A chatbot answers a question. Yesper does the work

An engineer at one of the firms we work with put it in a sentence: "A chatbot answers a question. Yesper does the work, just like a colleague." The sentence draws a category line, and the past two years have been lived on one side of it. The profession gave every desk something to ask, and the desks got good answers. But an engineering deliverable is not a question: it is a dozen interdependent parts that must agree, and end in one document an engineer signs. Holding them together is the actual job.

Freshly cast concrete surface with forested hills in the background

Every desk got its assistant in two years

ChatGPT launched on 30 November 2022 and reached, by a UBS analysis of Similarweb data, a hundred million monthly users in January; the bank's analysts wrote that in twenty years of covering the consumer internet they had never seen a faster ramp. Engineering moved nearly as fast. During 2023 Sweco built SwecoGPT, rolling the first version out to hundreds of employees in a single afternoon, and by this spring's annual report over 70 percent of the group uses it regularly. Skanska's American arm has run its own chatbot, Sidekick, inside the company's private cloud since December 2023. Ramboll trained RamGPT on its own project archive, reasoning that a firm with thousands of engineers no longer knows what it knows.

The help is real. Ask about what belongs in a nulägesbeskrivning, whether a parking-demand scenario looks reasonable, what a clause in the national annex actually requires, and a chatbot generally answers well; Yesper has an assistant mode for exactly those moments, when the engineer is holding the work and wants a fast, grounded second opinion. Notice, though, what every item on that list is: a part. A definition, a check, a clause, a figure. The answer comes back, and the deliverable has not moved.

It matches what we see at the firms we work with: assistants on every desk, sometimes an internal AI team as well, and the project calendar flat all the same. The measurements say it more precisely. When METR ran a controlled trial in early 2025, experienced developers working in their own codebases were 19 percent slower with an AI assistant, while convinced it had made them about 20 percent faster. Google's DORA report for 2024 found the same shape across thousands of software teams: individual productivity up, delivery throughput slightly down. In the 2025 edition the line finally turned, throughput rising with adoption for the first time, in the year agents arrived that take on whole tasks rather than questions. The help is real at the level of the task. The calendar was never set by how fast anyone could ask a question.

A deliverable is a dozen parts that must agree

The environmental impact assessment for Västlänken, the railway tunnel now being built under central Gothenburg, runs to 177 pages in the version submitted to the land-and-environment court in February 2016: fifteen chapters, covering groundwater diversion, surface water, vibration, construction noise, air quality, mass transport, and a Natura 2000 assessment of the salmon and kingfishers in Säveån. The colophon on page two lists six firms. The lead came from EnviroPlanning, the building antiquarian from Acanthus, the archaeologist from the state history museums, other specialists from Golder, Bergab and Tyréns, and the document itself points onward to the technical description, the hydrogeology report and its own appendices on nature and cultural heritage.

What makes it one deliverable rather than fifteen is that the parts constrain each other. The hydrogeologists identified somewhere between 5,000 and 7,000 buildings around the tunnel standing on clay or on timber foundations sensitive to a falling water table; a cross-disciplinary group of antiquarians, archaeologists, hydrogeologists, geotechnicians and structural engineers narrowed the list to about a thousand, then to some 500 culturally valuable risk objects. A hydrogeologist's map layer decided where the antiquarian walked. When the court approved the assessment in January 2018, its 266-page ruling set leakage limits per stretch of tunnel, five litres a minute past Skansen Lejonet, eighty at Station Haga, required infiltration wells tested and verified before excavation could begin, and added a condition of its own: the water pumped back down to protect the city's timber piles must not carry so much oxygen that it rots them faster instead.

No single chapter of that document is beyond a good assistant; every sub-study has a written method. The difficulty is that the chapters must agree, and keep agreeing while the design moves. Change the drawdown forecast and the risk inventory changes, then the monitoring programme, then the conditions the lawyers argue about. A chatbot repairs the part you point it at, and it has no way of knowing the other fourteen exist.

You hand a colleague the deliverable, not a question

This is the difference the engineer's sentence carries. You do not ask a colleague how to run one of the sub-studies; you hand them the assessment and get it back done. Yesper is the AI civil engineer for construction and infrastructure, and the description is close to literal: it takes the assignment, reads the brief, the site material and the governing documents, works through the parts while keeping them consistent with each other, checks its own output against the requirements, and returns a finished draft.

The responsibility does not move. The engineer reviews the draft and signs it, exactly as they would a colleague's work, and like a colleague, Yesper learns from that review: from what was accepted and what was corrected before signature. From the verdict, never from the client's data. The assistant and the colleague live in the same place, one for the questions, one for the work.

Only one of them changes what a week contains

A chatbot and a colleague are different categories, and a firm can reasonably want both: a fast answer when the engineer is holding the work, finished work when she would rather hand it over. But only one of them changes what a project week contains. A faster way to ask leaves the composition on the engineer's desk; something that holds the whole takes it off, and what moves into the cleared space is the work the title always promised: choosing which alternative to run, which assumption to trust, what to sign.

The first scoreboard, who has an assistant, has been filled in: when over 70 percent of a twenty-two-thousand-person consultancy uses its in-house chatbot regularly, having one no longer separates firms. What will separate them is how much of the signed work they dare hand over whole. At most firms that number is zero today. It is the only one left with room to move.

  1. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025): a controlled trial in which AI tools made experienced developers 19% slower, while they believed themselves about 20% faster.
  2. DORA / Google Cloud, Accelerate State of DevOps 2024: a 25% increase in AI adoption was associated with an estimated 1.5% drop in delivery throughput and a 7.2% drop in delivery stability.
  3. Sweco, Årsredovisning 2025: SwecoGPT launched in 2023; over 70 percent of employees use it regularly, with more than 200 AI assistants automating workflows.
  4. Microsoft, Sweco customer story (2024): the first version of SwecoGPT was developed and rolled out to hundreds of employees in an afternoon.
  5. Construction Dive, Sidekick: Skanska's AI construction chatbot (2024): in use across the US enterprise since December 2023, running in Skanska's private Azure environment.
  6. Reuters, ChatGPT sets record for fastest-growing user base (February 2023): 100 million monthly active users in January per a UBS analysis of Similarweb data.
  7. Trafikverket, Miljökonsekvensbeskrivning, Västlänken och Olskroken planskildhet (TRV 2016/3151, 2016): the 177-page assessment in fifteen chapters.
  8. Mark- och miljööverdomstolen, dom 2018-10-08 i mål M 1808-18, with the land-and-environment court's partial judgment of 31 January 2018: approval of the assessment, leakage limits per tunnel stretch, and the infiltration-water oxygen condition.
Benjamin Glaser Co-founder at Yesper. Writes about AI and the industry that builds the world. benjamin@yesper.ai

The difference from a chatbot only shows when the work actually gets done. Get in touch if you'd like to see Yesper take a whole task, not just answer a question.

Book demo