On 29 January 1957 the result of the competition for an opera house was announced in Sydney: out of 233 entries sent from 32 countries, the jury chose the sketches of Jørn Utzon, a thirty-eight-year-old Dane. The estimate came to 3.5 million Australian pounds, or 7 million of the later dollars, and the opening was set for 26 January 1963.
Construction began in March 1959, although Utzon had not finished the design and the engineers did not yet know how to build the roof. The government of the state of New South Wales feared that support for the project would weaken, so it preferred to start work before anyone changed their mind. Between 1957 and 1963 the concrete sails from the sketch went through at least twelve versions. The solution appeared only in mid-1961, when all the roof shells were designed as sections of a single sphere, which made it possible to assemble them from repeatable elements. By then the building site had been working for more than two years.
The bill grew along with each new version. In February 1966, after the new minister for public works withheld the payment of fees, Utzon resigned, left Australia and never returned. Queen Elizabeth II opened the building on 20 October 1973, ten years behind schedule, and the final cost came to 102 million Australian dollars.
The money came from a source that the first estimate did not contain. A state lottery launched for this construction brought in about 102 million dollars between 1957 and 1986 from 86.7 million tickets sold. In 2007 the building was placed on the World Heritage List, a year later Utzon died without having seen it finished, and today more than 10 million people visit the opera house site every year.
Was the 1957 estimate a mistake, or an honest valuation of the part of the building that could be drawn at the time?
The budget for an artificial intelligence assistant is drawn up in companies in a similar way: the price of the license is known, so that is what goes into the estimate, although the hardest part, the change in how work gets done, has not been designed yet. I argue that the license is the only line of the bill that is certain on the day of purchase, and that the return on investment depends on whether someone has decided in advance what the recovered minutes will turn into.
What exactly you buy for 30 dollars a month
Microsoft 365 Copilot, the assistant most often considered in companies working on Microsoft's office suite, costs according to the vendor's price list (as of November 2026) 30 dollars per user per month with annual payment and requires a separate license for a qualifying Microsoft 365 plan. For 200 people that makes 72 thousand dollars a year before anyone writes a first instruction. The price list is indicative, so you will learn the final amount only from the reseller's offer.
The assistant is a language model, a program that, on the basis of a huge number of examples, predicts the most probable continuation of a text, connected to the company's mail, calendar, files and meeting records. When answering, it reaches for the documents the given employee has permission to see. Two costs absent from the price list follow from this design. The quality of the answers depends on the order in the documents, and carelessly set permissions mean that the assistant may obligingly summarize for someone a salary file that has been sitting in a publicly accessible folder for years.
What was measured on 20 thousand British civil servants
A large public measurement was carried out by the British administration. From 30 September to 31 December 2024 about 20 thousand employees from twelve institutions received Copilot licenses, and the government's digital agency (Government Digital Service) described the results in a report dated 2 June 2025. Participants declared an average saving of 26 minutes a day, which the report converts into about 13 working days a year, and 82 percent of respondents did not want to go back to working without the tool. The authors themselves note that the numbers come from a survey and that nobody checked what the recovered time was spent on.
That reservation was tested by the British Department for Business and Trade, which evaluated its own pilot on a thousand licenses from the same quarter. Most users saved time, the most when writing, whereas scheduling meetings and creating graphics took longer than without the assistant. The average participant performed 1.14 actions with it per working day, and 22 percent of those surveyed noticed hallucinations, meaning invented information delivered in a confident tone. The report states that no evidence was found that the time savings translated into higher productivity.
Data from Denmark give a similar picture. The economists Anders Humlum and Emilie Vestergaard linked surveys of 25 thousand workers from 7 thousand workplaces with payroll registers and reported in a 2025 paper that users of chats based on language models save on average 2.8 percent of working time. In earnings and in the number of hours worked they found no significant change, and their calculations rule out an effect larger than 1 percent.
Why the recovered minutes do not reach the company's results
Twenty-six minutes a day arrive in pieces: four minutes on an email, six on a meeting note. Such crumbs fill up by themselves with other small activities, because nobody has told the team what to do with them. People behave rationally here: since the target has stayed the same and the work goes faster, they keep the difference for themselves, and it is hard to hold that against them.
The bill for 200 people shows the scale of the misunderstanding. Thirteen days a year per person gives 2,600 working days across the whole group, on paper the work of about twelve full-time positions, and that is the number you will see in the vendor's presentation. Yet none of those twelve positions can be pointed to as long as the recovered time lies scattered across two hundred calendars in portions of a few minutes.
A minute turns into money in two situations. In the first, the team has a backlog or demand it cannot keep up with, so faster work means more offers sent or tickets closed. The second occurs when the company pays for overtime or outside services that it can give up. In both, someone has to change the arrangements, for example shorten the response time promised to customers from three days to one, and that decision will be made neither by the assistant nor by the information technology department.
Four lines of the full estimate
An honest bill for the first year contains, apart from the licenses, four more lines:
- Training on the department's real tasks. Four hours for 200 people is 800 hours of work, or five months of one full-time position.
- Order in documents and permissions before the assistant starts searching them.
- The time of the person who will redesign the chosen process and answer for its result.
- Checking of answers by employees, given that in the department's pilot one respondent in five came across an invention.
My working rule, which you will verify with your own pilot, is this: in the first year, at least as much has to be allocated to people's time as to licenses. A company that plans only the first line repeats Sydney in 1957, because it prices the part that can be drawn.
The tool copes well with writing and summarizing meetings, and does less well at scheduling and preparing presentations, as the department's pilot showed. Two risks are more dangerous than the price list: wrong information sent to a customer without checking, and a document shown to a person who should not see it. I do not recommend licenses for the whole company at the start, because with 200 users it is impossible to establish where a result came from or who answers for it.
Recommendation and a first step for Monday morning
Start with a team of 20-30 people and one process whose result can be counted without a survey. Six weeks are enough to see a difference, and annual licenses for 25 people cost 9 thousand dollars according to the price list. On Monday, take four steps before you order anything:
- Choose a process with a countable result, for example the number of offers sent in a week or the number of days needed to close the month.
- Write down the result from the last four weeks, so that you have a reference point.
- Formulate one sentence about what will happen with the recovered time, for example “we answer requests for quotation in one day instead of three.”
- Appoint a person who, after six weeks, will compare the result with the reference point and decide whether to extend or end the trial.
If the sentence from point three cannot be written, the purchase should be postponed, because the saving will disperse across calendars without a trace in the results. One question remains, worth asking at the next management board meeting: who in your company has the right to decide what happens to an hour recovered by a team?