Key takeaways
- The gap between model capability and everyday adoption creates opportunities for startups.
- Context, integration and trust are essential to turning AI capabilities into useful products.
- Hands-on implementation can build an advantage if the learning becomes repeatable software.
- An adoption gap is not a moat: labs, incumbents and better models can close it.
- Evaluate AI products by what customers reliably accomplish, including the cost of checking and correcting outputs.
In May, OpenAI announced a company with more than $4 billion of initial investment dedicated to helping businesses deploy AI. Alongside it came an agreement to acquire Tomoro, bringing approximately 150 forward deployed engineers and deployment specialists into the operation. Investors included McKinsey, Bain & Company and Capgemini.
The OpenAI Deployment Company is a revealing signal. One of the companies building the world’s most capable models is investing heavily in people who will sit inside customer organisations, connect systems and redesign workflows. Better models alone are not enough to get the work done.
An impressive demo takes seconds. Changing how a business operates can take years. Between those two timelines sits an enormous opportunity for startups.
Cars, not faster horses
In a recent briefing to the Boston College investment Committee, Sequoia’s Pat Grady shared his view on the state of AI. His analogy is the arrival of the car: a new capability that changes what becomes possible. His briefing prompted me to revisit what that means for founders and investors.
The implication goes beyond making existing software easier to use. AI creates opportunities to take on more of the work itself. A product might help prepare a contract, investigate a customer problem or produce campaign assets, changing both what the customer buys and how the service is delivered.
But the arrival of a capability and its widespread adoption are different events. A model can draft a contract, generate a video or analyse a spreadsheet. Turning that into something people depend on requires context, access to the right systems, a way to check the output and enough confidence to use it again tomorrow.
Grady describes this distance as the diffusion gap.
In my earlier post, The Chokepoint Thesis: Moats, Affordance and Diffusion, I explored Sequoia’s framework for this moment. The diffusion argument deserves a closer look because it changes where founders should search for opportunities and how investors should assess them.
The question becomes: what can customers accomplish with the technology already available, and what is still stopping them?
Why the gap keeps widening
The challenge for anyone adopting AI is that the technology does not wait for the implementation plan. Each improvement makes another set of tasks possible while organisations are still working out how to adopt the previous wave. The gap can be replenished even as individual customers make progress.
There is an understandable temptation to assume adoption follows automatically once the technology is good enough. But businesses run on permissions, budgets, relationships, habits and processes that have often accumulated over decades. A more intelligent model does not automatically get access to the right customer records, know who can approve a refund or persuade a department head to change how their team works.
Headline adoption numbers can obscure this distinction. In McKinsey’s 2025 global survey, 88% of respondents said their organisations regularly used AI in at least one business function, but only 39% reported an impact on earnings before interest and taxes at the enterprise level. That does not mean the remaining organisations received no value. It shows how different widespread use and measurable business impact can be.
In a recent conversation with Sequoia, Box CEO Aaron Levie explains why this creates substantial work for application companies. Even a much more intelligent model needs connections to business systems, human involvement at particular stages and a way to operate within existing processes. Someone has to understand how the customer’s business actually works, and much of the opportunity sits in those details.
Where startups win
For a startup, the useful starting point is a specific task that matters enough for a customer to change their behaviour. The product has to make the result dependable and remove enough friction that using it becomes the obvious choice.
Legal AI offers a useful illustration. Suggesting a plausible contract clause is one thing. Knowing whether it reflects a client’s negotiating position, the firm’s preferred language and the relevant precedents is a much more specific problem.
A&O Shearman’s ContractMatrix, developed with Harvey and Microsoft, draws on organisations’ precedents and policies and lets lawyers review proposed amendments through tracked changes. Harvey’s customer case study reports around 2,000 lawyers using it daily. The product places AI within a familiar process and retains professional judgement at the point where it matters.
My takeaway is that context and control are central to the product’s value. A better model can improve the suggestions, while the surrounding system makes those suggestions usable. The customer needs both.
The same pattern appears in creative automation, an area we spend a lot of time thinking about at Remagine Ventures. Generating an attractive image is one step in producing a campaign. A marketing team also needs accurate product details, consistent branding, different formats, localisation, approvals and a route to publication.
Adobe’s GenStudio announcements reflect that broader need: reusable production workflows, connections to asset libraries and review tools, and brand checks alongside generative capabilities. Established companies clearly see the opportunity too. A startup needs to identify a customer or production process it can serve exceptionally well.
For a creative automation business, I would want to know how quickly a retailer gets approved assets into market, how much manual work remains and whether the customer returns every week. The number of images generated tells me relatively little about whether the company has solved the customer’s problem.
The customer experiences the checking, the handoffs and the exceptions as part of the job. If a product generates something instantly but leaves a person with an hour of corrections and coordination, the demo and the value proposition are telling different stories.
An opportunity is not automatically a moat
This is why I am increasingly interested in what founders learn during implementation. Spending time inside a customer’s business can reveal requirements that never appear in a benchmark: the spreadsheet everyone quietly relies on, the approval nobody documented, the exception that happens every Friday.
That work can be a source of insight and differentiation. It can also become a trap. Some adoption gaps will disappear with the next model release. Others will be closed by an incumbent adding a feature to software the customer already uses. Some require so much bespoke work that every new customer becomes another consulting engagement.
OpenAI’s deployment investment captures this tension. It validates the importance of implementation while bringing a powerful supplier closer to the customer. Founders should take both implications seriously.
The venture question is whether solving the adoption problem creates something that becomes more valuable and more repeatable over time. Does deployment become easier with each customer? Can the company learn from corrections and outcomes, where it has permission to do so? Does it become embedded in a recurring process? Is there a credible route from an initially narrow task to a larger customer relationship?
Early hands-on work can be exactly what a startup needs. But if customer twenty requires as much custom work as customer one, I would want to understand why. The learning needs to become part of the product.
The opportunity and blind spot for Israeli founders
A survey by the Israel Innovation Authority and the Brookdale Institute found that 95% of surveyed Israeli tech workers used AI tools regularly and 78% used them daily. That familiarity is an advantage: it helps founders recognise new possibilities and build quickly.
It can also create a blind spot. When your team lives inside coding agents and experiments with every new model, it is easy to underestimate how much support another user needs. An operations manager at a regional insurer or a controller at a mid-sized manufacturer may have a very different appetite for experimentation.
The diffusion gap is often defined by people whose working day looks nothing like the founder’s. They may care deeply about the result while having little interest in learning a new interface, writing better prompts or choosing between models.
For Israeli teams selling internationally, time spent with those customers is part of product development. Understanding their constraints, earning their trust and making the first deployment work deserve attention as early as the underlying technology.
What we look for at pre-seed
At Remagine Ventures, this pushes me towards founders who understand a buyer’s working day in unusual detail. I want to see a specific task that matters, a clear way to measure improvement and a product that becomes more useful as models advance. I also want to understand the full cost of delivering the result, including human review and customer support.
The same logic extends to consumer AI. There, the adoption gap often sits between trying something once and making it a habit. A creative tool, tutor or shopping assistant has to earn a place in someone’s life. Capability gets people interested; repeated usefulness gives them a reason to return.
You do not have to wait for another breakthrough to build something valuable. There are already tasks that AI can help people complete but which remain difficult to put into practice. Finding those tasks requires curiosity about the customer and a willingness to work through the details that a demo leaves out.
As an investor, the question I keep coming back to is: what can your customer reliably get done with your product that they could not get done before? The next model release will create another wave of impressive demos. I’m interested in the founders who turn that progress into something people depend on.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- The AI opportunity hiding between capability and adoption - October 13, 2026
- Weekly Firgun Newsletter – October 9th 2026 - October 9, 2026
- How Israeli tech changed since October 7 - October 8, 2026
