# Structuring a Modern AI Team

Denys Linkov, Wisedocs | AI Engineer World's Fair 2025 | 17:40

Source: https://www.youtube.com/watch?v=SbUxRluVRwk
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/structuring-a-modern-ai-team
Published: 2025-07-24
Tags: engineering-culture, enterprise, team-adoption

## TL;DR
- An AI team's structure should follow the company's bottleneck, whether that is product delivery, customer acquisition, retention, monetization, scalability, or reliability.
- Generalists fit early AI work because they can move across model training, serving, product requirements, domain work, and customer conversations; specialists become useful after the team exhausts its general knowledge.
- Hiring should add people who can hold context or act on it, while existing staff should learn to build prototypes, define domain use cases, write evaluations, and work directly with customers.

## Summary
Denys Linkov argues that companies should design AI teams around the work required to create value, rather than automatically hiring AI researchers. Technology is often available before an organization can adopt it, so domain knowledge, product judgment, customer contact, and execution matter. He divides companies into technology companies, verticalized or services companies, and tech-enabled companies, each with different gaps. Teams should identify their bottleneck and balance skills in model training, model serving, and business acumen against their budget. Generalists are useful while a company is still finding product fit. Later, specialists can push performance further. Existing teams can be upskilled through prototyping, evaluations, domain work, customer calls, and weekly learning sessions. New hires should provide context or act on it, with humans remaining accountable for the systems they build. Linkov also advises companies to test hiring assumptions from first principles and ask job-relevant interview questions.

## Key ideas
### The right AI team depends on the kind of company and its bottleneck
[01:10](https://www.youtube.com/watch?v=SbUxRluVRwk&t=70s)
Linkov separates organizations into technology companies, verticalized or services companies, and tech-enabled companies. A technology company may lack domain knowledge and business alignment. A verticalized company may have either strong or weak access to technology and expertise. A tech-enabled company often needs help with its technical challenges because technology is not its main product. Teams should then identify what is actually blocking progress: shipping features, acquiring or retaining users, monetizing, scaling, or maintaining reliability and observability.

### Technology adoption is often slower than technology development
[02:40](https://www.youtube.com/watch?v=SbUxRluVRwk&t=160s)
Linkov says that he believes society already has 90% of the technology needed to solve humanity's problems. He points to fax spending, the continued use of checks, and the slow adoption of contactless payments as examples. Medical systems and electronic medical records took decades to become digital after personal computing and the internet appeared. His conclusion is that technology itself is often not the limiting factor. The way a company uses technology should shape how it builds its team.

### Most companies do not need an AI researcher as their first AI hire
[03:45](https://www.youtube.com/watch?v=SbUxRluVRwk&t=225s)
Hiring an AI researcher can be premature before a company reaches a sufficient scale or has a specialized need. Training models from scratch, and even fine-tuning models of a certain capacity, may come after transformation work that connects AI to products and operations. Model providers such as OpenAI, Anthropic, and Google are different because advanced models are their product. Linkov asks teams to consider whether they would trade their existing domain team for five researchers from top labs, plus cash and draft picks. For many companies, the answer should be no.

### An AI team has to connect models to products and customers
[04:54](https://www.youtube.com/watch?v=SbUxRluVRwk&t=294s)
Linkov lists the work involved in creating value with AI: defining use cases, integrating with products, measuring return on investment, finding data, testing and refining workflows, building interfaces, selling the product, and making customers care. One person cannot normally do all of this. An AI team therefore needs cross-functional coverage, and the company should avoid treating the AI group as an isolated owner of a small deployment. Shipping the org chart into the product can create strange product behavior.

### Early machine learning teams can trade depth for breadth within a budget
[06:48](https://www.youtube.com/watch?v=SbUxRluVRwk&t=408s)
When Linkov built a machine learning team in 2021, he hired generalists and supported them with automation. The team had goals around serving hundreds of thousands of concurrent models, supporting multiple domains, keeping costs low, and enabling real-time training and serving. It built a custom MLOps platform, fine-tuned encoder models, created RAG as a service, and owned six microservices. Linkov prioritized model training, model serving, and business acumen, but set different skill thresholds rather than demanding the highest level in every area.

### Open source progress changes which skills a team needs
[09:09](https://www.youtube.com/watch?v=SbUxRluVRwk&t=549s)
For a team he built in 2024, Linkov found that open source and commercial models had advanced. The team could use commercial APIs, prompt tuning, and commercial-model fine-tuning, while also working with encoder and decoder models. An open source serving offering removed the need to build the team's own platform. The domain bar increased because the company processed medical records. Skills can be distributed across several people, so the team only needs a workable combination rather than one person who covers every specialty.

### Existing teams need inner-loop execution and outer-loop expertise
[10:45](https://www.youtube.com/watch?v=SbUxRluVRwk&t=645s)
Linkov uses inner and outer loops to think about reskilling and team structure. The inner loop contains daily work such as model training, prompting, product requirements, and model serving. The outer loop contains expertise that helps differentiate the company, including domain knowledge and the ability to build business cases. A weak technical inner loop harms execution. A weak domain loop makes product-market fit harder to find. The exact mix changes by company, but the framework helps teams decide where their priorities lie.

### Generalists fit discovery, while specialists fit later performance work
[11:53](https://www.youtube.com/watch?v=SbUxRluVRwk&t=713s)
Linkov says companies usually benefit from generalists while they are trying to find product fit and make basic progress. Generalists can adapt and do work beyond writing code. Once the team has established model training and serving and has exhausted its general knowledge, specialists can pursue the extra performance gains. He applies the same progression to the wider AI transformation: people need to learn to build, become domain experts, and work directly with other humans.

### Hiring should add context and accountability, not follow fashion
[13:56](https://www.youtube.com/watch?v=SbUxRluVRwk&t=836s)
Linkov gives two reasons to hire: a team may need someone to hold context, or someone to act on that context. AI agents may handle some work with a large context window, but people with expertise are still needed to check whether the context and execution are correct. Humans should remain accountable for the systems they build. He also rejects blanket claims that junior engineers are no longer useful. Companies should decide from first principles whether they need new graduates, experienced people, or retraining, then evaluate candidates with questions related to the actual job.

## Notable quotes
- "The way you build your team should reflect this by understanding the problems that you have." (03:45)
- "You win with generalists at the beginning when you're trying to find that fit, trying to make that basic progress until you get to the point where you exhaust the knowledge and you need to move into a more specialist model." (12:10)
- "People need to be hired for two main reasons. One is to hold context and the other is to act on context." (13:56)
- "Stop putting people through LeetCode that have nothing to do with the job." (16:00)

## Tools & references mentioned
- Wisedocs
- Shopify
- Duolingo
- Zapier
- Palantir
- OpenAI
- Anthropic
- Google
- Pascal's wager
- Ampere's wager
- Blackwell's wager
- Hugging Face
- Kubernetes
- RAG
- MLOps
- LeetCode
- YC

## Who should watch
- You have an AI mandate but cannot add unlimited headcount, and need to decide whether to retrain existing staff or hire specialists.
- Your team is building AI features but lacks a clear mix of domain expertise, engineering, model work, product judgment, and customer contact.
- You are setting interview criteria for AI roles and want to replace trend-following with job-relevant evaluation.

## Related talks

- [Hiring & Building an AI Engineering Team](https://aietalks.com/talks/hiring-building-an-ai-engineering-team) (Dr. Bryan Bischof, Hex, 29:07)
- [Small AI Teams with Huge Impact](https://aietalks.com/talks/small-ai-teams-with-huge-impact) (Vikas Paruchuri, Datalab, 17:36)
- [Insights on Building AI Teams](https://aietalks.com/talks/insights-on-building-ai-teams) (Heath Black, SignalFire, 20:30)
- [Moving away from Agile: What's Next](https://aietalks.com/talks/moving-away-from-agile-whats-next) (Martin Harrysson & Natasha Maniar, McKinsey & Company, 21:55)
- [AI Platform Engineering](https://aietalks.com/talks/ai-platform-engineering) (Patrick Debois, 28:18)