AEC AI has spent the last three years chasing point automations: a smarter takeoff here, a faster RFI draft there. The category that is actually compounding — and where 2026 capital and M&A are landing — sits underneath all of them.

Drawings, specifications, submittals, RFIs, and contracts are the industry's load-bearing data. Whoever turns them into durable, machine-usable structure wins the layer every workflow above depends on. The market is now pricing that thesis in. In the first four months of 2026 alone, Trimble announced it was acquiring Document Crunch; AECOM closed its $390M acquisition of Consigli; Neuron Factory took strategic capital from Trimble, Suffolk Technologies, Zacua and Imad Ventures; Primepoint closed a $10M seed; and Brickanta raised $8M from Northzone to put pre-construction decisions on structured footing. Zoom out, and six contech startups raised $126M in early 2026. A striking share of it is pointed at documents — not jobsites.

This post walks through why, what AEC-grade document intelligence actually requires, which companies are building it, and where we believe the moats will form.

Why documents became the wedge

The industry runs on documents the way banks run on ledgers. A single project produces tens of thousands of artefacts — drawing sets, specifications, submittals, RFIs, meeting minutes, O&M manuals, closeout binders — and each one encodes scope, risk, obligations, and the reasoning behind decisions. They almost never travel cleanly across phases or stakeholders.

The cost of that friction is measurable. PlanGrid and FMI found that construction professionals spend 14.1 hours per week on non-optimal activities tied to project information — 5.5 hours searching, 4.7 hours on conflict resolution, 3.9 on mistakes and rework — and estimated total U.S. labour cost at roughly $177.5B annually. McKinsey places the macro outcome in delivery terms: large projects typically run 20 percent longer and can finish up to 80 percent over budget, and most of that gap is upstream interpretation and coordination, not field execution.

A workforce cliff adds urgency. NCCER projects 41 percent of the construction workforce will retire by 2031, taking decades of institutional knowledge with them. The people who know where the answer lives in a 400-page spec are leaving faster than they can be replaced.

ISO 19650 has been standardising the governance discipline around information management in built-asset delivery. What it has not done — and could not do — is make documents machine-readable at scale. That is the gap document intelligence is now walking into. Generic Intelligent Document Processing has existed for a decade — OCR, classification, field extraction, export to systems of record. What is new is that the same stack, retooled for drawings and specs, is the fastest route to banked ROI in a conservative industry. Submittal review compresses from weeks to hours. RFIs auto-cite the right sheet and spec section. Contract obligations become a searchable register. The workflows do not change; the documents become legible.

That is why the capital is landing here.

The 2026 signal

The M&A and funding tape points in one direction.

Trimble acquires Document Crunch (April 2026). Trimble is folding a construction-specific contract AI into its Construction One delivery ecosystem. Software incumbents are acquiring document-intelligence platforms, not building parity in-house.

AECOM acquires Consigli ($390M, late 2025). An engineering firm bought an AI company rather than licensing from one. Consigli's "autonomous engineer" platform — MEP load calculations, clash-free BIM, tender document preparation with integrated risk analysis — is now in-house IP at one of the largest ENR firms. This inverts the traditional AEC software relationship and signals that proprietary AI may become a margin differentiator rather than a line-item expense.

Neuron Factory's strategic round (March 2026). The new cap table — Trimble, Suffolk Technologies, Zacua Ventures, Imad Ventures, and existing investor Colle Capital — reads as an AEC-industrial coalition. Capital is flowing into a "construction-focused knowledge graph" pairing GraphRAG with vector search across multimodal inputs.

Primepoint closes a $10M seed (April 14, 2026). A drawing-first platform that links every drawing element back to schedules, specs, and source documents — already funded to scale.

Brickanta raises $8M seed (January 2026). Northzone-led. Pre-construction bid and procurement documents as the wedge; RFP bundles in 15 minutes versus days.

TrunkTools' $40M Series B (mid-2025). Insight Partners-led, $70M total raised. Q&A over project documents, submittal review, and revision review — the three highest-volume document loops on any project.

The pattern is clear. The winners are platforms that turn unstructured AEC documents into structured, queryable, traceable objects — not tools that only automate a single human's Tuesday afternoon. And strategic buyers are paying real money to own that layer rather than integrate around it.

Adoption is the first step — the knowledge layer is what compounds

One of the most encouraging enterprise signals of the past year is Turner Construction's commitment to AI literacy at scale. The largest U.S. contractor by revenue signed a two-year, company-wide "wall-to-wall" agreement with OpenAI, giving all 11,000 employees access to ChatGPT Enterprise. The results are real: more than 400 custom internal AI applications and a reported 70,000 hours of annual productivity gains. That kind of organisational commitment — getting an entire workforce comfortable with AI as a daily tool — is a meaningful foundation.

The question is what comes next. General-purpose models are powerful interfaces. They help teams summarise emails, draft memos, prepare meeting notes, and rephrase content faster. That value is genuine. But the AEC-specific workflows where the largest cost and risk sit — submittals, RFIs, takeoffs, contract compliance, closeout — require more than a conversational layer. They require AI-readable data: parsed drawing sets that preserve cross-sheet structure, construction ontologies that encode the relationships between tags, schedules, specs, and submittals, evaluation frameworks that measure accuracy on domain tasks, and governed integrations with the CDEs where project data lives.

This applies whether the starting point is a simple custom GPT summarising a spec section or a complex end-to-end agentic workflow that drafts an RFI, cites the right sheet, and routes it for review. In both cases, the model provides the reasoning capability. The domain knowledge layer — structured, AEC-specific, grounded in project documents — is what turns that capability into compounding value.

Turner's move, and the broader wave of enterprise LLM rollouts across AEC, should be read as the industry crossing the adoption threshold. The next step is solving the knowledge layer underneath: turning documents from files people search through into structured intelligence machines can reason over.

What "AEC-grade" document intelligence actually requires

Generic IDP breaks on contact with AEC. A drawing sheet is not a document — it is a visual data structure with callouts, legends, schedules, cross-sheet references, title blocks, and revision clouds. Flatten it to text and you lose what matters most: the references that tie door tag A-101 to a schedule on A-601 to a spec section in Division 08. AWS's technical write-up of TwinKnowledge's architecture makes this explicit: drawing sets need a dedicated computer-vision pipeline to extract sheet metadata, page boundaries, and view/detail/schedule titles before retrieval even begins.

Five capabilities define AEC-grade in 2026.

1. Multimodal parsing as foundation

CV models that understand plan structure, not just OCR'd text. Nomic's AEC-Bench frames the problem directly: agents fail when tools flatten spatial structure, and cross-sheet reasoning is a core evaluation axis. The benchmark was released open source on April 2, 2026 with 196 tasks across 9 families, covering single-sheet, cross-sheet, and cross-document reasoning. AEC Foundry's own AECV-Bench takes a complementary approach, evaluating extraction and understanding across 23 document categories drawn from production environments — including adversarial samples from real-world failures. The emergence of multiple shared evaluation harnesses is a quiet but significant milestone — progress in this category can now be measured rather than claimed.

2. Graph layers on top of vector search

Every serious player — Neuron Factory, Primepoint, TrunkTools — is pairing semantic retrieval with explicit graph structure. Vectors surface the right chunk; a construction knowledge graph encodes the relationships (tag → schedule → spec → submittal → RFI → revision) that actually carry meaning.

3. Verifiable, cited answers

In AEC, an AI answer without a source is a future claim. Provision and TrunkText market cited accuracy numbers. Primepoint's "Ask Marvin" grounds every response in a specific drawing detail or spec section. Grounding has become the category's defining capability; hallucination-by-default is disqualifying.

4. Evaluation, not demos

TrunkTools' engineering team has published on end-to-end quality monitoring and retrieval ranking; Neuron Factory publishes an internal Eval Framework. With AEC-Bench and AECV-Bench now available as shared infrastructure, 2026 is the year procurement teams can ask any vendor "what did you score?" and expect an answer.

5. Governance enterprises can actually approve

Row-level access control, AES-256 encryption, SOC 2 — Howie and ContraVault AI lead with these. Document intelligence touches contracts, pricing, and legal exposure; governance is not optional.

This is the shape of a durable stack. The companies attracting capital and M&A interest have most of it — and are moving toward the rest.

The 2026 market map

We organise the landscape by job-to-be-done rather than by logo.

Drawing intelligence — parse, link, compare: Primepoint, TwinKnowledge, Nomic, Kreo (with Caddie AI), Togal.AI, Drawer AI, and Civils.ai. The foundation layer — without reliable sheet parsing, everything above is a demo.

Submittal and RFI loops: The highest-volume, highest-frequency documents on any project. TrunkTools (TrunkSubmittal, TrunkReview), BuildSync, Remy, Nonlinear, iFieldSmart, Submittal.app, Diesl, Nicky AI, Cassidy (RFI agent), and InspectMind AI. The incumbent in this segment is Autodesk Pype AutoSpecs, now tied into Autodesk Build inside ACC.

Contracts, risk, and compliance: Document Crunch (now Trimble) and ContraVault AI for tender and RFP risk, with a growing body of academic work on automated compliance checking across the lifecycle. This is where legal and finance functions first meet AEC AI.

Preconstruction — the economic hinge: Neuron Factory, Brickanta, Provision (pre-bid scope intelligence and cited Q&A over drawings and specs), and portions of Consigli (now AECOM) cluster here. Risk is priced and scope is locked before mobilisation, and document intelligence compounds across every downstream workflow.

Platform and ontology: Procore is embedding Copilot and agents into its platform (RFI Creation Agent, Daily Log Agent, Agent Studio). Autodesk Construction IQ is the other incumbent play. At the enterprise extreme, Palantir Foundry's construction ontology and the new Palantir + The Nuclear Company NOS platform demonstrate what happens when digital twin and document intelligence are pushed to their operational limit.

Adjacencies to watch

Karmen converting documents into schedules; Howie centralising project data into a searchable knowledge base.

The useful 2x2 for this category is not "point tool versus platform." It is (document breadth) x (verifiability). The interesting zone is the top-right: broad document coverage, every answer cited to source.

Where the moats actually form

Five patterns define durable advantage in AEC document intelligence.

Proprietary drawing parsing is a short-term moat

Generic vision-language models do not yet read a plan set well. Teams that have built CV pipelines tuned on real drawing sets, with hand-labelled evaluation data, have banked a head start. AEC-Bench will compress that lead — but not for the next 12 to 18 months.

The ontology is the long-term lock-in

Vector search will commoditise. A construction knowledge graph — tags, schedules, specs, submittals, RFIs, revisions, all linked — becomes customer-specific by the time it is useful. Primepoint, Neuron Factory, and Palantir Foundry are playing different versions of the same game.

CDE integrations are table stakes but non-trivial

Live, governed two-way sync with Autodesk Construction Cloud, Procore, Bentley, Egnyte, and SharePoint is where deployments succeed or stall. TrunkTools' Egnyte integration via the Workato SDK is a useful proof point.

Evaluation becomes a commercial asset

In a category where wrong answers create claims, the team with the best internal eval harness ships safer product faster. Expect eval scores to enter RFPs this year.

The customer-owned corpus is the long game

Every project a customer runs through a platform enriches their own graph — and moves them closer to an institutional brain that outlives retirements. This, more than any productivity metric, is the substantive answer to the 41-percent workforce cliff.

An open question frames the next twelve months: does the AI layer get owned by the software vendors (Trimble, Autodesk, Procore) or by the engineering and construction firms themselves (AECOM/Consigli, and whoever follows)? If AECOM's $390M move proves to be a one-off, incumbents consolidate. If a second major firm moves in kind, the balance of power in AEC software shifts structurally.

What to watch

Four signals will determine how this category matures.

Will a second top-10 ENR firm buy an AI startup? A one-off keeps incumbents in control; a second deal opens a new playbook.

Will Procore's Agent Studio open a genuine developer ecosystem — or remain a feature? Agent Studio has the platform position, but ecosystem effects require open access.

How quickly will AEC-Bench scores climb — and will vendors publish them? Transparency separates the companies with real capability from the ones with good demos.

Where will ISO 19650 governance meet grounded AI? Owner-operators are the next buyer segment, and they care about information management as compliance, not just productivity.

The industry is past the question of whether AI will land in AEC. The question now is which layer wins. On the current evidence, that layer is documents. Everything else sits on top.

Where AECFoundry sits in this picture

We build the domain layer underneath. Our work focuses on turning AEC-specific documents — drawings, specifications, submittals, RFIs, contracts — into structured, machine-usable intelligence, evaluated against real-world AEC complexity rather than sanitised samples.

If you are evaluating vendors, piloting an internal capability, or deciding whether to build, buy, or partner, we would welcome the conversation.

Talk to us | Explore AECV-Bench