Goldman Sachs published a report this month on AI among small businesses, and the headline number is the one every survey has produced lately: 86 percent of surveyed owners now use AI, up from 56 percent two years ago, while only 15 percent have fully integrated it into core operations. We have covered that adoption-versus-integration gap twice this month already. The more interesting finding is buried underneath it, and it is about what AI is being used for.
The shift
Use of AI for written marketing content fell from 74 percent to 65 percent. Meanwhile 42 percent now use it for financial analysis and business planning, more than double the 2024 rate. The report also found no widespread workforce reductions among the businesses surveyed.
Those two movements in opposite directions say something real. The first wave of small-business AI was writing: social posts, product descriptions, email copy. That is where the novelty was, and the decline suggests a correction, either because the output was mediocre, or because owners discovered that generic marketing copy is not actually the bottleneck in their business. The second wave is pointed at the numbers: cash flow, pricing, forecasting, reading a P&L.
Read the sample before you quote the number
Worth being precise, because this statistic is going to be repeated loosely. The respondents are 2,473 alumni of Goldman Sachs' 10,000 Small Businesses programme, out of roughly 15,000 surveyed, and they are American. People who complete a business education programme and answer a follow-up survey are not a cross-section of small businesses; they are an unusually engaged, growth-minded subset. So 86 percent is not 86 percent of small businesses. The directional finding, that use is shifting from copywriting toward analysis, is still credible, and it matches what we see.
Why this changes the risk
A bland AI-written LinkedIn post costs you very little. A confidently wrong cash-flow projection is a different category of problem, because financial analysis has properties that marketing copy does not.
It is hard to eyeball. You can tell immediately that a paragraph sounds generic. You cannot tell by looking that a twelve-month forecast quietly double-counted a quarter, or that a margin calculation used revenue where it should have used gross profit. The error is invisible precisely because the output format looks authoritative.
It compounds. A bad sentence is deleted. A bad number gets pasted into a pricing decision, a hiring plan, a loan application or a conversation with an investor, and by then nobody remembers where it came from.
And it is confidently delivered. Language models do not hedge in proportion to their uncertainty, which is tolerable for prose and genuinely dangerous for arithmetic presented as analysis. This is the specific failure mode behind fact-checking what AI tells you, and the stakes go up when the subject is your own money.
What to actually do
- Make AI show its work. Ask for the formula, the inputs and the assumptions alongside the answer. An analysis you cannot reconstruct is not an analysis you can rely on.
- Check one number by hand. Pick a single figure per output and verify it independently. If it is wrong, the rest of the output is suspect, and this takes two minutes.
- Keep it away from the source of truth. Use AI to interpret and summarize your numbers, not to calculate them inside your accounting system where the result becomes the record.
- Never let it near the irreversible step. Payments, bank detail changes, tax filings and anything a lender sees needs a human who understands the figures, which is the human-in-the-loop rule applied where it matters most.
- Mind what you are pasting in. Financial analysis means uploading financial records. Know whether your tool trains on that data and where it is stored.
- Write down where the numbers came from. If a figure in a plan originated with AI, note it. Future you will want to know.
The takeaway
The shift from copywriting to analysis is a maturity signal, and mostly a good one: businesses are pointing AI at work that matters instead of at filler. But the guardrails that were adequate for marketing copy are not adequate for financial modelling, and almost nobody has upgraded them. If AI is now helping you read your numbers, the useful habit is to make it explain how it got there, and to check one figure yourself every time.