Table of Content
What works today with AI in construction?
The AI that works today in construction reads and explains data you already have: drafting report commentary from job cost and WIP numbers, answering questions about costs and contracts with the source shown, suggesting cost codes on invoices, flagging unusual costs and giving forecast ranges. It works when the ERP data underneath is clean and a person approves anything that touches money, contracts or safety. It does not yet work as an autopilot that posts entries, approves change orders or sets prices.
TL;DR
- Most AI in construction content is about robots, drones and generative design. The faster return for most contractors sits in the data they already own: job cost, committed cost, billings, timecards and contracts in the ERP.
- This guide ranks ten uses by readiness: works today, works after data cleanup, pilot first, or avoid for now. Each row names the data it needs and the human check that must stay.
- Data readiness decides the outcome. In our example audit of an ERP, only two of six checks pass, and the guide shows which AI uses are safe to start anyway.
- Five worked examples use computed numbers: a query that double-counts cost, a draft report that invents a number, an anomaly flag on unit cost, a forecast range built from 12 finished jobs and invoice coding suggestions by confidence band.
- Set permissions by level: read, draft, propose, post. AI stays at the first three until you have months of evidence.
- All figures in the worked examples are illustrations built for this guide, not client results. Named industry sources are attributed and dated.
- CommercePundit connects AI to ERP data for contractors. See the last section.
Which AI uses in construction are ready, and which are not?
Status here means readiness on top of your own ERP data, judged by how checkable the output is. “Works today” does not mean no review. It means the output is easy to verify and a mistake is cheap to catch.
| Use | ERP data it needs | Status | The human check that stays | What goes wrong |
|---|---|---|---|---|
| Plain-English questions about job cost and WIP | Governed views of jobs, cost codes, budgets, commitments | Works today, on approved views | Saved, tested queries for key questions; row counts on every answer | Joins that double-count; wrong period or entity |
| Draft commentary for WIP, job cost and progress reports | The report numbers, prior period, notes | Works today | Project manager reads and approves; numbers checked against the source | Invented or misquoted numbers; confident wrong reasons |
| Contract and spec questions with quoted source | Contracts, specs, addenda as searchable text | Works today with citations | Reader opens the cited clause | Missed clauses; a quote taken out of context |
| Cost code suggestions on invoices and receipts | Past coded invoices, vendor list, cost code list | Works today with confidence bands | Review below the confidence line; spot-check above it | Vendor-specific habits learned wrongly; duplicate invoices |
| Anomaly flags on cost, hours and unit rates | Several jobs of history by cost code | Works after data cleanup | Estimator or PM investigates each flag | Noise from bad codes; flags ignored when too many |
| Margin and cost-to-complete forecast ranges | Finished-job history with percent complete and fade | Pilot first | PM owns the number; AI gives a range, not a figure | Small history; changed mix of work |
| Estimating from past bids and unit costs | Bid history, cost history, quantities | Works after data cleanup (see the estimating guide) | Estimator sets the price | Old costs applied to a new market |
| Equipment maintenance prediction | Maintenance history, hours, sensor or telematics data | Pilot first | Mechanic decides | Too little failure data |
| Posting entries or approving payments on its own | Everything | Avoid for now | Not applicable | Errors become ledger entries |
| Approving change orders or setting bid prices on its own | Contracts, pricing, risk | Avoid for now | Not applicable | Legal and commercial exposure with no accountable person |
Our broader guides cover the building blocks: construction automation without replacing your ERP, AI agents in ERP and AI workflow automation tools. This guide is about choosing which uses to start with, and proving the data is ready.
Why start AI for construction with ERP data and not robots or drones?
Robots, drones and computer vision are real, but they depend on equipment purchases, site conditions and field adoption. The office works on structured data with clear definitions: cost codes, budgets, commitments, timecards, billings. Mistakes there are traceable to a number. That makes it the cheapest place to learn what AI can and cannot do for your company.
Adoption data supports caution about readiness. In a survey of more than 1,000 technology decision-makers at architecture, engineering and construction firms across five markets, run in July 2025 and published by Bluebeam on October 28, 2025, 27% of firms said they use AI, and the top integration challenges were data sharing security (42%) and cost and complexity (33%). Bluebeam sells software, so treat this as a signal of where the market is, not a benchmark for your results. Associated Builders and Contractors published a free AI resource guide from Associated Builders and Contractors in February 2024 that lists more than 20 use cases across preconstruction, construction and building maintenance, and includes best practices for writing an office AI policy. Both point the same way: use cases are plentiful, and policy and data come first.
Is your ERP data ready? A worked audit
Run six checks before you point any AI at job cost. The counts below are made up for illustration, and the thresholds are suggestions to adjust to your own risk. They are strict on purpose: an AI that reads bad codes will explain them very convincingly.
| Check | Passed | Rate | Suggested threshold | Result |
|---|---|---|---|---|
| Cost lines with a valid, active cost code | 9,620 of 10,000 | 96.2% | 98% | Fail |
| Cost lines posted to an open job | 9,940 of 10,000 | 99.4% | 99% | Pass |
| Vendor records that are not duplicates | 1,164 of 1,200 | 97.0% | 98% | Fail |
| Commitments with a current change order status | 412 of 480 | 85.8% | 95% | Fail |
| Timecards with job, phase and equipment fields complete | 7,310 of 7,500 | 97.5% | 97% | Pass |
| Jobs with a recorded percent-complete method | 38 of 40 | 95.0% | 100% | Fail |
Two of six pass. What that means in practice: do not start with forecasting or anomaly flags, which depend on clean history. You can start with contract and spec questions, which use documents and not the ledger, and with report commentary on the numbers you already trust. In parallel, fix cost codes, duplicate vendors and change order status, because every later use depends on them. Our guides to construction job costing and the construction WIP report cover the definitions that should be settled first.
Example one: a plain-English question that doubles your cost
You ask, “What is the total cost to date on job 2418?” The assistant writes a query. If it joins the cost table to the change order table to show change order status, each cost line repeats once per change order. The numbers below are invented.
| Step | Result |
|---|---|
| Actual cost to date on job 2418 in the cost ledger | $184,200 |
| Change orders on the job | 2 |
| Answer from the naive join (each cost line repeated 2 times) | $368,400 |
| Correct answer after summing cost first, then joining | $184,200 |
The wrong answer looks professional. Three controls catch it: give the assistant approved views that already hold the right joins, save and test the queries behind your top fifty questions, and show the row count and the source view with every answer. Anything that does not tie to the cost ledger total gets rejected automatically.
Example two: report commentary that invents a number
Drafting commentary is the safest first use, because every number can be checked against the report. Feed the assistant only the report figures and ask it to use no others. Then check each number in the draft against the input.
| Sentence in the draft | Numbers used | All numbers in the input? | Action |
|---|---|---|---|
| Margin fell from 11.0% to 9.8% this month. | 11.0%, 9.8% | Yes | Keep |
| Cost to date is $1,845,000 at 60.8% complete. | $1,845,000, 60.8% | Yes | Keep |
| Rock excavation added about $212,000 of unplanned cost. | $212,000 | No | Reject until a person supplies the source |
The third sentence may even be true, but nothing in the input supports it, so it is blocked. A simple script that extracts numbers from the draft and compares them to the source catches this class of error, and it costs far less than the embarrassment of a wrong figure in an owner report. Our guide to construction reporting shows which reports to automate and where commentary belongs.
Example three: flagging an unusual unit cost
Not every useful AI is a chat assistant. A statistical check on unit cost is cheaper, more explainable and often more reliable. The history below is eight finished jobs for the same cost code, in dollars per linear foot of pipe. The values are invented.
| Measure | Value |
|---|---|
| History (8 jobs, dollars per LF) | 41, 44, 39, 47, 43, 46, 40, 45 |
| Median | 43.5 |
| Median absolute deviation (MAD) | 2.5 |
| New job at 61.5 | modified z-score 4.86: flag |
| New job at 48.0 | modified z-score 1.21: no flag |
The modified z-score is 0.6745 times the distance from the median, divided by the MAD, and a common rule flags values above 3.5. The job at 61.5 is 41% above the median and gets flagged for an estimator to investigate. The point is not the formula. It is that a flag comes with a reason a person can check, and that it only works if cost codes are consistent across jobs, which is why the audit comes first.
Example four: a forecast range, not a forecast number
Margin fade is how many points of margin a job loses between the 60% complete mark and the final result. Suppose twelve finished jobs show these changes in margin points: -5.2, -4.0, -3.5, -3.1, -3.0, -2.6, -2.4, -2.0, -1.5, -1.0, -0.5, -0.2. A job currently at 60% complete shows an estimated margin of 10.8%.
| Quantity | Margin at completion |
|---|---|
| Estimate today (no fade assumed) | 10.8% |
| Typical outcome (median fade of -2.5 points) | 8.3% |
| Lower end of the middle half of past jobs | 7.6% |
| Upper end of the middle half of past jobs | 9.4% |
The honest message to the owner is a range of about 7.6% to 9.4%, centered near 8.3%, with the caveat that twelve jobs is a small history and the mix of work matters. A machine learning model can refine this with more variables, but with this much data it will mostly add false precision. Use the simple version until you have several dozen comparable jobs. Our guide to the construction WIP report explains how percent complete and fade feed the schedule.
Where do cost code suggestions pay off?
Coding invoices to jobs and cost codes is a daily task with a clear right answer, which makes it a good fit for suggestions. The batch below is invented: 500 invoices, each given a suggested code with a confidence score, and review time depends on the band.
| Confidence band | Invoices | Suggestion correct | Review time per invoice |
|---|---|---|---|
| 95% or higher | 280 | 276 (98.6%) | 20 seconds |
| 80% to 95% | 150 | 132 (88.0%) | 45 seconds |
| Below 80% | 70 | 43 (61.4%) | 90 seconds |
Overall 451 of 500 suggestions (90.2%) are right. Reviewing every invoice with the suggestion takes about 5.2 hours, against 12.5 hours to code the batch by hand at 90 seconds each, a saving of 7.3 hours. The review rule matters more than the model: nothing posts without a person seeing it, low-confidence items get a full review, and a monthly sample of the top band checks that accuracy has not drifted. The same pattern of validating before posting applies to hours data, as our HeavyJob to ERP integration guide shows.
What should AI not do yet?
| Task | Why not yet | What to do instead |
|---|---|---|
| Post journal entries or release payments | An error becomes a ledger entry or a wire; segregation of duties disappears | AI drafts and explains; a person posts |
| Approve change orders or time extensions | Commercial and legal exposure with no accountable person | AI summarizes the clause and the cost impact; a manager decides |
| Set bid prices | It cannot know your appetite, relationships or current capacity | AI retrieves past costs; the estimator prices |
| Decide safety actions from camera data alone | False alarms and missed events both cost something, and worker privacy rules apply | Use it as an alert to a supervisor, with a clear policy |
| Answer from the open web about your contracts | No source control | Limit answers to your approved documents, with citations |
How do you set permissions for AI?
| Level | What AI may do | Example | When it is acceptable |
|---|---|---|---|
| Read | Look at approved views and documents | Answer a question about committed cost | Day one, with role-based access |
| Draft | Write text or entries that wait for approval | Report commentary, a suggested cost code | Day one, with review |
| Propose | Queue an action for a named approver | Suggest a journal entry or a payment hold | After months of clean accuracy data |
| Post | Act without a person | Auto-post a recurring entry within limits | Rarely, for narrow, reversible, low-value tasks |
Write these levels into a short AI policy before the first pilot, along with what data may be sent to which tools, who approves new uses and how errors are reported. The ABC guide mentioned above includes best practices for exactly this. Keep payroll, contract and personal data out of general-purpose public chat tools, and use tools that honor your role-based access and keep an audit log.
What does a 90-day plan look like?
| Weeks | Do this | Done when |
|---|---|---|
| 1 to 2 | Run the six-check data audit; list the top fifty questions managers ask; write the AI policy | Audit results, question list and policy exist |
| 3 to 4 | Fix the worst data issue (usually cost codes); build approved views for the top questions | Top questions answered from tested queries |
| 5 to 6 | Pilot report commentary on two reports with number checking; pilot contract and spec questions with citations | Managers approve drafts with few edits |
| 7 to 9 | Pilot invoice coding suggestions with confidence bands and a sample audit | Accuracy by band is measured and reviewed |
| 10 to 12 | Add anomaly flags on two cost codes; review what worked, retire what did not | A written decision: expand, hold or stop each use |
What should you ask an AI vendor?
- Which of my ERP systems and versions do you read from, and how (API, database view, file export)?
- Where does my data go, who can see it, and is it used to train models?
- Can answers cite the exact record or document, and can I see the query behind a number?
- How do you stop numbers that are not in the source from appearing in drafts?
- What is accuracy by confidence band on data like mine, and can I test it on a sample of my own?
- What can the AI do without a person, and can I turn that off?
- What does it cost at my volume of users, documents and invoices, including setup and data cleanup?
- What happens to my configuration if I leave?
If you are weighing a buy versus build decision for the ERP-connected parts, our AI integration services and AI development services teams build assistants, checks and integrations on top of the systems you already run. For estimating, see our guide to AI construction estimating software, and for contracts see AI contract review.
What mistakes do contractors make with AI?
| Mistake | What happens | Fix |
|---|---|---|
| Starting with the flashiest use | The pilot depends on data that is not ready | Start with report commentary and document questions |
| Pointing a chat tool at raw tables | Wrong joins and plausible wrong answers | Use approved views and saved, tested queries |
| Trusting a single accuracy number | Easy cases hide the failures | Measure accuracy by confidence band and by data type |
| No policy | Staff paste contracts and payroll into public tools | Write a one-page policy first |
| Letting AI post without review | Errors become ledger entries | Keep AI at read, draft and propose |
| No owner for data quality | Gains fade as codes drift | Name an owner for cost codes and vendor records |
Frequently Asked Questions
What is AI used for in construction?
Construction AI covers many uses, including drafting reports, answering questions about contracts and costs, suggesting cost codes on invoices, flagging unusual costs, forecasting ranges, estimating support, safety monitoring and equipment maintenance. How ready each one is depends on the data behind it and how easily a person can check the output.
What AI works today for construction companies?
The most dependable are drafting report commentary from checked numbers, answering contract and spec questions with a quoted source, suggesting cost codes with confidence scores and flagging unusual costs. Autonomous posting, change order approval and bid pricing are not ready.
Can AI replace project managers or estimators?
Not in the uses described here. AI drafts, searches and flags. Project managers and estimators still own the decisions on money, contracts and pricing, and they need to check what the AI produces.
How accurate is AI for job cost data?
It depends on your data and the task. Measure it on a sample of your own records by confidence band instead of trusting one headline figure, and check every number in a draft against the source report.
What ERP data does AI need?
It needs consistent job and cost codes, committed cost and change order status, timecards with job and equipment fields, percent-complete methods and documents in searchable form. Run a short data audit before any pilot.
Is it safe to put ERP data into ChatGPT or other public AI tools?
Not payroll, contract or personal data. Use tools that respect your access controls, do not train on your data and keep an audit log, and write a policy that states what may be shared with which tools.
How do I start with AI in a construction company?
Run a data audit, write a one-page AI policy, pick one or two checkable uses such as report commentary and contract questions, measure results for a few weeks and decide whether to expand each use.
How much does AI in construction cost?
Costs vary by tool, users, document and invoice volume, and the cleanup needed. Include setup, data work, training and review time, not just the license, and ask vendors for pricing at your volume.
Does AI work with Procore, Sage, Viewpoint or other construction systems?
It can, through the system’s APIs, database views or exports, depending on the product and version. Ask the vendor to confirm your exact system and how the connection is made, and test it on a sample.
Bottom Line
The AI that works in construction today explains and checks data you already own, with a person approving anything that touches money, contracts or safety. Prove your ERP data is ready, start with report commentary and document questions, measure accuracy by confidence band and keep AI at read, draft and propose until the evidence says otherwise. If you want that built on your systems, talk to us about our AI integration services, and read the rest of the cluster starting with the construction reporting guide and the construction automation guide.
