AI that argues back improves your work, and AI beats accountants at routine month-end tasks

Chatbots tend to agree with whoever is typing. They learn partly from people rating their answers, and people rate the agreeable answer higher, so a model asked whether a plan is good usually says yes and then polishes it. It also learns how you write your prompts, and it gives you back more of what you already think.

The fix is to ask for the opposite. One marketer had the AI invent customers who disliked the product, a new accounting certificate program, and their complaints produced ideas that the designers had never considered. Workers who deliberately use AI to challenge their own ideas produce noticeably better work.

The eagerness to please that makes a model agree also makes it follow an instruction to disagree, so one paragraph in the prompt is enough. For example, a procurement manager who wants to move a contract to a cheaper packaging supplier could paste in the proposal and write: "You are the finance director who must approve this, and you are looking for reasons to refuse. Give me your three strongest objections." A good model will come back with questions about the notice period in the old contract, the cost of testing new packaging on the line, and the plan if the first delivery is late.

Answering those questions takes half an hour. Hearing them for the first time in the approval meeting usually costs a second meeting and a week.

On four month-end tasks that meant digging through a company's working files, finding the right figures, calculating and reporting, the leading AI models were faster, more accurate and ten times cheaper than licensed CPAs. On complete, harder close tasks, AI still cannot close the books without supervision.

So the work moving to AI is the matching and searching that junior accountants learn on, and their job becomes checking the model and handling what does not fit. A finance team that stops giving juniors that work needs another way to train its future controllers.

Companies use about 50% more AI than in July and pay less for it in total. The average price that businesses pay for a million tokens fell 41%, to $0.68, as both companies cut prices and buyers moved routine work to cheaper models.

That gives a buyer leverage: ask for today's prices in any committed contract, and for the right to move work between models. Do not build next year's budget on those prices, though. Both companies spend billions to train and run their models, and the cuts will stop once one of them pulls ahead.

SAP has built an agent that posts journal entries and reconciles accounts, and Oracle and Workday are adding similar agents. Each acts on the records it finds. If one plant stocks a part in boxes and another in single pieces, an agent reordering 10 will get one wrong.

So the first weeks after switching an agent on are a cheap data audit. Let it only suggest, and have the team write down the cause of each wrong suggestion. Most will trace back to a few bad records, and fixing those costs far less than reversing an order.

The largest gains that finance teams report from AI are in forecast accuracy and the quality of decisions, the judgment work where finance has always been weakest. Cost savings come further down the list.

What separates the teams is whether they can show an auditor how the AI reached its numbers. Those that can are far more likely to report a big drop in errors: 33% of them, against 6% of the rest. A forecast that nobody can trace gets checked again by hand before anyone acts on it, so the time it saved is spent a second time. Recording what the AI used and how it reached each number is the work that lets its forecasts go straight into a decision.