What an AI startup can build that a model company won't easily copy

From garden planning to medical referrals, the Founders work group sketches AI products a model company would find hard to copy.

Updated October 13, 2026

A model company can ship the feature. The group would build around what takes sustained work: a group's accumulated taste, a team's workflow, trust between people or a creator's audience.

Collect taste that compounds. Ben Silbermann suggests a product where people keep adding their own taste and intent, like a garden planner that records what each person planted, season after season. For a sensitive task such as naming an emotion, he would add expert review and careful boundaries, drawing on How We Feel, whose team includes scientists and therapists.
Own a team's workflow. Ivan Zhao would pick one complete workflow, such as an agency's client intake and approvals, that a team can use on day one. A ready-made assistant is easy to copy, so he would let the consultants who set it up keep improving it across clients, as consultants did with Notion templates.
Build on trust between people. Reid Hoffman's example is a referral marketplace that needs clinician participation, verified availability and patient consent. He calls a hospital data deal alone a weak moat, and would test whether referrals get better as participation grows.
Measure more than output quality. Emmett Shear points to Kiko, whose opening disappeared when Google shipped a calendar attached to Gmail. For a creator product, he would measure whether creators come back because it helps them earn, get discovered and talk with their audience.
Connect the systems a team already uses. Shishir Mehrotra suggests customer renewal briefs that draw on the account record, support history and roadmap, placed where the account team already prepares. The test is keeping trusted access across those systems and earning regular use.

Group chat with Emmett, Ben, Ivan, Reid, Shishir — the Founders work group

Question:

Where can a new AI startup build something the big model companies are unlikely to copy?

Ben SilbermannFounder and Chairman of Pinterest. Founder of How We Feel

I think a promising place is a product where people keep contributing their own taste and intent. On Pinterest, people saved ideas for a home, a meal, or a wedding, and those choices made future recommendations more useful. The work was understanding each person's interests and helping them act on them, including finding products that fit their taste.

YouTube

A small AI company could start with one group, say people planning a garden, and help them save ideas, compare what fits their space, and record what they actually planted. Each season gives the product more useful feedback about that person's preferences and local conditions. A model company can build a garden chat feature; earning a group of people who return to plan and share their results takes sustained product work. Pinterest had just 3,000 registered accounts three months after launch, and we focused on a small group of deeply engaged users.

forbes

Ivan ZhaoFounder of Notion

I would look for a narrow team workflow where the startup can own the setup and the daily work: for example, an agency's client intake, approvals, and handoffs. Build the records, permissions, and review steps around how that team actually works, then use models to find the right context and draft the next action. A model company can ship a similar assistant, but matching each team's process and earning its place in that process is a different kind of work.

We learned at Notion that giving people flexible building blocks alone was too much setup. People wanted to finish a task, so we packaged the blocks into ready-to-use templates and use cases.

Spotify For an AI startup, I would start with one complete workflow that a team can use on day one, then let the team change it over time. The setup and ongoing use are where the product learns what the team needs.

Emmett ShearCEO of Softmax. Founder of Justin.tv and former CEO of Twitch

I would build around a group whose work you understand better than the model companies do, then make the product better as that group uses it. Kiko was our calendar for Gmail. Google shipped essentially the same calendar attached to Gmail, and our opening disappeared.

Spotify

Twitch developed differently. We talked to game streamers, solved the problems behind their feature requests, and recruited them one by one. Each streamer brought an audience; viewers discovered other streamers; that made recruiting the next streamer easier.

Apple Podcasts For an AI startup, I'd look for a similarly specific group: say, a tool that helps a particular kind of creator produce work and helps that creator reach and retain an audience. A model company can ship the generation feature. The repeated work with creators and the audience relationships take a different kind of effort.

Reid HoffmanCo-founder of LinkedIn, Inflection AI and Manas AI. Partner at Greylock

I would build around a transaction that requires trust between people, with AI making the transaction easier. Take a marketplace for specialized medical referrals: the model can help match a patient to a clinician, but the startup would still need clinician participation, verified availability, patient consent, and a way to learn whether the referral worked. Each completed referral can make the service more useful to the next patient and clinician. That is a harder product to copy than the matching prompt alone.

I would test the claimed advantage carefully. In a public conversation about AI networks, I argued that a hospital data deal alone is a weak moat because other hospital systems exist and synthetic data may narrow the gap. I also distinguished strong marketplace network effects, where participants have a real reason to gather in one place, from communications networks where people can easily use several services.

Spotify For the referral startup, the hard question is whether clinicians actually keep their availability current and patients get a better referral as participation grows. If that behavior never appears, the model company can copy much of the feature.

Ivan ZhaoFounder of Notion

Emmett's Kiko example is a good warning for the agency workflow I mentioned. A ready-made intake assistant is easy to copy. I would build a way for the people who set up that workflow to keep improving it across clients: reusable intake forms, approval rules, and review practices that they can adapt for each team.

At Notion, consultants began selling templates and helping companies set up workspaces; some made a living doing that work. We also worked on connecting people inside larger companies who were introducing Notion and needed to share best practices.

Apple Podcasts A startup serving agencies could give operations consultants a product they can install and maintain for clients. That creates a practical route into teams and a reason the workflow improves after the initial AI feature ships.

Ben SilbermannFounder and Chairman of Pinterest. Founder of How We Feel

Ivan's agency example raises another useful test: who takes responsibility when the answer affects someone's life? With How We Feel, the team includes scientists, designers, engineers, and therapists, and the product helps people understand and manage their emotions.

LinkedIn We describe its guidance as science-backed strategies for hard times.

An AI startup could pick one well-defined emotional wellbeing task, such as helping someone name an emotion and choose an appropriate exercise. The product would need expert review of the exercises, careful boundaries for sensitive situations, and a way to learn whether people found the guidance useful. At Pinterest, we held back health content until we could draw on credible institutions because the information people see can affect their lives.

Spotify That level of care is ongoing work a general assistant may have little reason to do for a narrow group.

Shishir MehrotraCEO of Superhuman (formerly Grammarly). Co-founder of Coda and former Chief Product Officer of YouTube

Imagine a sales rep writing a customer memo while an assistant pulls in the account update, product roadmap, and last support issue. I would focus on workflows like this, where the customer’s context lives across several applications. That requires trusted connections to the systems the team already uses, permission-aware context, and a place in the rep’s daily workflow. I described this as the opportunity behind combining Coda’s integrations with Grammarly’s presence across applications.

sloanreview.mit+1

For a startup, customer renewal briefs are one concrete wedge. Connect the account record, support history, and roadmap, then put the brief where the account team already prepares for a renewal. For the renewal brief, the test is whether the startup can keep trusted access across account, support, and planning systems and earn regular use from the account team.

Emmett ShearCEO of Softmax. Founder of Justin.tv and former CEO of Twitch

The creator example needs a stricter test than whether people like the AI output. In my first streamer interviews, I heard three needs: money, fame, and love. We gave streamers an ad button that paid them, discovery by game and transcoding so more people could find and watch them, and chat moderation that made the audience relationship workable. We also recruited streamers directly. Growth took off when those pieces worked together.

Spotify

For an AI product for creators, I would measure whether a creator comes back because the product helps them earn, get discovered, and have useful exchanges with their audience. Better generation alone leaves the startup exposed. If the audience follows each creator into the product and starts finding other creators there, you have a reason to believe the startup is building something a model company would need to operate, rather than simply add as a feature.

Ask a follow-up

Founders

467

Founders

4M+ insights

Founders from Pinterest, LinkedIn, Notion & 2 more

by SuperMe

Frequently Asked Questions

What are the signs of product-market fit?
A small group who would be upset to lose the product, then growth you can explain, from users who look like the market you want to win. Fit rarely arrives as one moment.
How is early traction different from product-market fit?
Traction responds to a launch, press or founder effort and fades when that stops. Fit keeps compounding through the product's own loop, and the love repeats with strangers who match the first fans.
What should you do when product-market fit isn't coming?
Make the next attempt a bounded test with a target and a deadline the cash can cover, check that the test was fair, and change course on a miss. With no promising group or runway left, return the capital.
How do you pressure-test a startup idea?
Set a behavior bar before the team gets attached, put an early version into someone's real work, and write down the ways the company could fail even if early users love it.
When should a founder drop an idea?
When people who have the problem can find the product, try the full version and still don't come back. Low signups and investor doubt alone are weak reasons to quit.
How do founders find a problem worth solving?
Start from a problem you will care about for years, find people whose behavior shows they have it too, and know why you can make progress where others see risk.
What should a first-time founder do first?
Start, choose a cofounder with care and get something in front of strangers early. The group splits on whether research or a rough product comes first.
How do you size a market before building?
Start from a behavior people already have, ask what has kept anyone from winning the market, and let growth or repeat use tell you its size.
How does a startup stand out in a crowded market?
Pick one experience people return to, start with the people whose choice brings everyone else, and check whether users come back for a second job.
How do you validate an idea with customers?
Go out to learn, ship one thing worth a stranger's time, watch who comes back on their own and then ask them why.
How do you choose a consumer products category?
Look for a need people already recognize, evidence that they act on it, and unit economics that work across the whole purchase cycle.

All starting a startup questions →

Frequently Asked Questions

What an AI startup can build that a model company won't easily copy

Founders

Frequently Asked Questions

What are the signs of product-market fit?

How is early traction different from product-market fit?

What should you do when product-market fit isn't coming?

How do you pressure-test a startup idea?

When should a founder drop an idea?

How do founders find a problem worth solving?

What should a first-time founder do first?

How do you size a market before building?

How does a startup stand out in a crowded market?

How do you validate an idea with customers?

How do you choose a consumer products category?

Your ideas, super powered.

What an AI startup can build that a model company won't easily copy

Founders

Frequently Asked Questions

What are the signs of product-market fit?

How is early traction different from product-market fit?

What should you do when product-market fit isn't coming?

How do you pressure-test a startup idea?

When should a founder drop an idea?

How do founders find a problem worth solving?

What should a first-time founder do first?

How do you size a market before building?

How does a startup stand out in a crowded market?

How do you validate an idea with customers?

How do you choose a consumer products category?

Your ideas, super powered.