Interviews · Interview
Charles Lamanna on What Microsoft Hires For in the Age of AI
Key takeaways
- Every employee becomes a manager of AI agents. At Microsoft, people create Autopilot agents with their own identity, personality and memory, and delegate work to them.
- Accountability stays with the person. In Lamanna's words, if your Autopilot makes a mistake, the buck stops with you, so you set the policies for what it may do.
- Everyone should be "model pilled and context pilled". Everyone has access to the same models, so the only way to stand out is the context you give them.
- Context engineering means choosing the right 50,000 tokens out of hundreds of terabytes. Microsoft calls its capability for this Work IQ.
- Databricks Genie and Microsoft Copilot work together through MCP so an agent can use a company's full context, from structured to unstructured data.
- Context is hard for three reasons: data that was never prepared, permissions, and APIs that were built for people instead of AI agents.
- The first thing to do is prepare your data: define golden records, build semantic models, delete stale data and secure it. That work pays off with every future model.
- Lamanna hires for two traits: being a learner with a growth mindset, and a thirst for doing a great job.
- Reading and writing individual lines of code is behind us, like assembly. Architecture, systems thinking, resiliency and business requirements still matter.
About this video
Every employee at Microsoft will get a direct report, and that direct report is an AI agent. In this conversation, Charles Lamanna, Executive Vice President of Copilot, Agents and Platform at Microsoft, explains what it means when we all become managers of AI agents, who is accountable when an agent makes a mistake, and why context, not the model, is now what sets a company apart.
We talk about context engineering in practice, how Microsoft Work IQ picks the right 50,000 tokens out of hundreds of terabytes, and how Databricks Genie and Microsoft Copilot work together through MCP so structured and unstructured data reach the same agent. Charles also shares what joint customers should do first, the two traits he hires for, and why reading and writing individual lines of code is giving way to architecture and systems thinking.
In this conversation, Charles explains:
- Your AI direct report: how Autopilot agents get their own identity, memory and policies, and why the buck still stops with you
- Model-pilled and context-pilled: everyone has access to the same models, so the context you give them is the differentiator
- Context engineering: why a 1 million token context window forces you to choose what goes in
- Genie and Copilot together: one full view of your data, from structured to unstructured
- Why context is hard: data preparation, permissions, and APIs built for people instead of AI agents
- The number 1 skill: a learner's growth mindset and a high bar for doing great work
- The future of coding: why syntax matters less, and architecture, resiliency and failover still matter
Chapters
- 00:00Intro
- 00:28Every Microsoft Employee Gets an AI Direct Report
- 01:18Who Is Accountable When an AI Agent Makes a Mistake
- 02:15Model Fatigue and Why Context Is the Differentiator
- 03:44Context Engineering, 1 Million Tokens and Microsoft Work IQ
- 04:58Databricks Genie and Microsoft Copilot Through MCP
- 06:01Why Context Is Hard: Data, Permissions and APIs for Agents
- 07:32What Joint Customers Should Do First: Prepare Your Data
- 08:23The Number 1 Skill Charles Lamanna Hires For
- 10:01One Project Everyone Should Build with AI Coding Tools
- 10:47Is Learning to Code Still Important in the Age of AI
Links and resources
Transcript
Generated from the video's captions and lightly edited.
Do you think it's still critical to have the coding skills?
Being able to read and write individual lines of code, I think that era is behind us. It's like it's like being able to read and write assembly 10 years ago. Like everybody should be model pilled and context pilled. They can send the same prompt and get the same answer. Why is your use of the model better than somebody else's? Because of the context you give to it.
What's the number one skill that right now is very critical?
The first is
At at the keynote you said something that really caught my attention. Every employee at Microsoft will get a direct report and then you shared your autopilot Apollo. I would love to learn more about this.
Yeah, everybody wishes they had like an assistant which could maybe help them with scheduling or moving projects forward or you know, create the presentation for the customer meeting tomorrow. In the past like that just wasn't accessible cuz that's not kind of how the world works. But now with autopilot people can kind of mint these things, give them their own identity, their own little like avatar, their own little personality and memories and then they help them be more efficient and more productive. And we think the the beauty of this is we're all going to be managers of these agents over time.
And that's increasingly what work will look like where you're maybe not doing it yourself by hand but instead delegating out to these things like an autopilot.
That's really exciting. So we'll have direct reports each of us. What if autopilot makes a mistake? Who will be responsible? Agent or me?
I would say it's going to be you cuz it's kind of like a it's like a classic organizational theory which is if someone on your team makes a mistake the buck stops with you. So if your autopilot makes a mistake the buck will stop with you as well and you'll ultimately be responsible and accountable. People aren't perfect either. Like we have we have systems, we have social contracts, we have risk controls and protections which make it possible for imperfect people to do things in a complex world. Same thing will help us have autopilots go do tasks as well.
You know, if like my autopilot is not going to go submit like a invoice for $10,000 without me looking at it. But you know, if it wants to do something for five maybe I let it.
You basically will define the policies around what autopilot is allowed to do.
Exactly. My autopilot, I don't let it talk to people outside of Microsoft because I don't want like emailing customers directly and I don't want customers emailing it. So it only is able to work like in company boundaries or with pre-approved people.
We used to be so excited about new models. Whenever the new models comes out, oh my god, we have to all switch. Here are all the benchmarks. But lately the shelf life of commercial model has been really decreasing, right? People already have fatigue of new models. From the other side, they are actually not driving the marginal improvement in businesses. And what we are seeing now that this really big improvements will come from context. Why do you think it's so hard to provide the right context to a model?
I mean I I would say right now everybody should be model pilled and context pilled. So the first is everyone will have access to the same model. So if you think about what it means to be a company, it's all about differentiating, providing more value, providing unique value. Anything that comes from a model anyone else can do. They can send the same prompt and get the same answer. Why is your use of the model better than somebody else's? Because of the context you give to it. So the only way to really differentiate and stand out moving forward is going to be by providing the right information in the right way at the right time to the model to get the outcomes you want.
And if you think about what makes a company unique, like what make like what is the difference between Boeing and Airbus or what's the difference between Pepsi and Coca-Cola? It's in their IP. It's in the intellectual knowledge and what's inside their employees' heads and how they work. And that is what defines a company. That's what makes them interesting and compelling to their customers. And if you want to make AI be a reflection of how your company operates or your team operates or you operate, you're going to have to give it context so it behaves appropriately.
The models are very smart, but the context window doesn't keep growing, right? Million tokens is more or less we're stuck around that. So if the model can only look at a million tokens to make a decision, figuring out what to put in those million tokens is really important. That's context. That's context engineering. And like I'll give an example of own use at Microsoft. So I have a OneDrive where I store all my files. it's about 700 gigabytes of content, presentations, videos, huge amount of content. I have a email inbox, which is about 90 gigabytes.
And I have in terms of documents and SharePoint and other files shared with me in the order of hundreds and hundreds of terabytes. So if I'm trying to use like Copilot and I want it to be able to answer a question that I really care about, none of the information that matters is in the model. It's in that context. Well, I can't feed hundreds of terabytes of data into the model. It'd cost a fortune and take forever. It just wouldn't work. I need to have some system to figure out what's the right million tokens in reality, what's the right 50,000 tokens based on what I'm trying to do to put into the model input.
That's what we call Work IQ at Microsoft. That's the con like the special context capability that we have in Copilot. and you're really not going to get a ton of value from AI if you don't have an approach to context.
Databricks we have similar technology with Genie. We are we're utilizing ontology and really looking into all of those different assets that you have. And now with our partnership we can bring Genie also to Copilot utilizing MCP. What are you most excited about the Databricks and Microsoft partnership?
Yeah. I mean I think the the first thing is everything we do at Microsoft and the same thing at Databricks is optimizing on what's best for our customers. So we have lots of customers who use Microsoft 365 and Office and Copilot and they use Databricks and we want to make sure they have a great experience. So it's kind of it starts from there. The reason we think it's really interesting is your context isn't just what's like in your data warehouse or your databases and it's not just what's in your files and your Teams transcripts and your chats and your emails.
It's all of that together. So by having Genie and Copilot work together our customers can get their full context from unstructured to structured and everything in between, semi-structured, all of that stuff, and have it be fed to Copilot or to other agentic systems, like say some code you're writing or some back-end agent that's trying to automate a process. It can tap into that full context.
Why do you think it's so hard for businesses right now to provide this
There's a few things. The first is just the scale and the size. Like like if you think about just the sheer amount of data that's out there and and often times it's not very organized. You don't have golden records. You don't have the right indexes. You don't have the right golden tables or semantic models. So there's a lot of data preparation and cooking that hasn't happened. The second thing is if if you think about like permissions and authorization. you can have the data, but should you have access to the data or is it safe for you to use the data with the model is also a big challenge.
and so like those are probably like two those are like classic data engineering problems, right? Do you secure it the right way and then you prepare it the right way? There were that was a challenge 10 years ago. It's a challenge today. Every company should have a great strategy around those two things. The third one which I think is more interesting is so much of our APIs and our programming was built for people or interactive applications. And I'll give like an example of search. If I go do a search as Charles, I want like 10 results with little tiny snippets because I want to be able to quickly scan it.
That's a great API in response for a person. Terrible response for an AI agent. Cuz an agent can't just read or can read more than just 10 responses. It can read a thousand responses and it doesn't just need a tiny little snippet. It even wants a bigger snippet or a summary of the document. and the way you expose that data even if you have it prepared is fundamentally different to get the most from an AI an AI agent. So rethinking and rebuilding our APIs and our data shape and our data formats to support the AI client is a growth that we all have to go through as an industry.
Databricks and Microsoft we have a lot of joint customers. What's one thing that you would recommend our joint customers right now go and do where we will really see the power of better together?
First get Copilot. That's the easy answer. but beyond that I would say prepare your data. Like cuz like data is only part of the equation. You have to do some work to go from data to context. That's the define your golden records, create the semantic models that matter, delete inaccurate or stale data, secure your data so that it's only retrievable by the right set of people. Like for the best model today and the best model six months from now and the best model six years from now, it will still require you to have done all of those things.
Those are some of the most high return and best leverage work you can do. So work on your data, get it in a good shape, prepare it over time.
Last question. For right now, we are going through a major transformation in careers. I'm sure you're hiring a lot of people. What's the number one skill that right now is very critical?
If I go look at who do you want to hire and grow and bet on in the organization, two big things I focus on. The first is are they a learner? Like are they willing to go challenge their assumptions, to experiment new things, explore new ideas? And you have to be a learner in technology because things change so quickly. Even in my own career, like right when I got started, that was the beginning of client server and then mobile really took off and then cloud then hyperscalers and now AI. In order to be like great across all of those things in just 15 16 years, you have to be willing to learn and reinvent yourself every single year.
So when we talk about this idea of like growth mindset versus fixed mindset in Microsoft, growth mindset is you're always willing to grow and improve. Fixed being you're kind of stuck in your ways. We always look for growth mindset. and the second is the thirst for doing a good job. If you want to you want to be driven and no matter what the responsibility is, you want to do a great great job with it, whether it's something simple like preparing telemetry for a product or something big like the new UI refresh of Copilot, you really just have to have this high bar and really push yourself to do the best possible job.
If you have those two things, just about anyone can be successful over time and it doesn't mean it'll always be like continuous improvement. You'll have your ups and your downs, but as long as you're always getting better and you're always trying to do a great job, things tend to work out.
Don't become comfortable with the skills that you already have. Continuously push yourself and utilize the new technologies. Maybe also do a lot of hands on, right? Learning not so much only about what reading what's in the news. what's one project that everyone should just go and build after watching this video?
I I would say that if you haven't used some of the agent coding stuff like GitHub Copilot or Claude code or Codex, get your hands on one of those things and just create a little app because it's like even if you're not a technical expert with just some natural language prompts and some guidance and coaching, you can create amazing things. And if you think about coding, which is one of the hardest things to get into and to learn, if it's that easy to get started and start creating things now with AI, just imagine what's going to happen as all parts of companies and departments and functions and domains enjoy the same type of AI value, you know, what's the future of sales and finance.
Anyway, so go get hands on, code something up, see what what it can do.
Do you think it's still critical to have the coding skills or it's really right now all about understanding your business requirements and anyone can truly build a useful application?
I I would say like architecture and system thinking is more important than ever. So you have to understand how it all fits together, but like being able to read and write individual lines of code, I think that era is behind us. It's like it's like being able to read and write assembly 10 years ago. Like I can't read assembly. I write in a higher programming language, a compiler converts it. We're going to write in a new programming language called natural language and that will then there'll be an AI to compile it into programming language which is then compiled into assembly.
So learning coding syntax probably not super important, but understanding architecture and systems and resiliency and failover, that will still be very important. I don't see a shortcut through that in the near term and the business requirements. If you don't build something that people want, it doesn't matter.