# Autonomy vs. Alignment: Scaling AI Teams to Deliver Value

Grant Wright, SEEK Ltd. | MLOps Coffee Sessions | Episode 44 | 50:40
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=Gr69acrT8HE
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/autonomy-vs-alignment-scaling-ai-teams-to-deliver-value
Published: 2021-06-30
Tags: engineering-culture, product-strategy, team-adoption

## TL;DR
- Grant Wright says SEEK moved from fragmented AI efforts to cross-functional teams that own employment information problems through to an API.
- SEEK starts with the capabilities a problem needs, adds roles when constraints appear, and changes its operating model as the organization grows.
- Grant argues that teams should release a thin slice with feedback loops early, rather than spend years pursuing a perfect solution without evidence of user value.

## Summary
Grant Wright describes how SEEK built and scaled AI teams across its employment marketplace. The organization first had to teach people what AI could do and prepare its data. Data scientists then worked inside product teams, but that model became difficult to scale because product, engineering, and data responsibilities were split. SEEK moved to cross-functional teams with product managers, data scientists, and engineers. Each team owns a problem space and delivers a service through an API, with written agreements about what the service does, who uses it, and how it is maintained. Grant is open about the disruption this caused, especially for data scientists who lost some autonomy. He says the model improved delivery by giving teams the context and engineering support needed to put ideas into production. As SEEK grows, it is adding shared infrastructure and standards while deciding which work should remain flexible. Grant also stresses early feedback, experimentation, and thin slices for large projects.

## Key ideas
### SEEK brought fragmented AI work together around shared problem spaces
[02:17](https://www.youtube.com/watch?v=Gr69acrT8HE&t=137s)
Grant says SEEK had AI efforts spread across the organization, with teams working locally and duplicating work. The company wanted to solve problems once, keep localization in the data rather than the code, and give capable teams ownership of larger problems. Grant joined SEEK after working in management consulting and became involved in its data and matching strategy. The group grew from about 30 data scientists, engineers, and product managers to more than 100 people. He says the team is now organized around cross-functional problem spaces, supported by leaders in machine learning, product, and engineering.

### Cross-functional teams connect product judgment, machine learning, and engineering
[08:44](https://www.youtube.com/watch?v=Gr69acrT8HE&t=524s)
SEEK chose teams that combine product vision and strategy, data and machine learning capability, and engineering. Grant says useful AI work needs all three. The team must understand a user problem, use suitable techniques, and build something performant enough to appear properly in the product. SEEK applies this model to search, recommendations, pricing, platform performance, and natural language processing over job descriptions, skills, and experience. Grant says the approach released ideas from data scientists who previously struggled to align roadmaps and obtain engineering capacity.

### The team evolved through stages as the organization learned what AI required
[07:29](https://www.youtube.com/watch?v=Gr69acrT8HE&t=449s)
Grant describes an early stage focused on teaching SEEK what AI was and preparing data for useful work. The next stage used individual data scientists inside product teams, which delivered local value but created weak links between product, engineering, and data. Product managers sometimes prescribed a method such as a random forest, while some data scientists worked on valuable ideas without product support. Experiments with an AI search platform showed that small teams combining product strategy, engineering, and data science could test and deliver work quickly. This led to end-to-end teams accountable for services and their performance.

### Changing from autonomous specialists to shared teams created real friction
[14:56](https://www.youtube.com/watch?v=Gr69acrT8HE&t=896s)
Grant says the reorganization changed how people worked. Data scientists who had substantial autonomy joined teams with product managers and engineers, and the new model required more collaboration, written delivery commitments, and attention to maintainability. People felt they were stepping on one another's responsibilities. Leadership had to explain why the model was being adopted and hold to selected rituals while accepting that the design would not be perfect. Grant says teams became more motivated once they saw useful outcomes being delivered. He also says the organization sometimes took too long to adapt.

### SEEK designs roles around needed capabilities rather than standard titles
[17:36](https://www.youtube.com/watch?v=Gr69acrT8HE&t=1056s)
Grant rejects the idea that there is one correct blueprint for AI teams. SEEK studied different approaches, then returned to first principles and asked what the organization needed to accomplish. It started lean instead of creating every commonly discussed role. The team added capabilities when a real constraint or gap appeared. This approach let SEEK adjust its structure as it grew, while keeping principles about accountability, collaboration, and delivery. Grant says the right design depends on the organization's situation and should change as scale and priorities change.

### API boundaries and written partner agreements prevent orphaned model files
[23:51](https://www.youtube.com/watch?v=Gr69acrT8HE&t=1431s)
SEEK's AI teams provide services such as search recall and ranking to product teams through APIs. The API defines a technical boundary, while a written agreement explains what the service does, what it does not do, what good looks like, who calls it, and how it will be maintained and monitored. Grant calls product groups partners because the relationship works in both directions. This setup prevents model files from being passed around without anyone knowing they are in use. It also gives teams ownership of something visible in a product.

### Shared rituals reduce duplication while leaving teams room to move
[28:23](https://www.youtube.com/watch?v=Gr69acrT8HE&t=1703s)
Grant says SEEK uses documented product visions, service guides, shared OKRs, and a weekly team showcase. Teams present what they are building and share lessons across the organization. The showcase was initially held on Friday afternoons, but the team changed it after finding that timing ineffective. Grant admits the process is imperfect because the organization wants to share important information without blocking teams. As SEEK grows, it is deciding which data assets, standards, and platform components should become more common while preserving flexibility where it helps innovation.

### Large AI projects need a thin slice and a feedback loop
[21:15](https://www.youtube.com/watch?v=Gr69acrT8HE&t=1275s)
Grant says SEEK learned that teams without a product owner or clear user outcome can spend too long trying to build the perfect solution. For a large AI search project, the company wanted more than a model in production. It also needed feedback loops and supporting systems for improvement. Grant now favors a thin slice that can reach the market, be tested, and generate learning. When a product team cannot provide enough constraint, he says the team should add its own boundaries and possibly time-box the work. Early evidence matters because a strong hypothesis may still be wrong.

## Notable quotes
- Grant Wright: "We need to think about our situation, we kind of need to go back to first principles and say what do we want to design." (17:55)
- Grant Wright: "If we're going to make them accountable for that they need all the capability to solve that problem end to end." (27:01)
- Grant Wright: "You've got to kind of, you can't just build it and assume they'll come." (32:55)
- Grant Wright: "You've got to have a view of the world that you can break these constraints and do something really valuable with it." (37:44)

## Tools & references mentioned
- SEEK Ltd.
- Indeed
- StepStone
- LinkedIn
- Google

## Who should watch
- You are deciding whether data scientists should sit in product teams, centralized groups, or cross-functional teams.
- Your AI projects reach modeling milestones but stall when they need engineering time, product support, monitoring, or maintenance.
- You are scaling across markets and need to decide which capabilities should be shared and which should remain owned by individual teams.

## Related talks

- [Who's MLOps for Anyway?](https://mlopstalks.com/talks/whos-mlops-for-anyway) (Jonathan Rioux, EPAM Systems, 1:10:15)
- [Scaling AI in Production](https://mlopstalks.com/talks/scaling-ai-in-production) (Srivatsan Srinivasan, AIEngineering, 51:56)
- [Enterprise AI Operations: The Missing Piece](https://mlopstalks.com/talks/enterprise-ai-operations-the-missing-piece) (Rani Radhakrishnan, PwC US, 41:28)
- [Vision and Strategies for Attracting & Driving AI Talents in High Growth](https://mlopstalks.com/talks/vision-and-strategies-for-attracting-driving-ai-talents-in-high-growth) (Ashley Antonides, Two Six Technologies & Olga Beregovaya, Smartling & Shailvi Wakhlu, Shailvi Ventures LLC, 30:34)
- [Design and Development Principles for LLMOps](https://mlopstalks.com/talks/design-and-development-principles-for-llmops) (Andy McMahon, Barclays Bank, 1:10:18)