Jagged Behavior: Why Your AI Model Needs to Be Tuned to Your Enterprise

TL;DR: Enterprise AI tuning closes the gap between what a generic model can do and what a specific business needs it to do. Ranveer Chandra, Vice President of Frontier Tuning and CTO of Agri-Food at Microsoft, built his answer around a problem he calls jagged behavior: the same model that solves a hard technical problem can fail at something a ten-year-old would get right, because it was post-trained on general synthetic data instead of an enterprise's own workflows. In this third episode of The AI Friction Files, recorded at IIT2026 in Long Beach, Chandra explains why organizing data into usable knowledge is the prerequisite for token efficiency, how an enterprise tunes a model, and where most organizations sit on Microsoft's three-phase path to what it calls a Frontier Firm. His core argument: AI isn't replacing people; it's multiplying what each person can do. The work now is learning to manage a team of agents instead of doing every task by hand.

Is AI Overhyped? Chandra Says No, and Explains Why

Asked directly whether AI is overhyped, Chandra's answer is blunt. "It still hasn't permeated in all parts of society. It's evolving super fast and the adoption rates are still growing. We are still in the middle innings of AI."

His evidence comes from his own work outside enterprise software. As CTO of Agri-Food at Microsoft, Chandra works closely with farmers, and the adoption gap there is stark. "They still are not using much AI," he said. "There are huge segments of society that are not using AI yet in a big way, and we need to get it everywhere."

The distinction matters for how enterprises should think about their own AI maturity. If entire sectors haven't started yet, being early isn't a sign that the technology is stalling, but rather a sign that most of the curve is still ahead.

Jagged Behavior: The PhD Student and the Ten-Year-Old Problem

The adoption gap isn't the only friction Chandra names. Even where AI is deployed, it doesn't behave consistently, and he has a specific explanation for why.

"One of the things I was talking about yesterday is how these AI models show this jagged behavior," Chandra said, pointing to a well-known observation from AI researcher Andrej Karpathy: that today's models can act like "a PhD student and a ten-year-old at the same time."

Chandra's explanation for the inconsistency is architectural. "This is because they haven't been post-trained in different environments," he said. "You can post-train it on general synthetic data, but you take it to an enterprise, it doesn't work as well." A model trained on the open internet has no exposure to a specific company's terminology, approval chains, or edge cases. It performs unevenly because it was never taught the shape of that particular business.

"So how do you then customize these models for that particular enterprise?" Chandra asked. "That's one of the capabilities we're building with Frontier Tuning."

Data Isn't the Same Thing as Knowledge

Before an enterprise can tune a model to its own operations, Chandra argues it has to solve a prior problem: most organizations don't have their data organized in a form a model can learn from.

"Data is a new oil," he said, echoing a comparison others have made before him. "But just the data by itself doesn't mean anything. It could be fragmented. It's unstructured. It's noisy." Raw data sitting in a system isn't automatically usable. It has to be turned into something a model can act on.

"It's not just data, it's knowledge," Chandra said. "Every enterprise has their own data, and from that data you need to extract the knowledge. You want to organize the data. The better the data is organized, the better the knowledge is organized, the better the AI is, and that will drive better token efficiency."

Token efficiency is Chandra's term for getting more useful output out of every token an enterprise spends on inference. "Tokens are important," he said. "You want to get the most out of every token you spend. There's a lot of work happening on token efficiency right now, and everyone needs to be conscious about how they're using these resources."

He also points to a second source of enterprise knowledge that rarely gets captured at all: what's in people's heads. "A lot of things are in people's minds, and AI needs to know a little bit of that to help them be more productive," he said. His guidance to enterprises is to codify what can reasonably be codified. "Not everything will be codifiable, but whatever can be codified, codify it, because then you'll be more productive. You should figure out what are the biggest opportunities in your enterprise."

How to Tune a Model to Your Enterprise

Chandra lays out a concrete starting sequence rather than a theoretical one.

The first step is building what he calls an eval set: "prompts and rubrics," where rubrics function as assessment criteria or metrics. "You want to evaluate how well existing models work" against the tasks that matter most to the business, before deciding what to change.

From there, Chandra suggests starting with what's already available. "You could start even with Copilot," he said, referring to Microsoft 365 Copilot. "Purchase it. It just immediately starts getting all your data in an enterprise secure way. It preserves all the security things you get from Microsoft. It comes with that. If it works, great, you're just off the races."

If the off-the-shelf result isn't good enough, the next step is what Chandra calls hill climbing with harness engineering: "You start modifying the context, you add more tools, you change prompts, things like that. You've taken your eval metrics from a certain point to the next point." Frontier Tuning is Microsoft's answer to that step: a system for teaching a model an enterprise's own workflows and conventions through structured reinforcement, rather than generic prompting alone.

The Three Phases to a Frontier Firm

Chandra frames all of this against a broader trajectory Microsoft calls the Frontier Firm, a term the company introduced in its 2025 Work Trend Index research and has continued to develop since. "A frontier firm is this firm that just does everything much more efficiently," he said.

Microsoft's own research lays out the same three phases Chandra described: most organizations today sit in phase one, where a person asks an agent a question and the agent responds. Phase two is where agents start working alongside people as what Microsoft calls digital colleagues, taking on specific tasks under human direction. Chandra's own example: one of the roles attending his meetings now is his boss's chief of staff, filled by an AI digital worker rather than a person. Phase three, which Microsoft's research says is still ahead for most organizations, is where every employee manages a portfolio of agents doing different jobs, a role Microsoft calls the agent boss.

"Everyone becomes an agent boss," Chandra said. "All of us will have multiple agents to do different things, and it's all about managing the agents to do their job." He connected this to a framing from Microsoft CEO Satya Nadella: "The future of employment is that every person will do macro delegation and micro steering."

Humans Aren't Being Replaced. They're Being Multiplied

Chandra pushes back directly on the narrative that AI eliminates jobs. "It's going to be humans in the loop too," he said. "Humans will be much more productive with everything in AI. Instead of coding being replaced, like people say, one can do the job of ten, but ten can do the job of a hundred."

"I'm of the other camp," he said. "I'm not saying that humans are not going to be there. Humans will be much more productive with AI, and they'll be productive at a different level with AI." His framing is additive rather than substitutive: AI takes over enough of the routine work that people get time back for strategic work they weren't able to reach before. "Imagine AI doing a lot of that, giving you time to do strategic stuff, expanding the amount of other stuff you could do."

That philosophy shapes how Chandra's own team validates what it builds. "We try things ourselves. We call it dog food," he said. "We are customer zero. Every department, my own team, tries things before we give it to even other teams within Microsoft, and we see the problems firsthand."

The Diagnostic: Where Your Enterprise Sits in the Tuning Process

Chandra's framework translates into a short set of questions any enterprise architect or AI program owner can run before investing further in a generic model deployment:

Do you have an eval set for the tasks that matter most to your business? If the answer is no, there's no way to know whether a model is underperforming or whether expectations were never defined.

Is your data organized enough to become knowledge, or is it still fragmented and noisy? Token efficiency depends on this. A model working against disorganized data burns tokens without a proportional return.

Have you tried the off-the-shelf option first? Chandra's own sequence starts with Copilot before moving to custom tuning. Skipping that step means paying for customization the enterprise may not need yet.

If the off-the-shelf result falls short, are you hill climbing systematically, or randomly changing prompts? Harness engineering, adjusting context, tools, and prompts against a fixed eval set, is what turns trial and error into measurable progress.

Which phase of the Frontier Firm is your organization really in, versus which phase leadership assumes it's in? Most organizations are still in phase one. Knowing that honestly is the starting point for planning phase two.

Watch the Full Conversation

This post covers the core of Chandra's conversation with Sujatha Padmanabhan at IIT2026. Watch the full episode of The AI Friction Files to hear it in his own words, on camera.

Back to Episode 1: Governance by Design: Why Retrofitted AI Guardrails Fail

Back to Episode 2: The CFO Token Budget Problem: Why AI Agent Governance Can't Wait

Part 3 in Sparq's series on The AI Friction Files.

Frequently Asked Questions

What is jagged behavior in AI models?

Jagged behavior, a term used by Microsoft's Ranveer Chandra, describes how the same AI model can perform an impressive, complex task and then fail at something far simpler. Chandra attributes the underlying observation, that models can act like "a PhD student and a ten-year-old at the same time," to AI researcher Andrej Karpathy, who has written about this inconsistency under the related term "jagged intelligence." Chandra's explanation is that models trained on general synthetic data haven't been post-trained on a specific enterprise's own workflows and context.

What is Frontier Tuning?

Frontier Tuning is a Microsoft capability, led by Vice President Ranveer Chandra, for customizing AI models to a specific enterprise's data, terminology, and workflows. It follows an initial step of building an eval set and rubrics, and applies hill climbing with harness engineering, adjusting context, tools, and prompts, to close the gap between generic model performance and what a business needs.

What is a Frontier Firm?

A Frontier Firm is a term Microsoft introduced in its 2025 Work Trend Index research to describe an organization built around on-demand AI intelligence and human-agent teams. Microsoft's framework describes three phases: agents that respond to direct questions, agents that work alongside people on specific tasks, and a final phase where every employee manages a portfolio of agents, a role Microsoft calls the agent boss.

Why does data organization affect AI token efficiency?

Poorly organized, fragmented, or noisy data forces a model to work harder for the same result, consuming more tokens without a proportional gain in output quality. Chandra argues that turning raw data into organized knowledge, including tacit knowledge that lives in people's heads rather than systems, is what allows a model to produce more useful output per token spent.

Is AI going to replace human jobs?

Chandra argues the opposite: that AI expands what one person can accomplish rather than eliminating the need for people. His framing, echoing a common industry comparison, is that one person could previously do the work of ten, and AI-augmented, that same person can now effectively do the work of a hundred, freeing time for higher-value, strategic work rather than reducing headcount.