Two years of pilots, press releases and promises. Then the invoices came due and the results did not. Here is what the backlash is really about, and what quietly replaces the black box.
For two years, enterprise AI sold a simple story: feed it your data, trust the output, watch the magic happen. In 2026 that story met the P&L. MIT's NANDA initiative found that roughly 95% of generative AI pilots never reach production with measurable impact. S&P Global reported that 42% of companies scrapped most of their AI initiatives in 2025, up from 17% the year before. This is not a technology failure. It is a trust failure.
Talk to the teams who pulled the plug and you hear the same thing. The tool produced a number, but nobody could say how. It gave a different answer on Tuesday than it gave on Monday, on the same data. When a planner asked why, the honest reply was a shrug. You cannot run an operation, let alone defend it to a board, on an answer you cannot reproduce or explain.
In demand planning and inventory, that is fatal. A forecast is not an interesting suggestion; it triggers purchase orders, ties up cash and sets service levels. If the number cannot be traced, every decision downstream inherits the doubt.
The companies still getting value from AI share one discipline: they separated the parts of the problem that benefit from learning from the parts that must be deterministic. Let machine learning sharpen the forecast, where more pattern-finding genuinely helps. But the decisions, how much safety stock to hold, when to reorder, how much cash to free, should run on transparent math that gives the same answer every time.
That is not a step backward from AI. It is what makes AI usable in a regulated, audited, cash-constrained business. The model proposes; a method you can read disposes.
Three questions separate a glass box from a black box. Can you rerun last month's result on the same data and get the identical number? Can you show, line by line, how any single figure was built? Will the number survive an audit six months later, when the person who ran it has moved on? If the answer to any of these is no, you are buying an opinion with a confidence interval painted on.
Most software asks you to trust it. The market has stopped agreeing to. The winners now show their work.
Picture your next S&OP meeting: one number on screen that nobody argues with. See it free, on your own data, in 45 minutes; and if it does not hold up, we part as friends.