Processes
Checking becomes the bottleneck of your commercial organisation
AI puts a quote on paper in minutes. So why does your inside sales team have no time to spare at the end of the week?
A technical quote that used to take half a day is on paper within minutes with AI. Neatly laid out, fluently written. Yet at the end of the week your inside sales team rarely has a day and a half to spare. Where does that time go?
The answer lies in a part of the work that belongs to every quote but rarely stands out: checking whether what it says is actually right. With AI, the work moves from making to checking. That checking takes more time than you would expect. It also gets harder and ends up with a handful of people in your company.
Time saved per task, not per week
The biggest promise of AI is saved time. Employees say so themselves in the Work Trend Index by Microsoft and LinkedIn (2024). Does the promise hold? The answer is more nuanced than you might think. Look at a single task and AI certainly saves time: a quote is on paper faster.
In 2025 two Danish economists looked at a full working week instead of a single task. Only a small part of the time saved was left: about an hour a week, or 2.8% of working time (Humlum and Vestergaard, 2025). Most of the time saved apparently goes to other work. Part of that work is new: checking what AI produces.
The big figures on time saved often come from experiments that measure one well-defined task. In 2023 two MIT economists gave a group of marketers and consultants a short writing assignment from their own field, such as a press release or a short report. The participants using ChatGPT finished 40% faster on average (Noy and Zhang, 2023). But they often copied the answer with few changes. A thorough check, of the kind a real quote demands, was hardly part of that 40%.
Reviewing takes the place of making
A more recent test did include the reviewing. In 2025 OpenAI set up GDPval, a large test with professionals from 44 occupations, from lawyer to sales manager. These professionals, with fourteen years of experience on average, each received an assignment from their own practice, such as drafting legal advice or a sales plan. The same assignment also went to an AI model. Other professionals then compared the two results without knowing who had made what. The model did surprisingly well: in almost half of the assignments the assessors rated the AI result as good as the professional's, or better (Patwardhan and colleagues, 2025).
A good result says nothing yet about time saved. So the researchers also tracked how long an assignment took. A professional without AI needed almost seven hours on average. The model delivered the same assignment in 4.5 minutes. Reviewing that result took a professional 109 minutes on average. Together, an assignment with AI takes just under two hours, 3.5 times faster than without AI.
Two hours is the best case, though: the model gets the assignment right in one go. Often it did not. Then the professional asked the model for a new attempt, or ended up doing the work himself after all. With those extra rounds, an assignment with AI took just under five hours. That is still 28% faster, but far less than the model's 4.5 minutes suggest.
More important than the time saved is how the time within one assignment is divided. It used to go almost entirely into making. Writing legal advice took a day. Reviewing was a small piece of work at the end. With AI that ratio is reversed. Making takes a few minutes. The work now sits in checking: reading, comparing, verifying and reworking what is wrong.
Without AI, making was a process of kneading. Your colleague wrote a draft, tested it against his experience, earlier quotes and the agreements with the customer, and improved it. He repeated that round until the quote was right. Checking was part of making. With AI he receives a finished text and has to verify whether his knowledge is in it. Such a check is hard, because when you review you mainly judge what is there. An error in the text stands out far sooner than a missing agreement (Mosier and colleagues, 1998).
A fluent text hides its errors
Checking does not only take more time, it also gets harder. A sloppy quote used to betray sloppy thinking. A typo in the price, a misspelt customer name or messy layout suggested the quote had been put together in a hurry on a Monday morning. Such sloppiness put the sales manager on guard. Today a quote from AI almost always reads smoothly, even when the content is wrong. Those signals of sloppiness disappear.
In 2023 Boston Consulting Group showed how far that can go. Researchers gave a group of consultants figures on the brands of a fictional company, together with notes of interviews with employees. The consultants had to advise the CEO on which brand to invest in. Half of them worked with AI, the other half without. There was only one right answer. It was hidden in details of the interviews. The consultants with AI reached the right answer 19 percentage points less often. Yet the assessors gave their advice higher marks for quality: they found it more convincing and better written (Dell'Acqua and colleagues, 2026).
Even a targeted check for errors misses a lot. In 2024 researchers had language models summarise news articles, write product descriptions and answer practical questions. Experts found a factual error in one in five of those texts. Random assessors spotted one in only one in twenty-five. The more confident the answer sounded, the fewer errors those assessors saw (Hosking and colleagues, 2024). Nobel laureate Kahneman gave this link a name long ago: cognitive ease. The more easily we read a text, the sooner we believe what it says.
If you review quotes, proposals and emails every day, you may recognise this. A quote from a junior using AI reads just as fluently as one from an experienced salesperson, even when the junior barely knows the customer. To find the errors anyway, the sales manager has to read more slowly and check for himself what it says. Reading slowly takes time, and that time was already scarce. He has to review more quotes, and each quote also needs more attention than before.
Where the work piles up
Despite all that reviewing, AI still saves time on a single quote. If the ratios from GDPval also apply to your inside sales team, a technical quote of half a day takes about three hours with AI, reviewing and reworking included. That makes room for about 40% more work in a week: with 28% less time per quote you can produce 1 / 0.72 = 1.39 times as many.
That sounds like a luxury problem. But the number of requests does not stay the same. And you do not control that flow. In the previous part you saw where it comes from: your colleagues make quote variants, your customers send questionnaires of eighty points, and buyers send their request for quotation to more suppliers than before. The extra room is used up sooner than you would think. Quotes wait on a desk until someone finds time to check them. Meanwhile the customer is waiting.
That checking gets stuck like this has also been measured. In 2026 Anthropic, the maker of Claude, described how the work of its own employees was changing. Much of that work runs through agents. Such an agent carries out a task on its own, like writing code or drafting a report. In August 2026 the agents of one employee together already worked six times as many hours as that employee. But those agents stood idle about 80% of the time, waiting for a response from an employee. One engineer summed it up drily: the humans are holding the agents up (Anthropic, 2026).
The academic journal Organization Science received 42% more submissions in 2025 than before ChatGPT, almost entirely papers with AI-written text. Every submitted paper first has to be read by an editor, even a weak one. The editors rejected almost 70% of the papers with a lot of AI text straight away. To handle the volume, the journal doubled its editors, from about thirty to sixty. Even so, some editors still received more than 250 papers a year (Gartenberg and colleagues, 2026).
The journal could still double its editors. A small company with one sales manager cannot simply add a second. Checking has become the bottleneck in two ways. The volume grows faster than people can review. Each text also needs more attention. So the sales manager has to be able to choose which quotes need his judgement and which can go out without him. In many companies the commercial organisation is not ready for that: sign-off sits with a handful of people. The rules they check against are written down nowhere. In the next part you read how to build that judgement into your organisation.
Follow the series
LET'S MEET
DOES YOUR COMMERCIAL ORGANISATION NEED A NEW DESIGN?
Happy to meet you for a short introduction.
