Nearly everyone uses AI now; very few can point to what it changed in their results. This page maps the distance between those two facts, in four levels ordered by how much of the organization the AI carries: where the market stands in 2026, where you stand, and what the next step is.
The state
Usage stopped being the question a while ago. In Autodesk's AI Pulse, a survey of 2,500 industry leaders across architecture, engineering, construction, manufacturing and media at the start of 2026, 98 percent use at least one AI tool and 84 percent say AI has raised their organization's productivity. Statistics Sweden counted 35 percent of Swedish companies with at least ten employees using AI in 2025, up from 25 the year before, and the EU's digital scoreboard, published this June, has adoption across the union growing by nearly half in a single year. Even construction, which finishes last in most digitalization surveys, is doubling annually: of the five hundred people working in Swedish construction whom Novus polls for the Ramirent report each January, 3 percent used AI in 2023, 8 percent in 2024, then 17, then 28.
The value has not kept up. MIT's NANDA project reviewed over three hundred public AI initiatives in 2025 and put the result in one line: 95 percent of the organizations investing are getting zero return. BCG surveyed 1,250 firms the same autumn and found 5 percent that build capability systematically and capture value at scale, 35 percent that are scaling up and starting to see returns, and 60 percent with next to nothing to show for the spend. McKinsey's annual survey completes the picture: nearly nine in ten organizations now use AI regularly, and about a third have begun scaling beyond pilots.
Beneath the survey numbers sits a pattern we recognize from our own conversations with firms in the sector. At the larger companies, two AI tracks run side by side: a top-down track, where group IT has procured ChatGPT Enterprise or Microsoft's equivalent with access for everyone and a short training, and a bottom-up track, where a team has built its own chatbot on SharePoint that answers questions about contracts and technical specifications. At smaller firms, AI is usually one person's discovery that spreads desk by desk. The MIT report measured how much of this happens off the books: workers at nine in ten companies use personal AI tools on the job, while four in ten companies have bought an official subscription.
None of the gap is explained by the models. The cancelled pilots and the systems carrying real work run on the same models, so the explanation has to sit somewhere else: in how much of the organization the AI carries. Nothing of yours at all, your documents, your systems and workflows, or your way of working with a memory of your projects. That is the scale the four levels below measure.
Overview
Grading an organization's maturity in levels is an idea with its own history, and it was born out of a procurement crisis. Through the 1980s the American military kept buying software that arrived late, over budget or never, and the Air Force asked a newly founded institute at Carnegie Mellon University in Pittsburgh to find a way of telling, in advance, which contractors would deliver. The man who took the job, Watts Humphrey, had spent 27 years at IBM and finished as the executive responsible for four thousand programmers; he arrived at the institute in 1986 with what he later called an outrageous commitment, to fix how the world builds software. His team wrote roughly a hundred questions for grading a supplier's way of working, borrowed a five-level scale from the quality movement, and published the result in 1987 as a maturity framework. Its description of the lowest level could have been written about AI use today: success depends on the competence and heroics of individuals, and cannot be repeated unless the same individuals do the next project.
The framework stuck because the diagnosis travelled beyond software: what separates organizations is how much of the work the organization itself carries, rather than its individuals. Humphrey received the 2003 National Medal of Technology for it, and the maturity models you meet today in everything from information security to BIM trace back to those five levels. The four levels below apply the idea to AI in construction and infrastructure. They are ordered by how much of the organization the AI carries, they mix business development, ways of working and technology, and they are not exclusive: most firms stand with a foot on several. The models themselves are the same on every level; the levels measure everything around them.
The short version first; each level then gets its own section.
AI that helps the individual, ad hoc.
AI that finds and answers.
AI that works in the background.
AI built into how you work and deliver.
Level 1
ChatGPT, Gemini. You open a chat window, type a question, get an answer. It helps the individual, right now: an estimator drafts an email, a project manager summarizes a long meeting note, a bid manager translates a technical description.
Even at enterprise scale the level looks the same. When Skanska's US business rolled out its chatbot Sidekick, in use across the company from December 2023, the point, its data solutions lead explained, was giving people a secure environment to work in; the biggest Nordic consultancies have stood up assistants of the same kind under their own names. Useful from the first week, and still, at bottom, one person asking and one answer coming back.
It is also the level with the largest shadow. The MIT report's sharpest finding was that workers at nine in ten companies use personal AI tools on the job, mostly without being asked to, while only four in ten companies had bought an official subscription. Procured at group level or downloaded onto a private phone, the economics are the same: the individual user's skill is the entire asset, and when the person changes jobs, the capability walks out the door with them.
The MIT report is blunt about why the level stalls:
Chatbots succeed because they're easy to try and flexible, but fail in critical workflows due to lack of memory and customization.
MIT NANDA, The GenAI Divide, 2025
The tools are doing exactly what they are built for: general models, general tasks, individual productivity. What is missing is anywhere for the gains to accumulate. Nothing at this level sticks to the organization, which is why it so rarely reaches the accounts.
Level 2
RAG, short for retrieval-augmented generation, is AI that searches your own documents and answers from what it finds, with a citation down to page and paragraph. This is where AI starts talking with the organization rather than with one person at a time.
A design manager asks what the technical description requires and gets the answer with the page reference. A site manager checks a contract clause without opening the contract. A bid manager pulls up earlier bids on similar jobs while the phone call is still going. Searching that took twenty minutes becomes asking that takes twenty seconds.
The usual implementation is a chatbot wired to SharePoint, OneDrive or Google Drive. Microsoft ships a built-in variant for SharePoint, and Glean, Notion AI and a shelf of similar products belong to the same family; under the hood, the documents are indexed in a vector database and the model answers from the excerpts that match the question.
It is currently the most common form of AI in European construction. When the BIM platform Vitus polled a hundred BIM managers, VDC leads and construction professionals across 21 countries in spring 2026, document copilots were the most widespread application, at 43 percent, and the capability most in demand was not generative design but data validation and quality control. Before this industry wants anything else, it wants to be able to trust its own documents.
How far the searching reaches depends on what is built around the model. A general chatbot finds the words you typed. A construction-specific system knows that a question about ventilation should also pull in VVS, V-drawings, air handling and AMA VVS & Kyl 22 / QJB.12, and widens the search on its own.
The level's boundary sits in the verb. Level two finds, quotes and answers, and the acting stays with you: when the answer arrives, the deliverable it belongs to has not moved an inch off your desk.
Level 3
Agentic AI is the industry's term for AI that carries out assignments instead of answering questions. An orchestrating model breaks the task into parts, calls specialized agents and your systems, runs the parts in parallel or in sequence, and comes back with a result, with nobody clicking between the steps. You stop asking for help and start delegating errands.
The money is arriving now. Gartner expects 40 percent of enterprise applications to carry task-specific agents by the end of 2026, up from under 5 percent a year earlier. In Autodesk's survey, 16 percent of organizations already have agentic AI and another 43 percent plan to be running it within a year. Since late 2024 the systems even share a standard, the Model Context Protocol, for how they connect to the places where data lives.
In construction the assignments are concrete. A tender review that takes in 240 pages and returns a requirements table in five minutes. Work preparations compiled every Monday morning from the weather forecast, the contract terms and previous projects. Change-order identification that reads incoming material against the contract as it arrives, and AMA and PBL reconciliation that runs on every new delivery without anyone asking for it. Ten of these can run in parallel in the background while the teams work on the rest.
It is also the level where most initiatives will die. Gartner expects over 40 percent of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and missing risk controls, and judges that of the thousands of vendors now selling agents, only around 130 sell the real thing; much of the rest is relabelled chatbots and automation scripts. The MIT report locates the mechanism behind the mortality: systems that do not learn the workflow they sit in stay stuck in pilot. A fair share of what is bought as agents this year will be written off, and the write-off will say more about the buying than about the technology.
What surviving looks like is already on the record. Haskoning, the Dutch engineering consultancy, has trained over 4,000 employees and runs more than 200 agents in daily use, supporting everything from tender reviews to contract analysis, according to its director of digital technology in Autodesk's study. The survivors of the cull will share the traits Gartner and MIT keep circling: integrated with the systems where work actually happens, verifying their results, and learning from the outcomes.
Level 4
The fourth level has two load-bearing parts. The first is custom-built capabilities, often called Skills: a specific review, a recurring calculation, a way of working that someone captures once and that then runs hundreds of times across projects, the same way every time, for everyone.
The second is organizational memory. Every new project is indexed as it happens, file by file, contract by contract, parties and phases, so that when an estimator opens a new tender request the relevant history is already active in the context. The archive stops being something you search and becomes something the system arrives already knowing.
The AI also stops waiting to be asked. Work can start on a schedule, every Monday morning; on an event, when a new version of a document is uploaded; or on continuous watch, flagging deviations in incoming material. By the time you sit down at the screen, half the job is often done and waiting for review.
Among firms we work with, this is where the shape of the role changes. An estimator gets decades of previous projects activated in every new tender. A QHSE manager runs one standard across twelve companies on Skills somebody built once. The engineer's day shifts from producing the steps to directing them: setting the assignment, reviewing the result, signing what goes out.
It is also where the returns are heading. BCG measured agents at 17 percent of total AI value in 2025 and expects 29 percent by 2028, and McKinsey's high performers are at least three times more likely than their peers to be scaling agents in a given business function. Behind both curves sits the same plain rule: the value appears where the AI stops being general and starts carrying your organization.
The map
Lay the survey numbers onto the ladder and the market draws itself. Nearly everyone stands on level one. A growing minority, in European construction roughly four in ten, has some form of level two in production. Level three is where this year's investment is pouring in and where next year's cull will fall, and level four is sparse, which is exactly where the value is collecting.
The climbs happen in public now. Skanska did not stop at Sidekick: an assistant that answers from the intranet's own documents followed, then two more built on the firm's risk case studies and safety manuals, four assistants by spring 2025. A level a year, roughly, for a firm that started early.
What makes the ladder expensive to climb is that the levels live in different tools, from different vendors, with different security models and contracts. ChatGPT Enterprise or Microsoft's assistant at the base; a SharePoint chatbot or Glean for the finding; n8n or Zapier with model nodes for the flows; and for the fourth level usually no standard product at all. Each new tool is its own procurement, its own security review and its own silo of context, and the context is the whole cargo: what has to be carried up the ladder is your documents, your systems, your way of working, your project history. Humphrey, who drew the first of these maps, left a warning for everyone using one: "If you don't know where you are a map won't help."
Yesper
Yesper is the AI civil engineer for construction and infrastructure, built to carry all four levels in one place: general AI to ask, search and answers with citations from your own documents, agentic workflows that deliver whole assignments, and Skills and organizational memory that learn how your firm works. Swedish codes and standards are built in; one security model and one contract cover all four levels, and the memory grows the longer you use it.
The result is one staircase instead of four procurements, with a partner for the climb.
Sources
Twenty minutes on a call: you describe where you are, we share a first read. No commitment, and no proposal until we both think there is something worth pursuing.
Book a call