Imagine a good cook who already knows how to follow a recipe, check the pot, and fix a mistake. That cook can often learn a new dish faster than a total beginner. Microsoft did something similar with artificial intelligence (AI), software that can learn patterns and make helpful predictions. It took one AI model, a learned pattern system inside software, that had been trained to write code, the instructions people write for computers, and kept teaching it until it could do useful work in Microsoft Excel.

The simple idea

Think of a carpenter who moves from building shelves to building cabinets. The job is new, but the steady habits still help: measure, cut, check, adjust. Microsoft used the same kind of carryover with an AI model.

The model had already learned agentic coding, which means taking several steps, using software tools, and checking progress while writing code. Microsoft then continued training that same model for spreadsheet work in Microsoft Excel, instead of starting from a brand-new model.

A spreadsheet is a grid of boxes used to organize numbers, words, and facts. In Excel, the model had to do work such as write a formula, which is a spreadsheet instruction that calculates an answer, sort data, meaning arrange facts or numbers in order, and check whether the result made sense.

Rather than start over, Microsoft kept training a coding model for spreadsheet work. Reuse and Retrain Coding model Keep training Excel specialist
Rather than start over, Microsoft kept training a coding model for spreadsheet work.

Why code skills helped with Excel

At first, code and spreadsheets may look unrelated. Code looks like lines of text. Excel looks like a table. But both kinds of work often need the same basic habit: take a careful step, see what happened, then choose the next step.

This is why the earlier coding training mattered. The model had learned to act more like an agent, which is a software helper that can plan steps, use tools, and respond to results. A tool is a small action the software can use, such as running code or changing a cell in Excel.

The new Excel work did not erase the old coding skill. It built on it. The model carried over the useful part: how to work through a task in stages instead of trying to guess one final answer all at once.

📌
The key surprise. The important skill was not only knowing code; it was knowing how to take steps, use tools, and check the work.

Microsoft built a practice gym

A pilot does not learn only by reading about flying. A pilot also uses a flight simulator, a safe practice system that behaves like a real plane. Microsoft built a similar safe practice space for Excel.

That practice space is called a reinforcement-learning environment. Reinforcement learning is a way to train software by letting it try actions, see scores, and improve over many attempts. An environment is the practice place where those attempts happen.

The environment behaved like real Excel. The model could try spreadsheet actions, run them, see what changed, and get feedback. This mattered because spreadsheet work is active. The model needed to do things, not only answer questions.

What the model practiced

Picture a student doing many workbook exercises, not just one final exam. The Excel model practiced real spreadsheet workflows, which are series of connected steps people use to finish work.

It might write a formula, fill a table, sort rows, compare values, or repair a mistake. After each attempt, a grader, which is software that scores whether the result is right, checked the answer.

Then the model's weights were adjusted. Weights are the internal settings that shape how an AI model responds. You can think of them like tiny dials inside the model. Training nudges those dials so the model is more likely to make a better choice next time.

  • Write spreadsheet formulas that calculate answers.
  • Sort and organize rows of data.
  • Use Excel actions as tools, not just describe them.
  • Check results and fix mistakes.
  • Repeat thousands of practice tasks to improve steadily.
The model tries an action, sees the result, is scored by a grader, and its dials are adjusted. Practice Loop Try action See result Grader scores Adjust dials repeat
The model tries an action, sees the result, is scored by a grader, and its dials are adjusted.

How the model changed over time

Imagine saving a video game before a hard level. That saved point lets you continue from what you already learned instead of starting over. In AI training, a checkpoint is a saved version of a model at a certain moment.

Microsoft used the trained coding model as the starting checkpoint. Then it kept training that checkpoint on Excel tasks. Across later checkpoints, the model got better.

On tests, the percentage of tasks the model completed correctly climbed from below 72 percent to about 86 percent. That steady rise matters because it shows learning over repeated practice, not just one lucky result.

Across checkpoints the share of tasks completed correctly rose from below 72% to about 86%. Score Climbs Below 72%early Risingmore practice About 86%later start later
Across checkpoints the share of tasks completed correctly rose from below 72% to about 86%.

What the 2026 results showed

A smaller delivery van can sometimes do the same daily route as a huge truck, while using less fuel. That is the practical result Microsoft was looking for: strong performance without wasting computing power.

In 2026, Microsoft found that the Excel-trained model was on par with a top general-purpose model for the most common tasks. A general-purpose model is a model built to handle many kinds of work. The comparison was based on real user feedback from production, meaning the live service people use, not only a lab test.

The Excel-trained model was also more efficient, meaning it used less computing work, and cheaper to run. Because it was smaller, it could run on older, less expensive chips, which are the parts inside computers that do the calculations, as well as on the newest chips.

The real decision for leaders and builders

If you run a kitchen, you do not buy a new oven for every dish. You ask whether the oven you have can be used well, and whether cooks need better practice for a new recipe. Leaders and builders face a similar choice with AI.

The choice is not always between one huge model and starting over from nothing. Sometimes the better path is to reuse a strong model, train it in a realistic practice environment, and make it a specialist for one product or task.

This works best when the team controls the model, the software it works inside, and the tests that show whether the work is correct. Microsoft has that whole stack for products such as Excel. A stack is the set of connected layers that make a system work together.

✅
Practical lesson. Before building a new giant model, ask whether an existing model can become a cheaper specialist through focused practice.

How this connects to AI agents

A helpful office assistant does not just say what should happen. The assistant opens the file, checks the numbers, makes the change, and confirms the result. That is the idea behind an AI agent.

An Excel helper that plans steps, uses spreadsheet tools, checks results, and fixes mistakes is an agent. It is not magic. It is software trained to work through a task with feedback.

Microsoft is now extending this approach to more products, including Microsoft 365 Copilot Chat, Microsoft Outlook, and others. The larger lesson is simple: you do not always need a fresh giant model for each new job. You can reuse a good model, give it realistic practice, and create a focused helper that is capable and cheaper to run.

Key takeaways

  • Microsoft reused a coding model instead of starting a new Excel model from scratch.
  • The model's step-by-step coding habits transferred well to spreadsheet work.
  • A realistic Excel practice environment let the model try actions, get scored, and improve.
  • The model's test pass rate rose from below 72 percent to about 86 percent across checkpoints.
  • For common Excel tasks, the specialist model matched a top general-purpose model while costing less to run.
  • The same idea can guide future AI helpers in products such as Microsoft 365 Copilot Chat and Microsoft Outlook.
An Excel helper that plans, uses tools, checks results, and fixes mistakes is an agent. An Excel Agent Excel agentworks in steps Plans stepsnot one guess Uses toolsspreadsheet actions Checks resultsdid it work Fixes mistakestries again
An Excel helper that plans, uses tools, checks results, and fixes mistakes is an agent.

Glossary

Artificial intelligence (AI)
Software that can learn patterns and make helpful predictions or decisions.
AI model
A learned pattern system inside software that turns an input into a useful output.
Agent
A software helper that can plan steps, use tools, check results, and continue working toward a goal.
Reinforcement learning
A training method where software tries actions, gets scores, and improves over many attempts.
Environment
The practice place where an AI model tries actions and sees what happens.
Checkpoint
A saved version of an AI model at a certain point in training.
Weights
The internal settings that shape how an AI model responds.
General-purpose model
An AI model built to handle many different kinds of tasks instead of one narrow job.
Stack
The connected layers of technology that work together to make a system run.