AI strategy: Think bigger, start with focus

Patrick Collison's thoughts on building a business offer a useful starting point: Which customer problems can we solve better with AI, and how do we find out whether our idea works?

Imagine a product presentation. A digital adviser answers questions, recommends products and prepares a suitable quote. The team is impressed. Then someone asks whether it accounts for special terms. Whether its recommendations match the product range. And who steps in when an answer is wrong.

These questions determine whether the demonstration can become a useful service. An AI strategy therefore needs to describe the path from a customer problem to reliable everyday use. That includes the idea, the knowledge behind it, the design and the operation.

In a conversation with Y Combinator, published on 31 July 2026, Stripe co-founder Patrick Collison argues for more ambitious ventures while emphasising the value of personal knowledge. His cache analogy describes understanding that is immediately available. It is not a measured speed comparison between humans and AI. His founding story is also revealing: Stripe launched publicly relatively late, but worked with real customers early on. Y Combinator conversation and transcript

Our takeaway for businesses: Choose an important problem, think through the whole solution and test the critical assumptions early. One continuous example can show what this means in practice.

1. A good AI strategy starts with a better customer decision

Consider a manufacturer whose products need explanation. The following example is hypothetical.

Its website contains product descriptions, technical specifications and application examples. Yet prospective customers frequently need to contact sales. Which version fits my situation? Which components can I combine? Which accessories do I actually need?

One possible task for AI would be to summarise the existing text. A more ambitious idea would be to guide customers through their choice: capture requirements, ask useful follow-up questions, explain suitable options and hand unresolved issues over to sales.

This changes the quality objective. The goal is an understandable recommendation that someone can act on. Whether that works depends, among other things, on the product knowledge available and how the conversation is designed.

Experienced product specialists can identify when two apparently suitable components do not belong together. Sales teams know the question customers regularly forget to ask. Designers can make it clear what a recommendation is based on and how users can correct their input.

This knowledge belongs in the solution. It needs maintained sources, understandable decision rules and examples of correct and incorrect answers. That makes expertise usable by the system and makes its quality possible to assess.

2. An ambitious idea needs a clearly defined first deployment

Over time, the digital adviser could connect product selection, configuration and quote preparation. A sensible first deployment could cover a single product family with clear selection rules. The scope and number of participants need to fit the task.

The team could initially work with selected customers and sales colleagues to test whether the advice makes a difference. Three assumptions would be critical:

• Customers can express their requirements clearly through the conversation.
• The available product data is sufficient for technically correct recommendations.
• The handover contains the information sales needs for the next step.

Each assumption requires different evidence. An attractive interface cannot answer the second question, just as a correct data model cannot answer the first.

This fits the core of Lean Startup: testing assumptions through use, measurement and learning. The method cannot be reduced to the smallest possible ideas or the cheapest possible implementation. Lean Startup methodology

In this example, that would mean asking people to complete real selection tasks, having specialists assess the answers and observing where users abandon the process or need support. The team should decide in advance which errors would rule out wider use and which results would justify expansion.

Even a small test can lead to a major decision. Perhaps the customer dialogue takes too much effort, while the same system works well as an advisory aid for sales. That would be a specific finding to guide further development.

3. Quality becomes visible at the handovers

Suppose the first test succeeds. The advice is now going to be available on the website. The product team needs to establish how updates to the range will reach the system. Sales must actually receive the enquiries. And users need a clear way forward when the adviser cannot give a reliable answer.

This is where design and technology come together in practice. A recommendation can be technically correct yet difficult to understand. A convincing conversation can fail if its outcome lands in the wrong inbox. A good first version can become outdated if nobody maintains the sources.

A shared review plan helps prepare for everyday use:

The underlying system integration should therefore be considered at the concept stage. This includes responsibilities, access rights and deciding which actions may happen automatically.

Such evidence also helps when choosing a partner: a defined scope of use, a clear benefit and a realistic assessment of the effort required for implementation and operation. Interest in AI alone says little about whether a business will adopt a particular solution.

4. Growth signals offer perspective; your own impact needs evidence

Stripe provides an interesting data point. According to the company, 20 per cent of startups incorporated through Stripe Atlas in 2025 processed their first customer payment within 30 days. In 2020, the figure was 8 per cent. Stripe's 2025 annual update

These platform figures show a change within the groups observed. They demonstrate neither that AI alone caused it nor that lasting profitability follows. They also cannot simply be applied to all new businesses or Germany's established small and medium-sized companies.

For our hypothetical manufacturer, other measures would matter. Do customers more often find a technically suitable option? Are there fewer avoidable follow-up questions? Do enquiries lead to more useful sales conversations? And how much effort goes into checking, maintenance and operation?

A baseline is needed before the test. Where possible, a comparable group should use the existing selection process. Differences in product range, campaigns or customer groups need to be considered in the analysis.

This also makes it possible to describe a quality improvement beyond time savings: customers understand their options better, decisions are easier to follow and sales receives better information. The test shows whether those improvements actually occur.

5. Creative Engineering connects the idea with its impact

For Davies Meyer, this is the role of Creative Engineering: developing customer understanding, creative concepts, design, technology and evaluation together.

In the example, the work begins with customers' questions. These shape an advisory logic, which is translated into an understandable dialogue. Product knowledge and connected systems make it usable. Evaluation shows which parts help and which need to be revised.

Within this process, AI can help draft conversations, produce alternative wording or develop test scenarios. This allows additional approaches to be explored. Which ones are technically sound, fit the brand and work for users needs to be tested deliberately.

We take responsibility for concept and quality. In our work, that means defining criteria, explaining decisions and assessing the implementation against them. Responsibilities for product data, approvals and operation are agreed with each client.

The first step can be captured in a sentence: “We want these customers to be able to complete this task better, and this is how we will know.” When the team can make that sentence specific, the project has a clear direction.

Which customer problem would you like to approach differently? Bring the task and the current process. Let's outline a useful first deployment together.