An example I made up
Nadia Hashim's AI tool did two weeks of work in under an hour. The award was made a week later than planned. That week was spent on the one check that mattered. Here's how it went.
The situation
Nadia Hashim is head of procurement at a regional chilled-food distributor with its head office in Singapore. The contract for the trucks that deliver to and from the company's Singapore warehouse ends in six weeks. Three bids are in. Her director wants the award agreed at Monday's management meeting.
What the AI produced
Comparing bids like these used to take Nadia's team about two weeks. BrushfinchAI, the comparison tool the company uses, does it in under an hour. Its summary ends with one line:
"Bid B gives the best value. It costs 12 percent less over three years than the next bid, at the same service levels. I am 88 percent sure."
The table behind it is tidy, and the figures add up. It looks ready for Monday.
The check
Nadia asks two questions before she takes it anywhere. What did the comparison use? The three bid documents, and nothing else. Who hasn't seen it yet? The warehouse manager and the finance head.
She sends it to both. The warehouse manager reads it the same afternoon and phones her. The finance head has no change to make. Bid B prices 12-metre trucks. Two of the warehouse's three loading bays take trucks of 10 metres at most. Nothing in the bid documents asked about the bays, so the comparison had no way of knowing.
The decision, and who was told
Nadia doesn't put Bid B to Monday's meeting as the award. She asks Bid B and the next bidder to confirm a truck size against the bay plan by Thursday. Then she tells her director what is decided now and what waits. The shortlist of two is decided on Monday. The award waits for the next Monday meeting, after both bidders confirm a truck size. The cheapest bid may not fit the site.
What happened after
Bid B comes back with smaller trucks at a higher price. The gap to the next bid falls from 12 percent to 4 percent. The award still goes to Bid B, one week later than planned. Three months later, deliveries run with no queue at the bays. The team keeps one rule: "Before a fast comparison goes to the meeting, the people who run the site read it."
Everything in this example is made up, including the numbers and the timings.
Speed is the time saved on the way to an answer
BrushfinchAI saved Nadia's team about two weeks, and that saving was worth having. I'm not arguing for slowing AI down.
Progress is a different thing. Progress is a decision that holds once the people it touches have been asked what they know and the result has arrived. On that test, the work had not moved forward when the comparison was finished. It moved forward when the warehouse manager read the comparison, and the result arrived three months later, when deliveries ran with no queue.
The risk with a fast answer is that it can reach the meeting before the people who know the work have seen it. When a comparison takes two weeks, the people on the site often hear about it along the way. When it takes an hour, nobody hears about it unless someone asks them.
So the time saved is worth most when part of it goes to that check. Nadia spent a few hours of the time she saved on the two readers, and a week on the bidders' answers.
What one study found
A study published in March 2026 looked at a related question with real people. In a pre-registered experiment with 758 consultants from the Boston Consulting Group (BCG), those using GPT-4 worked 25.1% faster on tasks inside the AI's capability. On one business case designed to fall outside that capability, with 373 consultants, 84.5% of the group without AI reached the correct recommendation. In the two AI groups, 60% and 70.6% did. Graders rated AI-assisted recommendations as more persuasive, including wrong ones. (Dell'Acqua and colleagues, Organization Science, March 2026.)
The finding on accuracy comes from one task, done by early-career consultants with GPT-4 in 2023, so I don't treat it as a general rule. But it relates to the same gap as Nadia's case. A confident, well-written answer can be wrong and still persuade the people reading it. Sounding right is a style, not a signal.
What still has to happen
If your teams report that AI has made their work faster, that's good news. It tells you how fast the work moves. It does not tell you whether the decision holds. For that, ask who has read the answer since it arrived, and what they know that the documents don't.
AI can speed up existing work. People still decide what needs to change.
The AI can produce a comparison in an hour. A person still has to ask what it used and who hasn't seen it, before it goes to the meeting. Someone still has to tell the director which part is decided now and which part waits.
Speed is worth having. It becomes progress when some of the time it saves goes to the people who know the work.
I give keynotes, briefings and workshops on this.