Almost every white-collar professional now has access to some kind of AI tool, provided by their employer in the hope that it will make their work faster, better, or more accurate. Most use it to rewrite emails and generate the occasional funny image for a Teams channel. There is, in a sense, nothing wrong with this use of AI, but so much is being left on the table.
I started using ChatGPT about two weeks after it launched, which places me, according to OpenAI’s own retrospective, somewhere in the first 0.1% of users. That fact is less impressive than it sounds. Being an early user just meant I was curious, not that I was doing anything productive with the tool. For the first three years I used AI the way most people are using it now: for emails, for clarifying questions I half-remembered from a textbook, for entertaining myself. The shift to using it for actual work didn’t happen until I started my current position in late 2025, when the combination of a new role, new software, and a firm-issued Copilot license forced me to start asking better questions of the tool.
In the six or seven months since, I have learned enough to notice that almost no one around me uses AI for anything beyond what they were using it for in 2023. The most common use case is still writing emails. The second most common is rewriting emails someone else wrote. The third is asking it to define a concept the person already understood and just needed to confirm. These are reasonable things to do with AI. They are also, to put it bluntly, a waste of the tool
.
The first thing I used Copilot for at work that I would now describe as genuinely useful was a side-by-side comparison of two large XML files. The files in question were eFile returns that get sent to the IRS, roughly 80,000 characters each, full of nested elements that look like a wall of brackets. I needed to know what had changed between an earlier version and a later one, and the change was not obvious. I uploaded both files to the secured, client-safe version of Copilot that my firm makes available, asked it to identify the differences, and within seconds had a clear, structured list of exactly what had moved, what had been added, and what had been removed. The same task, done by hand, would have taken the better part of an hour and produced at least one missed entry.
That example, comparing two structured files, is the most narrowly defined version of what I want to gesture at. The broader category is something like: any task that involves comparing, transforming, or reformatting information that already exists in your possession. Copilot can take a timesheet kept in one format and convert it into the import format your firm’s time-tracking software expects. It can read a forty-page memo and produce a one-page summary in the voice of your practice group. It can take a small workflow you have described in plain English and write a single-page HTML application, runnable in your browser with no developer involved, that handles a piece of your week you would otherwise be doing by hand. None of these are theoretical. All of them are things I have done or watched a colleague do with the tools the firm already pays for. Nothing outside of a simple chatbot subscription.
This is the place where most articles about AI at work go off the rails, because the writer starts talking about ChatGPT, or Claude, or some agentic startup that just raised a Series A. You have probably heard of Anthropic, OpenAI, Google, xAI, and a half-dozen others. You may have read about their models with interest, but with no real intention of using any of them at work. None of that matters very much if you are reading this from a work laptop. Your firm has almost certainly made a decision about which AI you are allowed to use, and the decision was almost certainly Copilot. Microsoft sells the operating system, the productivity suite, the calendar, the email, and now the AI as a single bundle, and Microsoft’s enterprise contracts are written in a way that compliance officers can defend on a Monday morning. If you work in a Google shop, the answer is Gemini, for the same reasons. Either way, the tool you actually have is the tool this essay is about.
There are good reasons your firm has narrowed the field. The free, consumer-facing versions of these tools, the ChatGPT you might use at home, the public Claude site, the Gemini chatbot signed into your personal Google account, make no enterprise-grade promises about what happens to the text you type into them. For a firm that handles client financial data, legal records, audit working papers, or anything else governed by a confidentiality obligation, that is not a small problem. Health-care clients have it worse. Some HIPAA-regulated organizations are required to use specific certified data centers, and if the model you want to use is not hosted in one of those data centers, the answer from compliance is no, and that decision is final. Compared to that, the average professional-services firm has a relatively easy set of rules: use the enterprise tool the firm has licensed, do not paste anything you would not put in a discoverable email, and you are mostly fine.
The thesis of this blog is that, within those rules, there is more capability than almost anyone is using. Copilot alone, without ChatGPT, without Claude, without any tool your firm has not approved, is enough to put a thoughtful user well ahead of nearly all of their colleagues. Not because Copilot is the best AI on the market; it isn’t. But because the gap between “uses Copilot to rewrite emails” and “uses Copilot to compare two 80,000-character XML files or build a single-page web app for timekeeping” is not a gap of access. It is a gap of applied creativity.
In the next several posts I will work through some of the specific techniques I have found useful: a method for using AI to check its own answers by hiding the result and asking it to solve the same problem alongside yourself; a walk-through of building a single-page browser application using only Copilot and a text editor; and a piece on configuring Copilot’s agent features with a small JSON file so that everything it produces comes back in your firm’s brand colors. None of these require you to be a developer. None of them require any tool your firm has not already given you.
If you would like a small thing to try this week, here is one. Find a task you did recently that involved comparing two versions of the same document, a contract redline, two drafts of a memo, two exports from a system, and ask Copilot to identify the differences. Then check its work against your own. The point is not whether the answer is perfect; the point is to see, for yourself, how close to useful the tool you already have already is, and to fill in the rest with techniques covered in future posts.
Samuel L. McHargue