AI Paved Paths
Encode shared building blocks and guardrails so humans and agents can move safely
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 91%
AI Paved Paths is a platform model for scaling work when more people and agents can build across multiple systems. A central team identifies the capabilities that recur across local problems, then packages them as preferred infrastructure, templates, source-of-truth data, and encoded guardrails. These foundations should carry teams roughly 80 percent of the way without forcing every builder to rediscover security, identity, testing, data interpretation, or design standards. Local teams still own the business problem and can extend the path where necessary. The mechanism increases velocity by removing repeated work while reducing the risk of fragmented systems, inconsistent experiences, and unreviewed AI output. Human judgment remains responsible for whether the result is useful, high quality, and safe.
Origin
Elizabeth Stone described Netflix strengthening common infrastructure as humans and agents work across more systems. She used engineering, data science, and design examples to show how paved paths and templates can encode what good looks like without eliminating local ownership.
Core principles
- 01Shared platforms should get most teams most of the way to a solution
- 02Guardrails work best when encoded in tools and workflows
- 03Source-of-truth data must replace tribal knowledge
- 04Humans remain accountable for impact and quality
- 05Common foundations should preserve room for local problem solving
How to run it
- 1
Map repeated building blocks
Look across business domains for capabilities, data, and decisions that teams repeatedly recreate. Separate genuinely shared foundations from domain-specific work.
Pro tip Start with the recurring work that can get most teams roughly 80 percent of the way there.
Watch out Do not centralize a capability merely because two teams happen to use similar tools.
- 2
Establish trusted foundations
Define the source-of-truth data, common infrastructure, and preferred access patterns builders should use. Make the correct starting point easy to find.
Pro tip Treat old tribal knowledge as a signal that an important rule should be made explicit.
- 3
Encode the guardrails
Build security, identity, testing, data-use, and quality expectations into the paved path. Reduce the need for each person or agent to remember every rule independently.
Pro tip Encode design language and analytical interpretation as well as engineering controls.
Watch out A written policy alone will not scale when thousands of builders and agents are acting quickly.
- 4
Preserve accountable judgment
Make a named human responsible for the problem, the quality of the output, and the decision to ship. Add review where AI output cannot yet be trusted on its own.
Watch out Agent-generated work does not remove human responsibility for the result.
- 5
Enable local extension
Let product teams adapt the shared foundation to their specific business problem. Feed broadly useful improvements back into the platform for other teams.
Pro tip Judge the platform by how quickly it enables good local outcomes, not by adoption alone.
Watch out A mandatory one-size-fits-all stack can recreate the friction the platform was meant to remove.
In the wild
Netflix is hiring engineers who can look across business domains and identify common building blocks for a world where agents operate across multiple systems. The shared path can supply source-of-truth data and guardrails while domain teams remain focused on personalization, advertising, content delivery, and other local problems.
→ Teams can move faster without rebuilding core capabilities or relying on one expert's tribal knowledge.
Netflix's experienced designers develop templates and express what good user design looks like so people without formal design training can still build coherent products. The design system protects the end-to-end member experience as more functions gain the ability to prototype and ship.
→ More builders can contribute without producing conflicting design languages or a fragmented experience.
Common mistakes
Relying on tribal knowledge
Expecting every builder to find the one person who knows the rule does not scale when thousands of people and agents are working across systems.
Confusing a platform with control
Forcing every local problem into one rigid stack can suppress the speed and judgment the shared foundation is meant to enable.
Automating away accountability
Treating agent output as ownerless work leaves no human responsible for its impact, quality, or safety.
Is it for you?
Best for
Organizations scaling AI-assisted work across many teams, systems, and business domains.
Not ideal for
Small teams with one narrow product and little repeated infrastructure or coordination cost.
From the episode
Netflix CPTO on AI and the future of product and tech roles
Elizabeth Stone