Using AI to Build Your 2027 Budget

Where I am coming from

I spent my first twelve years in legal working in accounting and finance. The last fifteen I have spent in overall firm administration, which put me on both ends of the same numbers: overseeing actual expenses, building the budgets, gathering the data that went into them, challenging what came back, and explaining the variances afterward.

For the past nine months I have been digging deep into what AI can do with financial data.

I want to be clear about what motivates that, because in our profession the question sitting behind every AI conversation is whether it comes for someone’s job. Everything I have done with AI and its implementation so far pushes me the other way. I am not looking to eliminate people. I am looking to enhance the ones we already have, who are already excellent, who sometimes feel overwhelmed by larger projects (like annual budgeting), by giving them back the bandwidth to focus on the parts of the work they are exceptional at, and taking away the mundane data processing that so many of us find burdensome.

Budget season is a good place to test that, because it is mostly data processing wrapped around a comparatively small amount of judgment.

Tools

I have used Copilot extensively. For a budget project of this kind I would prefer Claude, but Copilot deserves its due first.

Copilot is the safe option for a lot of firms, and safe is not a criticism here. So many firms already live inside Microsoft 365 that Copilot arrives in an environment already vetted, already licensed, and already holding the files. That removes an enormous amount of friction, and friction is what kills these projects before they produce anything.

My preference for Claude on this particular job comes down to how the data gets there. In Claude I would create a segregated Project for the budget and point it at where the needed files already live. The alternative is uploading a large volume of proprietary firm documents into an AI environment, and for a budget project that volume is not small: years of financial statements, transaction detail, the chart of accounts, and budget email. Pointing at files where they already sit is a materially different posture than making copies of all of it. On the other hand “pointing at them” also means giving permission for Claude to access that file location.

Whichever platform you choose, the tier matters more than the brand. This is firm financial data and firm email, so it belongs under an enterprise agreement whose retention and training terms you have actually read. This is NOT something you want to do inside any free AI account or platform. (General rule of thumb: If it’s free, your data is what’s being sold and used.)

The challenges you will run into

Three problems show up in nearly every firm, and they are versions of the same thing: the record of what happened does not match the way people described it.

Accounting inputs versus theoretical categories. Budget categories get invented during the budgeting process, largely by people who do not work in accounting. They are reasonable groupings and they reflect how the firm thinks about its money. They simply do not correspond cleanly to how accounting books the spend. So every year there is a mismatch between what the budget said and what the ledger recorded, and reconciling it is manual work nobody has time for.

Mis-recorded expense items. Beyond the structural mismatch, individual items get recorded wrong. An invoice gets mis-coded in accounting. Or something gets mis-categorized a level up, during the higher-level budgeting process. Either way, the number you are trending against next year is describing something other than what you think it describes.

Missing depth of history. If you have not been at the firm long, you likely do not know the contract renewal cycles. And often times, even if you have been there for years, those cycles may never have been recorded anywhere well enough to consult during a budget cycle. Renewals arrive as surprises, and price increases arrive already effective. The normal alternative during budgeting? Plug in a reasonable guess percentage increase and call it good.

A path forward with AI

Here is the sequence I would use. It looks like more than it is, because most of the effort sits in the first four steps, and those are gathering rather than analysis.

Feed it five years of past budgets, financial statements, and budget versus actual comparisons. Five years is enough to distinguish a cycle from a coincidence, and these files are small.

Feed it two years of actual transaction history. This is the heaviest data in the project, which is why the window is shorter. It is also where the mis-coded transactions live, so you want it at the transaction level rather than summarized.

Feed it three years of emails about the budgeting process. This is the step people skip and the one I would insist on. The email carries the context behind the decisions, the reasoning that never survives into the workbook. It also does double duty. In those same threads the model will pick up vendor contract renewal dates, and it will likely surface actual renewal proposals sitting six months out that nobody has flagged yet. Keep in mind, asking AI to pull budget and Vendor emails is a pretty large ask and does require some finesse and patience to get what you need. Secondary warning, you’re not choosing the individual emails, which means it’s going DEEP into email and making judgements about what pertains to budgeting, expense, spending, etc. With Copilot, assuming you’re in M365, it already has access to all of it anyway.

Feed it your GL code listing. Mark which codes are active and which are not. If your firm keeps notations for retired codes showing where those items now point, include those too. This listing is what lets the model connect the budget’s language to accounting’s language, and the retirement notes are what keep a five-year trend from breaking at the exact point a code changed.

Then ask it to analyze everything you have given it, and let it explain what it found. Read the findings and give it input on the ones that look off. It will make mistakes. Specifically, when it cannot find a definite answer it will produce an assumption based on the data instead, and it does not always draw a bright line between the two. Telling it which findings are wrong, and why, is not overhead. It is the step that makes everything after it worth having.

Then let it build out the basis of a budget. Not the final budget. The basis. Give it more feedback and ask it questions about its assumptions rather than its totals, because the assumptions are where the errors live.

One thing to watch for here, and it may be the most valuable output in the whole exercise: when it drafts something like this it will usually produce a methodology document whether or not you asked for one, showing how and why it made the assumptions it made. That can be a lot of information to process, and the temptation is to skim it. Do not. That document is your backup in March when someone asks why we chose $900 for XYZ this year.

Alongside it you will likely get a second “artifact” called a decision log, recording the input you gave it while the various output documents were being built. Every correction you made, every assumption you rejected, every piece of context you supplied that was not in the source files. The methodology document explains the model’s reasoning. The decision log records yours. Between the two you can answer a question that is otherwise impossible to reconstruct months later: not only why the number is what it is, but who said so, and when.

Summary

None of this hands you a budget you can submit. It hands you a basis, a set of findings you have already corrected, and a written record of the reasoning behind every assumption in it. You still choose the numbers and you still defend them.

What it takes away is the part that was never a good use of an administrator’s time: the reconciling, the reading, the cross-referencing between systems that do not agree, the hunt through two year old email for a renewal date. That work has to happen for a budget to be any good. It just does not have to happen at your desk, at night, in October.

So, back to where I started. The value here is not headcount. It is giving capable people back the hours they were spending on data processing, so they can spend them on the parts of this job that drive their passions for excellence.

Richard L. Wood, MBA is principal at kapchur consulting, where he helps professional services firms deploy operational and financial solutions through AI implementation. He can be reached through the contact page.