AI Engineer World's Fair 2026
How to build an AI-Native Health Company
Loading slides…
Read the talk
Building an AI-Native Health Company
Maven Clinic’s transition shows how AI changes delegation, planning, code review, and the reliability controls needed for scheduling and reimbursement.
From a talk by Dan Feng
Where does AI adoption begin?
Is your company already AI-native, or are you still thinking about starting? In Dan Feng’s opening audience poll, a few hands go up for each position. Most of the room falls somewhere between them: adoption is underway, but the transition is incomplete.
Maven Clinic provides digital healthcare for women and families across maternity, fertility, parenting, and menopause. Feng dates the start of Maven’s AI journey to roughly two years before the talk. Its work has produced Maven Intelligence, an orchestration layer that enables AI across the company’s products for employees and clients.
The motivation is competitive as well as technical: choosing not to adopt AI does not stop competitors from adopting it. Feng illustrates this with a tractor analogy. The advantage belongs to farmers who learn to operate the machinery; possessing a new technology matters because it changes what its users can accomplish.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Change tools, products, and working practices
There is no single definition of an AI-native company and no predefined playbook that completes the transition. Maven’s approach gives that broad ambition three practical dimensions.
- Internal tools. Start with work people already do: daily summaries, meeting management, and Jira task creation. Before doing something manually or delegating it, ask, “Can I use AI to do it?” Maven’s leaders, including its CEO, increasingly use AI to complete routine tasks themselves. Adoption changes who performs the work as well as which tool they use.
- Customer products. Build AI into the product to improve the user experience and potentially lower operating costs. Feng’s example is always-available chat support, which he presents as a faster, cheaper way to help customers than human support.
- Culture and process. Change the way people work so that the organization can benefit from the new capabilities. Tool access and product features are only part of the transition; working practices need attention too.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Support people at different adoption speeds
A company does not adopt new technology at one uniform speed. Feng distinguishes three groups, each needing a different kind of support:
- Early adopters. Give them access to tools and encourage them to share what they discover. They need little persuasion; their experience can help the rest of the company.
- The majority. Put the most effort here. Build shared infrastructure and easy-to-use tools, remove adoption friction, and improve the experience through continuous feedback. Tool preferences can change: Feng describes widespread use of Cursor in the previous year, followed by many employees switching to Claude Code in the year of the talk. Maven supports both so that people can work with tools they are comfortable using.
- Hesitant adopters. Understand their worries and meet them where they are. At the same time, clearly communicate the company’s direction so that support does not leave expectations ambiguous.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Move ownership closer to the person solving the problem
When the desired outcome is clear, AI can take on much of the implementation. That changes the value of a handoff. In Maven’s earlier workflow, a senior engineer assessed a problem, designed a solution, and delegated implementation so several engineers could work in parallel. Feng now sees engineers increasingly implement their own solutions with AI. Once they have worked out what to build, delegating it introduces coordination overhead that they can often avoid.
That shift raises the expected scope of individual ownership. Maven wants new engineers to solve problems independently, taking on responsibilities previously associated with a technical lead. The expectation is that they can carry a problem through to implementation without depending on another engineer to define and delegate each task.
The hiring criteria follow from that expectation:
- Interest in AI and continued learning. The field changes quickly, so engineers need to keep learning and help their teammates stay current.
- Product understanding. As AI expands what an engineer can implement, the boundary between product management and engineering becomes less distinct. Understanding the product helps an engineer contribute beyond the software mechanics.
- System knowledge and judgment. Deep understanding of the system and the ability to handle complicated, ambiguous problems remain valuable.
Rewards need to reinforce the same behavior. Maven has begun asking about AI contributions in performance reviews. The emphasis is on using AI to multiply a person’s impact, with that expectation extending beyond engineering to the rest of the company.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep the vision long and the delivery commitment short
Extensive upfront planning made sense when implementation was expensive to change. Teams could spend weeks or months refining requirements and finalizing designs before building. Yet large projects still failed to get everything right at the beginning. Feng describes some AI-assisted building as taking only minutes. In that setting, alignment becomes the expensive part, and the planning process needs to reflect the changed cost of implementation and correction.
A one-year vision still has a role: it provides inspiration and direction. Feng suggests a planning exercise in which the team assumes future models can do everything it would like, then imagines what the product could become. That assumption is a way to expand ambition, not a forecast of model capabilities or a detailed delivery commitment.
The concrete work stays close to the present:
- Define the next delivery window. Focus on what the team will deliver in the next two to four weeks. Product managers and designers specify what engineers need for the current sprint.
- Build and release. Engineers focus on that scope and release by the end of the sprint, or sooner. Meanwhile, product managers prepare the next batch of requirements.
- Allow correction. If a decision made at the start of the sprint proves wrong, change course and fix it. The process should accommodate inexpensive revision.
- Keep documents useful for communication. Maven prefers one- or two-page product requirements documents and technical design documents over lengthy specifications, so people can discuss and iterate on them.
The awkward horizon is the middle: three- to six-month plans. Several model releases may arrive during that period and change what is feasible. Feng therefore puts less emphasis on those plans, even though quarterly and half-year planning are familiar habits. The organizational adjustment is learning to work with a directional long-term ambition and specific near-term commitments while accepting uncertainty between them.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Expand coding assistance from tasks people can verify
AI coding tools offer a practical starting point because implementation is one of their strengths. But adopting them does not require immediately automating every stage of development or becoming comfortable building software from a phone. Maven expanded its use through experience with lower-risk tasks.
The progression was gradual:
- Start with unit tests and documentation. These tasks offered outputs that engineers could readily verify and a low-risk way to learn the tools.
- Build confidence and guardrails together. Use that experience to develop the team’s rules, skills, and guardrails.
- Make AI use the expectation across engineering tasks. When an engineer chooses not to use it, investigate why. The exception becomes useful feedback about where the tools or workflow are not working.
- Shift engineering attention. Feng reports that AI coding tools now handle nearly all of Maven’s implementation, while engineers concentrate on review, architecture, and evaluation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep review meaningful as code output grows
More implementation capacity creates a review bottleneck. Feng’s rough observation is that engineers who once wrote hundreds of lines of code per day can now produce thousands with AI. A review process designed around the earlier volume cannot simply absorb that increase unchanged.
Maven has tried several AI code review tools. They help, but the team is not comfortable relying on them completely; engineers still provide valuable feedback. The immediate response is to change how human review is used.
- Let engineers identify simple changes. An engineer who understands a pull request and feels confident that it does not need another reviewer may merge it without peer review. The engineer remains accountable for the change.
- Keep reviewed changes small. Feng recommends no more than 500 lines of code per PR requiring review. This is Maven’s stated review practice, not a universal threshold for safety; the purpose is to make meaningful inspection possible.
- Use stacked PRs for larger features. Divide the work into multiple pull requests so an engineer can continue implementing while earlier changes are being reviewed.
The failure mode to avoid is rubber-stamp approval: a reviewer cannot meaningfully inspect the change but approves it anyway. That creates false confidence because the team treats the approval as evidence that the code was checked. Maven continues working on AI review tools, but their promise does not make an empty human approval useful.
At the time of the talk, AI assists with nearly every stage of Maven’s software development. End-to-end automation—from design through implementation and release—is the goal. Feng extends that ambition to monitoring live traffic, detecting issues early, and fixing them automatically. He explicitly says that this autonomous operating loop is still being developed, not an achieved capability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choose reliability controls by failure consequence
Generative systems introduce uncertainty beyond the behavior explicitly implemented in conventional software. Hallucinations cannot be ignored, but Feng argues that trying to eliminate every one can be costly and may be unnecessary for some tasks. Define acceptable and unacceptable failures at the beginning of system design, then choose controls that fit their consequences.
| Workflow | Example failure | Consequence and tolerance |
|---|---|---|
| Appointment scheduling | An appointment attempt fails, and the user can press the button again. | Feng suggests one failed attempt in 1,000 could be tolerable when the user can retry; this is a hypothetical tolerance, not a measured failure rate. |
| Reimbursement | A $200 claim is paid as $50, or a $50 claim is paid as $200. | Either incorrect amount is unacceptable in Feng’s example and triggers escalation. |
The scheduling example still represents a poor experience, but it has a straightforward recovery path. An incorrect payment needs stronger controls because both underpayment and overpayment create a consequential error.
For receipt processing, Maven adds a decision process before proceeding:
- Review the same receipt with different models. Obtain multiple assessments of the same input.
- Require agreement. Move forward only when the models’ results agree.
- Offer human assistance when the result remains unclear. Explain that the system is having difficulty processing the receipt and ask whether the customer wants to connect with a human agent.
The workflow therefore has a way to stop and seek help instead of forcing an uncertain result through. Rigorous release testing provides another layer of reliability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Test repeated behavior before release
Maven has hundreds of integration tests covering known use cases, and the team keeps adding cases. Each test runs many times because an LLM can behave differently across executions. A single successful run is insufficient evidence that the behavior is dependable.
For repeated integration-test runs, Feng gives 90% as an example of the consistently high pass rate the team requires. That figure describes an example testing requirement, not measured production reliability or permission to issue incorrect reimbursements. The testing process asks whether the system can repeatedly handle a case, rather than whether it happened to succeed once.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluate conversations and recalibrate after launch
Evaluation continues after release. Maven’s automated evaluation system assesses each conversation using predefined rubrics for good and bad behavior. It generates results that the team reviews, making production conversations part of the quality process rather than ending evaluation at the release gate.
A dedicated group also manually spot-checks conversations. Human review can reveal a problem in the system, but it can also reveal a problem in the evaluation itself: a rubric may be too strict or too loose. The feedback loop therefore improves both the product’s behavior and the criteria used to judge it.
For new-feature launches, Feng suggests manually reviewing about 20% of conversations instead of relying on ordinary spot checks. This is increased scrutiny during a launch, not the baseline sampling rate. The combined process gives the team confidence to release while still acknowledging that hallucinations remain possible.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
First-party announcement of the orchestration layer Feng names, with rollout timing and governance context. The announcement itself is not cited in the talk.
Current official documentation for one of Maven's supported coding tools; useful for exploration, not evidence of the reported internal results.
Current official introduction to the other coding tool Feng names, including setup and example workflows.
Full transcript
Loading transcript…