Doing strategic planning yourself at night with Excel and ChatGPT: where it works and where it breaks

By the team at Human Ready · Updated August 2026

This setup is not a confession. At a company doing a few hundred million in revenue, the CFO doing strategic planning personally, in a spreadsheet, with a language model open in the next tab, is often the correct answer: it is fast, it costs nothing, and it puts the person with the best judgement closest to the numbers. It also breaks in three specific places, and the breaks are predictable enough to plan for. This page is about which parts of the stack to keep, which parts are quietly costing you, and what to fix before anyone sells you software.

Why is the CFO doing this personally in the first place?

Because at mid-size there is nobody else, and that is structural rather than a staffing oversight. A CFO at a mid-market media company described the shape of it plainly (translated from Portuguese, as all buyer quotes on this page are):

"We're a relatively unsophisticated and small company. I have one person in strategic planning sitting inside the finance area. But the macro strategy, looking at the market, spotting opportunities, that's done by us at board level. Because there are few of us to discuss it."

In a large enterprise, strategy and FP&A are staffed functions and the executive receives the analysis. At mid-size the executive is the analysis function, on top of running the company. The work does not get delegated because there is nobody to delegate it to, so it moves to the only unoccupied hours in the week.

That relocation has a cost nobody books. A managing director at a mid-market industrial group told us he arrives at the office at 8am, an hour before anyone else, because it is the only uninterrupted thinking time he gets. At 9am the floor opens and the day turns reactive. His own summary: "The difficulties of day-to-day absorb so much of our time that sometimes there is no time left to think." McKinsey's research on decision-making puts executives at nearly 40% of their time on decision-making, most of it, by their own account, poorly used. The 8am hour is what happens when a person tries to reclaim that percentage manually.

What does the Excel and ChatGPT stack genuinely do well?

More than most vendors will admit. Four things hold up in practice.

Competitive and peer benchmarking. This is the strongest case, because the alternative was never a platform, it was a request that got deprioritised. The same mid-market CFO:

"I used to ask my finance director for a comparative analysis. Poor woman, she was closing the month: 'I'll pick that up next week.' Now in two minutes I not only have the whole comparison, I keep asking questions. Were there new deals? And this, and that. Of course it can have gaps, but I have all the assumptions in front of me."

That work used to fall off the list entirely. Now it happens. That is a genuine gain and no software purchase is required to keep it.

Framing before analysis. Structuring the question, listing what the board will push on, stress-testing an assumption before building anything around it. A model is a good thinking partner precisely because it argues back at no cost.

Documents. Contracts, covenant terms, auditor letters, analyst reports. The source fits in the context window and you can check the output against it by eye.

A spreadsheet as the model. Excel is not the weak link here. A driver model built by the person who understands the business, in a tool they can audit line by line, beats a platform model nobody trusts. Keep it.

Where does it break?

Three places, in the order they usually bite.

1. Consolidation and validation, not analysis. The failure is boring and expensive. From the same CFO, describing her budget process:

"What I want is a macro that kneads all the spreadsheets together and does the validations. One year we had errors, because someone forgot, and there was no cross-check saying 'look, this is falling sharply, it should be rising.' This is done manually. We don't have time for it. It's archaic. There's no other way to put it."

Twelve area owners, twelve files, manual cross-checking, one missed validation. The analysis was never the problem.

2. Numbers that cannot be reproduced or traced. Ask a language model the same question about the same data twice and you can get two different figures: different session, different day, different model version. Finance runs on reconciliation, and a number that will not reproduce cannot be reconciled or defended when someone else's version disagrees. The accuracy floor is measurable: in the FinanceBench benchmark (arXiv:2311.11944), a GPT-4-Turbo retrieval setup answered 81% of financial questions incorrectly or refused to answer. The failure mode that matters is not the refusal. It is the confident, fluent, wrong figure that reads exactly like a right one. What general AI does and does not do on company data is worked through in can ChatGPT or Claude replace a financial analyst.

3. The hour it runs on is yours, and it does not scale. The DIY stack has one operator. It works until the question arrives during the board meeting rather than the evening before it. The same managing director described his monthly board meeting as always containing one question he cannot answer in the room. There is nothing wrong with his data and nothing wrong with his spreadsheet. The question is simply new, and answering it needs a person, and the person is presenting.

Speed is the part that carries money. The mid-market CFO again, on why she rates it above precision: "It has to be very fast, because if it takes another month, you've already lost a month of cost adjustment."

What should you fix before buying anything?

In this order, because the first two are cheap and often enough.

  1. Automate the consolidation and the validation checks, not the analysis. This is where errors actually enter. It is also the least glamorous thing on anyone's roadmap, which is why it stays manual for years.
  2. Write down the definitions. One page: what margin means, which entities roll up where, which adjustments are standing. Every tool you might buy later, and every model you prompt today, inherits this ambiguity if you skip it.
  3. Keep the language model away from the numbers, and close to everything else. Documents, framing, drafting, external research. Never the source of a figure you will act on.
  4. Time the loop. From "leadership asks" to "answer on a slide", measure it for a month. If the honest number is a week and the decisions it feeds are worth more than that, you have a business case. If it is a day, you do not, and no vendor should convince you otherwise.

When do you actually need something more?

Three conditions, and you want at least two of them before spending money. The data lives in more than one system and someone reconciles it by hand every month. The questions now come from people other than you, which means the single-operator model has already failed. And the cost of being a month late is larger than the cost of the tool, which is the only version of this calculation that matters.

If none of those hold, stay where you are. A spreadsheet and a language model, run by a CFO who knows the business, is a genuinely good answer at this size, and buying a platform to replace judgement you already have is how mid-market companies end up with expensive software nobody logs into.

Where Advisor fits

For companies past those conditions, Advisor, built by Human Ready, is the version of this stack that survives being handed to someone else. It connects to the systems the data already sits in, models the business concepts once, computes every figure with deterministic engines rather than generating them, and returns an analysed answer with drivers decomposed and sources traceable. The conversational part stays, because that is the part that works. What changes is that the numbers behind it can be reconciled, and the answer arrives during the meeting instead of after it: humanready.io.

Related reading in this series:

Sources

Buyer quotes on this page come from Human Ready's ongoing conversations with finance leaders at European mid-market and enterprise companies, anonymised to role and company profile, and translated faithfully from Portuguese.

Page maintained by Human Ready. Last reviewed August 2026.