A study of 199 startup founders suggests AI adoption depends as much on managerial willingness to delegate work as on beliefs about model capability.

Source note: Innessa Colaiacovo and Rembrand Koning. “Delegating to AI: Beliefs and the Organization of Work at New Ventures.” Harvard Business School Working Paper 26-093, June 18, 2026. https://www.hbs.edu/ris/Publication%20Files/26-093_0fabe2ed-1d80-4577-9e75-0ff771256850.pdf

The lazy story about AI adoption says managers will delegate work to AI once the tool is good enough. Improve the model, reduce the error rate, demonstrate task performance, and the organization will naturally reorganize around it.

This paper argues that the story is incomplete. Founders do more than observe AI capability and update a spreadsheet. They choose what work to trust, what workflows to redesign, what hires to delay, and which parts of the company should be run through human judgment, human-AI collaboration, or AI alone.

That choice varies a lot, even among startup founders who use AI heavily and operate in firms where the upside should be obvious.

Why This Paper Matters

Most AI productivity research studies workers or tasks. It asks whether a consultant writes a better memo, whether a developer ships code faster, or whether a support agent resolves more tickets with a model beside them.

Those are useful questions, but firms do not adopt technology task by task in a vacuum. Someone has to decide whether a task should be delegated, whether the workflow should change, whether the company should hire fewer people, and whether AI belongs in the operating system of the firm or merely in employees’ toolbelts.

Colaiacovo and Koning study that managerial layer directly. Their sample is 199 founders of North American technology startups. These are not AI skeptics in sleepy organizations. The firms average 16 employees, are about 3.5 years old, over half have raised venture capital, most are software producers, and 66% offer a product or service with a GenAI component.

Even in that AI-forward sample, the results are uneven. Some founders believe AI lets them run much leaner companies. Others report little or no labor saving. The difference is not explained only by whether they think AI performs well on a specific task. It also reflects whether they are willing to delegate work to it.

The Idea in Plain English

AI adoption is a delegation problem as much as a capability problem.

A manager can believe an AI system is competent at customer service and still hesitate to let it handle real customer interactions. Another manager can see the same capability and decide that the right answer is not AI alone, but a human working with AI. A third might use AI informally for drafts and search, while never redesigning the workflow around it.

The paper’s practical distinction is between AI access and AI integration. Informal use means employees have AI tools and use them when convenient. Formal integration means the firm deliberately builds AI into internal workflows.

That distinction matters. The founders who formally integrated AI report much larger perceived savings and more changes to hiring and management than founders whose firms use AI informally.

What the Researchers Tested

The paper has two main empirical pieces.

First, the authors surveyed founders about how generative AI is used inside their firms and how much additional labor they believe they would need to produce the same output at the same quality without it. This part documents heterogeneity in perceived AI impact and compares firms with formal AI integration against firms where AI use is informal.

Second, the authors ran a survey experiment around customer-service delegation. Founders were asked how often AI could handle customer-service inquiries as well as a skilled human, then how likely they would be to delegate that work to three modes: a human alone, a human working with AI, or AI alone.

The founders were then randomly shown either optimistic or pessimistic information about AI performance in customer service. The authors measured whether this information changed performance beliefs and delegation preferences.

This design lets the paper separate two things that are often blurred together:

  • A belief about whether AI can perform a task well.
  • A managerial preference about whether that task should be delegated to AI.

What They Found

The paper’s headline finding is that startup founders report large average AI benefits, but the benefits are distributed very unevenly.

Among founders using AI, the average founder estimated that without generative AI the firm would need 55% more employees to maintain the same output and quality. But the median estimate was only 17%, and the most common answer was zero. About 30% of founders said they would not need any additional employees without AI.

That is the shape of the result. AI is transformative for some firms and marginal for others, even within a founder-heavy technology sample.

Formal Integration Is Associated With Bigger Perceived Gains

Founders who formally integrated AI into workflows reported much larger savings than informal users. They estimated that without AI they would need 91.4% more employees on average, compared with 24% for firms where AI use was informal. Their estimated hours savings were also higher: 82.8% more hours needed without AI, compared with 33% for informal users.

The same pattern appeared in hiring and management. Formal integrators were more likely to say AI helped them avoid or postpone hiring, hire or build AI agents, hire for different skills, train employees, reduce headcount, or create new roles.

They were also more likely to report management benefits beyond routine automation: better team coordination, more management advice or resources, easier onboarding, and fewer escalations because employees could solve problems with AI support.

The paper is careful here. This evidence is not causal. Firms that formally integrate AI may differ in other ways. But the association supports the paper’s main claim: access alone is not the same as organizational adoption.

Founders Prefer Human-AI Collaboration

In the customer-service vignette, founders were not choosing only between people and AI. The dominant mode was collaboration.

On average, founders reported a 74.2% probability of delegating customer-service work to humans working with AI, compared with 42.2% for humans alone and 39.3% for AI alone.

That pattern matters because it cuts against the clean replacement narrative. For these founders, AI adoption often means changing the work relationship, not removing the human by default.

Capability Beliefs Explain Only Part of Delegation

Founders who believed AI could handle more customer-service inquiries as well as a skilled human were more willing to delegate to AI. That part is intuitive.

But performance beliefs explained only part of the variation. Among founders with similar beliefs about AI capability, willingness to delegate still ranged widely. For example, founders who believed AI was at human parity on about half of customer-service inquiries still differed sharply in whether they would delegate work to AI alone.

The information treatment confirms the point. Optimistic information raised performance beliefs by about 5 percentage points relative to pessimistic information and increased willingness to delegate to AI alone by about 4 percentage points. Yet large residual variation remained after the update. Revised performance beliefs explained only a minority of revised delegation preferences.

In plain English: better evidence that AI works moves managers, but it does not make them all behave the same way.

Why It Happens

Delegation is not a narrow technical decision. It bundles together risk tolerance, trust, customer expectations, team identity, workflow design, accountability, and management style.

A founder may believe AI performs well and still worry about brand damage, error recovery, internal learning, customer relationships, or loss of visibility into the work. Another founder may see the same technology as a way to compress coordination, onboard people faster, and delay hiring. Both can hold similar beliefs about model performance while making different organizational choices.

This is why AI adoption creates uneven firm-level outcomes. The tool may be widely available, but the willingness and ability to redesign work around it is not evenly distributed.

What This Means for Builders

For builders selling AI into companies, the paper is a useful warning: a demo that proves capability is not enough.

The buyer also needs a delegation path. What work should move first? Which human stays in the loop? What does review look like? When does AI act alone? Who owns mistakes? What gets measured after adoption?

Products that answer those questions can become operating systems for work. Products that only expose a model capability risk being treated as optional employee software.

The paper also suggests that “AI agent” products should show the managerial control surface along with the task result. If managers differ in delegation preferences, adoption depends on giving them ways to set boundaries, observe behavior, intervene, and increase autonomy gradually.

What This Means for Buyers and Operators

For operators, the main lesson is to audit delegation as well as tool usage.

It is not enough to ask whether employees use ChatGPT, Claude, Copilot, or internal agents. The better questions are:

  • Which workflows have AI formally integrated?
  • Which tasks are still human-only by habit rather than necessity?
  • Where is human-AI collaboration the right default?
  • Which tasks could move to AI alone with review, logging, or escalation rules?
  • Where are founders or managers blocking delegation because of trust, accountability, or workflow ambiguity?

The paper also explains why two companies with the same tools can see very different returns. One treats AI as a faster assistant. The other uses it to redesign coordination, hiring, onboarding, and task ownership. The second path is harder, but it is where the larger reported gains appear.

What to Watch Next

The next wave of research should connect stated delegation preferences to observed behavior inside firms. The paper’s survey experiment is useful because it isolates beliefs from preferences, but it does not directly observe long-run delegation choices across many workflows.

It would also be useful to know what creates stronger delegation preferences. Is it founder background, prior automation experience, customer risk, product type, team composition, or exposure to better operating examples?

Finally, the field needs more work on governance. If formal AI integration is associated with larger gains, the next question is how companies can formalize AI without creating brittle automation, hidden quality risks, or overdelegation.

Limitations and Caveats

The sample is intentionally specific: North American technology startup founders willing to complete a detailed survey. The findings should not be treated as representative of all companies.

Some results are based on founders’ counterfactual estimates of labor savings rather than measured productivity data. The comparison between formal integration and informal use is associative, not causal. Firms that formally integrate AI may already have different cultures, workflows, leadership styles, or technical capabilities.

The survey experiment focuses on customer service, so the delegation findings may not transfer cleanly to engineering, sales, finance, legal, or product strategy. The information treatment also captures a short-term belief update, not the full process through which managers redesign organizations over months or years.

Still, the paper’s core point is hard to ignore. AI capability is only one input into adoption. The organizational decision to delegate work is its own layer of strategy.

Source

Colaiacovo, Innessa, and Rembrand Koning. “Delegating to AI: Beliefs and the Organization of Work at New Ventures.” Harvard Business School Working Paper 26-093, June 18, 2026. https://www.hbs.edu/ris/Publication%20Files/26-093_0fabe2ed-1d80-4577-9e75-0ff771256850.pdf