problem and may deepen the expertise crisis.

Companies need people who can (1) effectively operate AI systems and (2) make non-routine judgments about when and how to use such systems. Those in the second group—domain experts with deep contextual judgment—are scarce and difficult to replace.

One executive we interviewed described a golden rule at their organization: “Do you have the expertise to challenge the output? If not, you are not allowed to use generative AI in your solution.” This is wise—but creates a bottleneck. Who has that expertise, and where is it located? Right now, it is centered in the people doing the most complex, non-routine, highest-value work. If you ask them to spend a big percentage of their time validating AI outputs, they’re not doing what they were hired for. And if you don’t ask them to validate outputs, you won’t know whether those outputs are trustworthy.

Generative AI produces output that is fluent, confident, and competently formatted; errors don’t carry the sloppiness that once signaled them. A reviewer without deep subject matter expertise cannot reliably spot these errors.

Where are you losing subject matter expertise faster than you’re developing it? And how will you preserve—or build—that expertise for the decade ahead? The winners will be companies that protect deep expertise from efficiency pressure. They will maintain some people doing the complex, non-routine work that builds judgment, even when AI could do parts of it. This is not inefficiency; it is insurance against confidently delivered errors with potentially catastrophic consequences.

The Way to Win at Plug and Play Is with Indispensable Data

In MIT CISR’s digital business models research,[foot]For an overview of this research, see Peter Weill and Stephanie L. Woerner, What’s Your Digital Business Model?: Six Questions to Help You Build the Next-Generation Enterprise (Harvard Business Review Press, 2018).[/foot] the modular producer model—in which a company provides plug-and-play products or services through other companies’ digital ecosystems—has become strategically important as ecosystems increasingly rely on such offerings.[foot]P. Weill, I. M. Sebastian, S. L. Woerner, and G. Benedict, “Business Models in the AI Era,” MIT CISR Research Briefing, Vol. XXV, No. 10, October 2025, https://cisr.mit.edu/publication/2025_1001_BizModelsAIEra_WeillSebastianWoernerBenedict.[/foot] But modular producers’ average results are unremarkable, and often these providers are substitutable: The ecosystem leader can swap them, and the customer barely notices.

We found that top-performing modular producers focus on three differentiators—continuous innovation, connectability, and brand strength[foot]S. L. Woerner, I. M. Sebastian, and P. Weill, “All Companies Need a Modular Producer Strategy,” MIT CISR Research Briefing, Vol. XIX, No. 2, February 2019, https://cisr.mit.edu/publication/2019_0201_ModularProducer_WoernerSebastianWeill.[/foot]—with a goal of becoming irreplaceable. In the AI era, companies we interview are reporting a fourth differentiator: proprietary data that is not replicable. Such data is the asset that converts a plug-in service or product into infrastructure.

Belden, an industrial networking company, saw its traditional networking hardware becoming commoditized as value shifted toward AI and other advanced technologies. Rather than compete with technology giants on AI itself, Belden focused on providing trusted, edge-specific data from industrial environments. Belden’s sensors and networking hardware capture operating conditions at a granularity and provenance that are hard to reproduce from outside the plant floor. The role is deliberately narrow: too edge-specific for a giant to bother replicating, too critical to their AI to bypass. Today, in an NVIDIA solution to which Belden contributes, Belden complements NVIDIA and remains indispensable.[foot]M. Mocker and I. M. Sebastian, “The Strategic Role of the Partner Ecosystem in Belden’s Transformation from Products to Solutions,” MIT CISR Working Paper No. 469, July 2026, https://cisr.mit.edu/publication/MIT_CISRwp469_BeldenPartnerEcosystem_MockerSebastian. [/foot]

Whether you can build AI is no longer a strategic question; increasingly, every company can. What matters is what others cannot build without you.

No One Will Remember How Many AI Agents You Built

With frontier models, everyone has access to the same intelligence, at the same price, on the same day. Prototyping is democratized: some concepts can now be discussed in the morning and turned into a working demo by the afternoon, with less code, fewer specifications, and less human effort. Companies are turning to AI agents for every problem—even those better solved by a conventional algorithm.

But there’s a hidden cost with AI agents. A more autonomous agent has a larger space of possible behaviors and thus a higher burden of proof. The effort saved during prototyping is deferred, not eliminated; it comes due when the agent is prepared for production. Evaluation becomes open-ended. Monitoring becomes supervision by a decision-maker, not inspection of an output. Governance means underwriting a capability, not signing off on a specification. The more autonomous the agent, the harder it is to prove predictable outcomes—and unprovable systems don’t scale, pass risk committees, or reach customers. What looks cheap to build becomes expensive to trust, and companies are optimizing in precisely the wrong direction: accumulating agents they will never be able to deploy.

The uncomfortable consequence is that competitive advantage in AI will not come from the number of agents you build; agents are available to everyone, and a differentiator owned by everyone is a differentiator for no one. Advantage comes from one place only: the ability to take AI to production and scale it, with outcomes you can prove. Evaluation, monitoring, governance, and reliability are the unglamorous disciplines that turn a demo into a dependable system. The winners will be the ones who shipped.

AI Will Disappoint for Years—Then Exceed Expectations

AI is billed as the most transformational technology of our lifetimes. Yet companies are already reporting skepticism, cutting budgets, and asking hard questions about bottom-line value. Such disappointment is not new, but considering rollouts of earlier technologies, it’s typical.

When cars were invented, there were no paved roads, seatbelts, windshield wipers, headlights, fuel stations, driver’s licenses, or traffic lights. The technology alone created little value. Only when thoughtful complements were introduced—new infrastructure, regulations, and skills—did cars become transformational. Value from digital technologies developed in a similar fashion. Companies that reimagined business processes when digitizing captured the most value, while those that simply automated existing ways of working saw minimal returns.

AI requires even more profound organizational change than previous technologies. AI agents, for example, differ vastly from digital systems that execute predetermined processes in that they can make judgments, pursue goals without predefined workflows, adapt, and increasingly work in concert with other agents to solve complex problems. Consider their use in drug discovery: AI agents can shift the work from executing predefined protocols to pursuing goals such as “identify molecules with this property,” fundamentally changing how scientists work.

Nearly everyone tells us they personally get tremendous value from their favorite AI tool. This gap between organizational disappointment and individual enthusiasm suggests the problem is not with the technology but how we’ve structured work related to it.

Getting value from AI, and particularly from agents, requires redesigning workflows, roles, and handoffs between humans and AI. Just like managing human colleagues, managing an AI agent is real work: coaching, correcting, evaluating, and escalating. Like human colleagues, agents need onboarding, regular review, and ongoing calibration. As both humans and agents will make mistakes, companies need real-time learning feedback loops and governance to identify, correct, and learn from them.

In a decade, AI will exceed our expectations. We’ll see genuinely personalized customer experiences, faster drug discovery, more proactive health care, and personalized learning. Yes, some jobs will disappear—but history shows that transformational technologies create new roles we cannot yet imagine.

The bumpy road ahead requires three changes: redesigning work so humans and AI each do what they do best; updating regulation, ensuring it is both enabling and protective; and sharing value equitably.

The Debate Continues

In 2015, MIT CISR published five propositions about managing in the digital economy.[foot]P. Weill, J. W. Ross, and S. L. Woerner, “Thriving with Digital Disruption: Five Propositions,” MIT CISR Research Briefing, Vol. XV, No. 7, July 2015, https://cisr.mit.edu/publication/2015_0701_FivePropositions_WeillRossWoerner.[/foot] A decade later, many have held up: Value chains did become less relevant. Platforms did become table stakes. Those that didn’t hold universally involved organizational trade-offs rather than logical failures.

We expect a similar outcome for these five propositions on AI value creation. We invite you to debate them with your peers and your executive team. What would your company do differently if they are true?

© 2026 MIT Center for Information Systems Research, Woerner, Sebastian, Benedict, Van der Meulen, Mocker, Reynolds, and Weill. MIT CISR Research Briefings are published monthly to update the center’s member organizations on current research projects.