Harvey reached a $15.5 billion valuation in a profession whose work was already written in the form language models read best. Construction’s information lives in programmes, models, drawings, codes and specifications. Until a layer makes those readable to software, and links them to each other, AI in construction will keep working at the edges of the job.

On 9 September, Harvey raised $550 million at a valuation of $15.5 billion.

By Memoori’s count, AI start-ups serving architecture, engineering and construction raised $616 million in the first half of this year. One legal AI round came close to that half-year total.

The easy reading is that investors like lawyers more than builders. A better one starts with where AI has taken hold so far.

Coding came first. Menlo Ventures estimates US enterprises spent $4.0 billion on AI coding tools in 2025, 55% of all spending on AI for specific departments. Legal AI drew $650 million. Construction does not appear in the report.

The pattern is modality. AI took hold first where the work is already text or code, because that is what language models read natively. Contracts and judgments are text. Much of a construction project is something else.

Ask an AI assistant whether the fire doors on a hospital’s third floor comply, and whether they will be in before the floor is handed over. The answer is spread across the door data in the model, the plans, the specification, the fire strategy, the building regulations and the programme. Some of that is prose. None of it points reliably at the rest.

Construction firms are buying anyway. Deltek’s 2026 survey found 78% of North American architecture and engineering firms use generative AI, and only 38% report a measurable business impact. We bought the seats. The seats have not added up.

Anthropic’s data on Claude suggests how shallow that use runs. In its first Economic Index, in February 2025, architects and engineers used Claude more than their share of the US workforce would predict. In March this year, Anthropic published a measure of how much of each occupation’s model-feasible work actually shows up in that use. Computer programmers scored 75% and lawyers 17%. Architects scored 8%, civil engineers under 1%, and drafters and cost estimators zero.

All of these are desk jobs, so the easy explanation that construction work is physical does not apply. AI in construction has started at the edges of the job.

“Harvey for construction” is now a standard pitch. It can copy the assistant and the embedded team. It cannot copy the starting condition: a domain already written in the models’ language, with its cross-references maintained for 150 years.

Construction’s information sits in five forms. Three are not prose. Programmes are networks, building models are geometry and relationships, and drawings are a visual language. The other two, codes and specifications, are text that only works through references to other documents, and on most projects nobody keeps those references in a form software can follow. Law had both pieces before its AI arrived: text, and a maintained record of how its documents link. Domain-specific AI in AEC is waiting for the layer that gives construction the same.

Law was already in the models’ language

In November 2022, Allen & Overy put Harvey’s tool in front of its lawyers. By February 2023, 3,500 of them had asked it around 40,000 questions in a trial. The announcement set one condition: a lawyer reviews every output.

Within months, anyone could rent a model comparable to the one underneath. Harvey’s edge was the fit. Law runs on contracts, judgments, statutes and memos. A model that reads and writes text could work on almost all of it from the first day.

The vendors added people. Harvey’s forward-deployed staff include about 180 former practising lawyers, typically with eight to ten years in practice, its chief product officer said in September. Legora, its Swedish challenger, spent months building inside the law firm Mannheimer Swartling, and its chief executive ran the first 30 customer onboardings himself. CMS piloted Harvey for 12 months with 300 lawyers before rolling it out to 7,200.

Legal AI reached daily use in the world’s largest firms.

Even in law, text was not enough

Then its output reached courtrooms. Damien Charlotin, a legal researcher, has logged 2,184 court decisions worldwide dealing with AI-fabricated material, as of 10 October. Lawyers are involved in 864 of them.

In 2024, Stanford researchers tested leading legal research tools against their vendors’ claims to be free of hallucination. Lexis+ AI produced a hallucinated answer on about 17% of queries, and Westlaw’s AI-Assisted Research on about 33%. The products have changed since, so read those rates as history. The diagnosis has held. The study’s authors concluded that in law, relevance is not a property of text alone.

The passage that reads closest to a question may come from a dissent, or from a case overruled last year. Retrieval finds text. It cannot tell you whether the text still governs.

Lawyers have had help with that since 1873, when Frank Shepard began printing gummed stickers listing the later cases that had cited each judgment. Today’s versions, Shepard’s at LexisNexis and KeyCite at Westlaw, are called citators. Editors read each new judgment within days and code what it did to each case it cites, from followed to overruled. A citator is a cross-reference layer. It turns a library of documents into a map of how they relate.

The leading legal AI vendors converged on it. LexisNexis answers inside Harvey are checked against Shepard’s under an alliance agreed in 2025. Legora had no licence to either publisher’s editorial work. So its staff cut the spines off printed law reports and scanned every judgment, verifying each by double entry. In September it put its own citator into limited beta.

On 29 September, a US federal appeals court affirmed that Westlaw’s headnotes, the editors’ summaries of each case, are protected by copyright. “Headnotes are not law; judicial opinions are,” the court wrote. The judgments are increasingly free. The editorial layer that relates them is the property.

Five forms, and the links between them

Construction has neither advantage in full. Each of its five forms loses something when a model reads it as plain text.

Programmes. nPlan, which has trained on more than 750,000 of them, puts it this way: “A construction schedule isn’t a document, it’s a graph.” Reduce one to a list of activities and dates, and a date held by a constraint can look the same as one driven by logic. A 2026 benchmark of general graph problems found reasoning models falling below 50% accuracy once a graph passed about 120 nodes. One study of 180 real programmes counted more than 320,000 completed activities between them, close to 1,800 on average. In a 2025 study on Primavera P6 data from two residential projects, GPT-4o-mini, given no examples, extracted the right construction information about half the time. A small model trained on parsed schedule tables reached 84%.

Building models. An IFC wall’s position is defined relative to a chain of other placements, and its properties hang off separate relationship objects. A door’s fire rating may sit in a standard property, a custom one or only the type name, depending on the office template. A May 2026 benchmark asked 1,027 questions of 37 real IFC models. Single-pass querying answered 19% correctly and declined to answer more than three quarters of the geometric questions. An agent that worked through the model’s structure step by step reached 56%.

Drawings. On a fire strategy plan, each door’s rating is often a tag placed beside it. Extracted text keeps the tag and loses the door. DrawingVQA, a small benchmark of 33 issued-for-construction structural drawings published this year, found Gemini 2.5 Pro answering 56% of expert-level questions correctly, against 93% for experienced professionals. AEC-Bench, published this year by a vendor in this market, found retrieval was frequently the bottleneck, and most agent runs reached for a plain-text extraction tool. AECV-Bench, a benchmark my company published, found a related split on separate tests: up to 95% on text and document questions, and at best 39% on counting the doors on a floor plan exactly. My company, AECFoundry, builds the kind of layer I am describing.

Codes and regulations. Codes are text, and they work the way law does, through references and local amendments. Law has editors keeping those links current. For codes the record is patchier: not all jurisdictions even report which edition they have adopted. As of April, US states sat on five different editions of the International Building Code, each with its own local amendments. The Eurocodes leave 1,573 parameters to be set country by country. In England, the Approved Documents are guidance, and the law sits in Schedule 1 of the Building Regulations; Scotland, Wales and Northern Ireland set their own. In a 2026 study, AI agents turning code rules into checks that run on building models scored 0.40 out of 1 on unseen models, on a measure that penalises both missed and false results, and 0.21 on the most complex rules.

Specifications and standards. A specification section cites standards, and those standards cite others. The International Code Council notes that when a code cites a standard, that standard’s own references carry legal effect within the scope of the code’s reference. The 2021 International Mechanical Code alone lists more than 255 standards from 25 organisations. A common break is a substitution. The section names a product “or approved equivalent”, the contractor proposes another, and the approval sits in a submittal response while the section still names the first.

The harder problem sits between the forms. An activity in the programme installs elements in the model. Those elements appear on drawings, are specified in a section and must meet a code clause. When the documents disagree, the contract decides which one governs, and the answer changes from contract to contract. On US federal work the specification governs the drawing. Under NEC4 the project manager resolves each inconsistency. After award, an RFI response, an instruction, a revised sheet or a variation can move any of these links.

In law, the citator holds the links. On a project, people do. Linking a programme to a model for 4D planning still “requires considerable manual effort”, a 2022 study found. In NBS’s 2023 survey, 56% of respondents exchanged information in IFC and 36% in COBie, the formats meant to carry some of these links. The rest live in trackers and registers, or in the heads of the planner and the design manager. At handover, the documents are archived. The links between them leave with the team.

What scale will not bring

The next model will read some of this. Perception is improving fast, and spatial reasoning is still its weak point: on one test of spatial reasoning across several images, the best model scored 49% against 97% for people. Anthropic’s own documentation says Claude’s location outputs are approximate and its counts may be imprecise.

Representation helps today. Turning piping and instrumentation diagrams into graphs raised accuracy by 18% over images, and cut token costs by 85% against the raw files. Letting an agent work through an IFC model’s structure nearly tripled single-pass accuracy. Scale has absorbed gains like these before, and may absorb some of these. A firm deploying AI this year cannot wait to find out.

Scale cannot learn what was decided after training. Which response changed which detail, or which edition of a standard this contract invokes, is settled on the project. An agent can try to reconstruct those links on every query, at a cost each time. Each attempt leaves no record of who decided what.

Why no platform can hold the links

If construction only needed a citator, it could plug into one, as Harvey has through its LexisNexis alliance. It cannot.

In law, the judgments are public and two publishers dominate the editorial layer. Construction is close to the reverse. In the US, even the codes are opening up: in April, a federal appeals court held that a for-profit platform publishing building standards incorporated into law is likely making fair use. The scarce material is each project’s own information, written by several parties and held in several systems. The programme sits in one tool, the model in another, the drawings and RFIs in a common data environment, and the specification somewhere else again. On most large projects no single system holds all of it.

Access to it is now contested. Procore’s policy for developers building on its platform, effective 30 September 2025, bars them from using its API data to train or benchmark AI models, and from bulk-exporting it without Procore’s express consent. That month Trunk Tools lost its API access; Procore said the changes protect the integrity and security of its customers’ data. Publishers set terms too. NEC’s publisher bars indexing its contracts into AI systems without consent, and the ICC’s terms bar data mining of its codes.

The platforms are building agents over the form each one holds. Procore bought Datagrid in January and released five agents on it in May, and Trimble agreed to buy Document Crunch in April. Autodesk has announced a drawing compliance review for early next year. Terms differ. Autodesk revised its terms in December 2025 so that customers may train on their own data, though not to replicate Autodesk’s products. Bentley lets users decide whether their data trains its AI.

A platform can make its own form readable. The links run across all of them. Platform renewals now settle, in the small print, who can read which part of the project, and on what terms.

Law shows where the value goes when someone else keeps the links. Thomson Reuters’ legal business grew about 10% organically in the second quarter, growth the company credits mainly to Westlaw and its AI assistant. LexisNexis reported underlying revenue up about 10% in the first half. At least one of Harvey’s own litigation workflows requires a LexisNexis subscription. In law, whoever keeps the cross-references gets paid by almost everyone who builds on them.

What transfers

Build the layer above the platforms, over information the firm already has the right to hold and export. Each form keeps its native structure: the programme stays a network, the model keeps its geometry, the drawing keeps each symbol in place and the clause keeps its exceptions. Then link them, and make every answer cite the sheet, clause, activity or element it rests on, at the revision it used. Any agent, including a platform’s own, can then be checked against it. Part of the duty already exists. In England, the golden thread rules for higher-risk buildings require an accurate, up-to-date record of who changed what and when.

This has been tried. BIM and the common data environments of ISO 19650 both promised linked information, and asked people to key the links in by hand, often with no line in the fee for it. Software can now propose links from the documents teams already produce, for the people who made the decisions to confirm. Nobody has yet published how well that works at scale.

Keep it in the firm. Kirkland & Ellis has committed $500 million over three to four years to its own AI platform, according to Bloomberg Law. Few AEC firms can fund anything like it. A seat licence keeps nothing in the firm. Whatever route a firm takes, the contract should say the firm owns the links and can take them to another platform or another provider, including away from whoever built them.

Leave the signature with the professional. When the regulator for solicitors in England and Wales authorised the first AI-driven law firm, Garfield.Law, in 2025, it attached conditions. The system may not propose case law, and named solicitors stay accountable for every output. Under most construction contracts, a named person decides what the documents mean. The layer should hold those decisions and surface the conflicts still waiting for one. It should never make them.

The liability is not moving either. ENR reported on 9 October that AI exclusions have not yet appeared on design-firm liability policies. ASCE’s policy says AI cannot be held accountable. The engineer who signs off the calculations still is.

Harvey’s legal engineers are lawyers first. The construction equivalent is planners, BIM managers, design managers and engineers who know what each form means, working inside the firm until its own people can keep the layer. In July, Trunk Tools advertised for a manager to lead five to eight experts embedded in one top-20 general contractor, tapering to part-time support as the client’s people take over. Arcadis invested in Nomic after a six-month trial with about 150 engineers in 12 countries.

Start where the links cost most: one live project, and one recurring question that crosses at least two forms, such as whether what is being installed matches what was specified and approved.

A citator codes a case once for every lawyer who later relies on it. A project’s links serve one team. The structures do repeat. IFC is an international standard. Every P6 programme shares one table structure. MasterFormat and Uniclass give specifications and model objects shared codes. Capture that grammar once and the per-project work should taper. The tenth project should then cost less than the first, because what was learned on nine, and every correction the firm’s engineers made, carries forward. No construction firm has yet published that comparison.

The rest of the job

The programme, the model, the drawings and the clauses that carry the liability are written in languages the models do not yet read together.

Until a firm makes those forms readable and writes the links down, its AI will keep getting better at the part of the job that was already text.