How to teach Copilot to write in your voice (and why it matters more than any other AI skill for finance professionals)

META: Most finance professionals use Copilot badly because they skip one step: teaching it how they write. Here is how to fix that. # How to teach Copilot to write in your voice (and why it matters more than any other AI skill for finance professionals) Most finance professionals I speak to have the same complaint about Microsoft 365 Copilot: it produces bland, generic text that sounds nothing like them. Board commentaries that read like they were written by a committee. Variance explanations stripped of any judgement. Management account narratives so flat they could belong to any company in any sector. They are not wrong. But the problem is not Copilot. The problem is that nobody told it how they write. This is the single highest-value AI skill a finance professional can develop right now: teaching the model to replicate your voice. Not your department’s voice. Not “professional finance tone.” Yours. I have done this. Not as a theoretical exercise, but because I needed consistent, high-quality financial writing produced at speed. The difference between generic Copilot output and voice-trained output is not incremental. It is the difference between a tool you tolerate and a tool you rely on. ## Why voice training matters more than prompt engineering The AI conversation in finance has fixated on prompt engineering: crafting the perfect instruction to get the perfect output. That matters. But it is second-order. First-order is context. A large language model produces generic output when it has no reference point for what good looks like in your specific context. Give it three examples of your actual writing and it shifts from guessing to mimicking. The quality improvement is immediate and obvious. This matters in finance more than most functions. Finance writing carries weight. A board commentary sets the tone for strategic decisions. A variance explanation can trigger operational changes or capital reallocation. An audit response needs to be precise, defensible, and consistent with how your function communicates. Generic AI output in these contexts is not just unhelpful. It is a risk. ## The three-sample method The process is simpler than most people expect. Pick three examples of your own financial writing. Good candidates: - A board commentary or management accounts narrative - A variance explanation with your typical level of detail - An audit response or controls documentation Paste these into Copilot Chat. Ask it to describe your writing style: sentence length, tone, level of formality, how you handle technical terms, how you structure explanations. Copilot will return a summary. Read it carefully. If it has missed something important, tell it. If it has overstated something, correct it. This is not a one-shot exercise: the back-and-forth is where the precision comes from. Save the final summary as a style card in the Prompt Gallery. This is your reusable voice profile. ## From style card to permanent voice A style card is useful. A custom agent is better. Microsoft 365 Copilot’s Agent Builder lets you create a personal AI assistant with your style instructions baked in permanently. No coding required. You set the instructions once and every interaction with that agent follows your voice profile automatically. For finance professionals producing repetitive documents, this is where the real efficiency gain sits. You are not pasting your style card into every prompt. You are not hoping Copilot remembers the context from three conversations ago. The agent carries your instructions forward, every time. The practical difference: instead of spending ten minutes editing Copilot’s generic output to sound like you, you spend two minutes reviewing output that already does. ## What Copilot actually pulls from One thing most finance professionals do not realise: Copilot does not operate in isolation. It connects to Microsoft Graph, which means it can pull context from your documents, emails, and files in OneDrive and SharePoint. This matters because your writing style is not just about sentence structure. It is about the level of detail you typically include, the metrics you reference, the way you frame commercial context around the numbers. When Copilot has access to your existing documents, it has a richer base to work from. The implication: keep your key finance documents in OneDrive or SharePoint where Copilot can reach them. If your board packs live in a shared drive that Copilot cannot access, you are cutting off the context that makes voice training effective. ## The vendor demo versus real life Here is where I need to be direct about what the vendor demonstrations do not show you. Microsoft’s demos make this look frictionless. Upload samples, get a style card, deploy an agent, done. In practice, the first output will not be right. It will be closer than generic Copilot, but it will not be right. Finance writing has nuances that general-purpose AI struggles with. How you handle negative variances versus positive ones. Whether you lead with the number or the narrative. How much context you give before the conclusion versus how often you lead with the conclusion and back it up. These are the things you refine through iteration. Generate a draft. Read it critically. Identify what feels wrong. Update your style card or agent instructions. Generate again. This loop is where the quality comes from, not from the initial setup. Expect three to five rounds of refinement before the output consistently feels like yours. That is normal. It is not a flaw in the tool: it is the nature of voice replication. ## Voice commands and hands-free drafting Copilot Voice adds another dimension. Available through the Microsoft Copilot app on desktop and mobile, and within Teams, Outlook, and Word, it lets you dictate rather than type. For finance professionals, the practical application is drafting on the move. Dictate a variance explanation during a commute. Start a board commentary narrative while reviewing reports. Ask Copilot to summarise data or flag trends, all by voice. The voice features connect to the same Microsoft Graph context, so your dictated prompts pull from the same document base as typed ones. And if you have built a custom agent with your voice profile, the output matches your style whether you type or speak the instruction. The key constraint: voice features are currently available in a limited set of languages through Microsoft 365 Copilot. Check your licence covers it before building workflows around it. ## Where this fits in a finance AI strategy Voice training is not the whole picture. But it solves a specific, high-frequency problem that blocks adoption. Most finance teams I have worked with abandon AI writing tools within weeks. The reason is almost always the same: the output does not sound like them, so they spend more time editing than they save. Voice training fixes the economics of that equation. Once the output is recognisably yours, the use cases compound: - **Monthly board commentaries** go from 90 minutes of writing to 20 minutes of review - **Variance explanations** maintain consistency across periods without manual enforcement - **Audit responses** follow your established tone and level of detail - **Internal reporting narratives** scale without losing the finance director’s voice None of this replaces judgement. You still read every output. You still own the conclusions. But the drafting work, the part that consumes hours every month, compresses dramatically. ## The refinement loop that most people skip The difference between finance professionals who find Copilot useful and those who dismiss it as a gimmick almost always comes down to one thing: whether they invested in the refinement loop. Setup is not training. The initial style card is a starting point. Real voice training happens over weeks, not minutes. Every time Copilot produces something that does not quite land, that is data. Update your style card. Adjust your agent instructions. Add another example to your reference set. This is the same discipline finance professionals apply to any process improvement: measure, adjust, repeat. The tool improves because you improve the instructions. Not automatically. Deliberately. ## What needs to be true first Before any of this works, three things need to be in place. **Your documents need to be accessible.** If your financial writing lives in local folders, email attachments, or systems Copilot cannot reach, voice training has nothing to work from. Centralise your key documents in OneDrive or SharePoint. **Your licence needs to cover it.** Microsoft 365 Copilot is a paid add-on. Agent Builder and Prompt Gallery are part of that subscription. Voice features have additional availability constraints. Confirm what your organisation’s licence includes before building processes around features you may not have. **You need to commit to the refinement.** The initial setup takes 20 minutes. The refinement that makes it properly useful takes a few weeks of periodic attention. Most people do the first part and skip the second. That is why most people think the tool does not work. ## What comes next Large language models will get better at voice replication. The gap between first draft and final output will narrow. But the finance professionals who build their voice profiles now will have a compounding advantage: months of refinement data that new adopters will need to replicate from scratch. This is not about being an early adopter for its own sake. It is about recognising that AI-assisted writing in finance is a when, not an if. The professionals who teach their tools how they think and communicate will produce better work, faster, with less friction. The ones who wait will eventually do the same setup. They will just be 12 months behind.