Every conference in the last two years has had an "AI in construction" keynote. Most of them are 60% hype, 30% pilot programs that didn't scale, and 10% genuinely useful working examples. The ratio is improving, but the industry is still in the "everyone's talking about it, fewer are using it" phase.
This article skips the hype and focuses on where AI is actually earning its keep in construction today. What's working, what isn't, and how to pilot it on a live project without breaking workflows that already function. It's written for working project managers, estimators, site leads and builders — not for conference-keynote audiences.
The state of AI in construction, honestly
Two broad categories of AI are relevant to construction work.
Multimodal large language models (Claude, GPT, Gemini) — the general-purpose models that can read drawings, answer questions, summarise documents, draft emails, and so on. These are where most of the recent progress has come from. The step change over the last 18 months is that they can now look at a rasterised drawing and reason about what they see, not just read the text layer.
Specialist construction ML tools — narrower systems trained on specific construction data. Object detection for takeoffs, computer vision on site cameras for progress and safety monitoring, predictive models for scheduling or cost overruns. These have been around longer but are often more brittle outside the exact scope they were trained on.
Most working AI in construction today combines the two: a specialist layer for detection/classification, a general-purpose LLM for reasoning and user interaction.
Seven jobs where AI is already working
1. Drawing revision comparison
Arguably the most mature AI workflow in construction right now. Vision models can look at old and new revisions of the same sheet and produce a structured summary of what changed — and, critically, categorise changes (design-impacting vs title-block-only) so the team can triage. On big consultant packages this turns a five-hour review into a fifteen-minute spot-check.
This is working today, it's cost-effective (fractions of a dollar per drawing), and the accuracy is high enough to trust as a first pass. It's still a first pass — a senior reviewer signs off — but the time savings are real.
2. Drawing-aware project search
Natural language Q&A over your project data. "Find all overdue RFIs." "What did the consultant change in Rev C of the hydraulic set?" "Show me the approval status of the lobby details." Grounded against your actual drawings and annotations, these queries now return answers in seconds.
The keyword is "grounded". A generic chatbot without project context will guess. A properly-built project assistant will cite the specific drawing or annotation its answer came from, and should refuse to answer when it can't find relevant data.
3. Automatic drawing classification on import
When a consultant sends a bundle of 200 PDFs with drawing numbers hidden in title blocks, the fiddliest part of the upload is keying in numbers, titles and disciplines. Vision models now do this well enough to pre-fill 70–90% of the fields, leaving you to review rather than transcribe.
4. RFI drafting and summarisation
Not replacing the RFI process — accelerating it. AI is good at turning three bullet points into a properly-structured RFI with the right context, and at summarising long RFI threads into a decision log. Still needs a human approver, but saves meaningful time.
5. AI takeoffs (in specific scopes)
Dedicated tools (Togal, Kreo, Stack) now do credible first-pass takeoffs for scopes where symbols are standard and drawings are clean. Fixture counts, door/window schedules, straightforward linear measurements — real time savings here. Complex services drawings and non-standard symbology still struggle. See the AI takeoffs article for a fuller take.
6. Site photo analysis and progress tracking
Computer vision on site imagery to monitor progress, detect safety violations (missing PPE, unsafe scaffold), and match built work to the drawings. Works well on repeat scopes (high-rise floor plates, modular builds), weaker on bespoke work.
7. Schedule risk prediction
ML models trained on historical project data to flag tasks that are likely to run late. Useful on large, data-rich projects with thousands of similar historical tasks. Less useful on smaller or more bespoke jobs where the model can't generalise from few examples.
Seven jobs where AI isn't ready (yet)
It's as important to be clear about this as the positive list, because overreach is what burned a lot of early AI pilots.
- Anything safety-critical that stands alone. Egress compliance, structural adequacy, fire sizing — AI can assist a qualified engineer but cannot replace them.
- Contract drafting and legal language. Generative models produce plausible but often wrong contract clauses. Don't use them as primary drafting.
- Cost estimating at final pricing stage. Takeoffs for bid/no-bid decisions, fine. Final tendering without a human estimator reviewing line by line, not yet.
- Regulatory interpretation. Building codes, standards interpretation — the LLM will give you a confident answer that's subtly wrong. Use it for drafting, not for decisions.
- Negotiating with people. Claims, variations, disputes — these are human conversations that depend on relationship and context AI doesn't have.
- Generating construction drawings from scratch. Design generation is active research. Generating a buildable set of plans from a brief isn't a product yet; treat anything marketed that way with scepticism.
- Autonomous site robotics. Autonomous layout, autonomous demolition, autonomous bricklaying — real research, not mainstream production. A few niche products are emerging; most are pilots.
How to introduce AI without breaking your workflow
Teams that succeed tend to share a few habits.
Start with one workflow. Pick the single highest-pain, highest-frequency task and pilot AI there. Usually that's either revision comparison (if you deal with lots of drawings) or RFI management (if you deal with lots of questions).
Keep humans in the approval loop. For anything that affects cost, compliance or contract, the AI's output is an input to a human decision, not the decision itself.
Measure cycle time, not accuracy. Accuracy debates are endless. The honest question is "did this AI tool reduce the time to complete workflow X?" If yes, keep it. If no, bin it.
Pay attention to data governance. Which model is being used? Where is your data processed? Does the vendor retain it? Does it train on it? For most commercial work, cloud-hosted models are fine. For secure or government work, those questions matter a lot.
Don't build on the cutting edge of the cutting edge. The tool that just shipped last week isn't the one to bet a project on. Wait for at least a couple of months of real-world use before committing a live job.
How Doclio uses AI
We're opinionated about where AI fits and where it doesn't. Two places in Doclio use AI today.
Drawing revision comparison. When you upload a new revision, Doclio runs a Claude Sonnet vision model over the old and new sheets in the background and pre-computes a structured summary of what changed. Results are categorised as "design changes" versus "title block only" so you can triage quickly. Because it's cached at import time, the comparison is instantly available when you open the drawing.
Drawing-aware AI assistant. A chat assistant that answers questions about your project using your actual drawings, annotations, and reports. Grounded in your project data with a Supabase-authenticated scope, it cites back to the specific drawing or annotation it used. When it can't find the answer in your project or our knowledge base, it can optionally search the web (via Tavily) and cite sources. Built on Claude Haiku 4.5 for text queries.
Both run on a monthly credit system so you're not paying for AI you don't use.
What we don't claim to do:
- Generative design.
- Autonomous site operations.
- Financial forecasting.
We focus AI on the two places it reliably pays back — reading drawings and answering project questions — and leave the rest to specialist tools.
What's next
The interesting frontier over the next 12–18 months is in workflow automation: AI not just answering questions, but taking actions on your behalf. "Draft an RFI about this detail and send it to the architect." "Update all open defects on Level 2 and generate a punch report." "Compare this new revision and notify the trades whose scope is affected."
The underlying model capability is already there; the constraint is product design — making sure AI actions are transparent, reviewable and reversible. That's the next wave, and it's close enough that teams should be building the muscle now, not in two years.
---
Try Doclio's AI features on a free project — drawing revision comparison and the AI assistant are included on every plan, starting from Starter.