AI agents for RFIs and submittals in construction: what they do, what the evidence shows, and where people stay in charge

Every building project runs on a stream of questions and approvals. A request for information (RFI) is a written question from the builder to the design team, such as "the drawings show two different beam sizes, which is right?" A submittal is the product data, shop drawing or sample a contractor sends for approval before ordering materials. Both are slow, repetitive and easy to get wrong, which is why software vendors are aiming AI agents at them. An AI agent here means software that does a multi-step task, such as reading a spec, checking a document against it and drafting a response, rather than only answering a question. This article looks at what these tools do today, how widely contractors use them in the US and Canada, and how to adopt them without handing over professional judgment.

The problem: paperwork that stalls the job

Construction teams lose time to information gaps. A widely cited 2018 survey of 600 construction professionals by PlanGrid and FMI found that poor communication accounted for 26% of rework and poor project information for another 22%, and that workers spent more than 14 hours a week on "non-optimal" tasks such as hunting for documents ¹. That study is old and was commissioned by a software company, so treat the exact figures with care. The pattern it describes still matches what contractors say today. At the 2026 DBIA design-build conference in Cleveland, one vendor reported that submittal review "came up again and again" from general contractors and subcontractors alike, and called it a continuing bottleneck ².

The work is hard to speed up by hand. A submittal reviewer may compare a data sheet against hundreds of pages of specifications, checking thicknesses, ratings and certificates. An RFI writer has to find the conflicting drawing sheets, describe the issue clearly and route it to the right person. Late or vague RFIs delay answers, and late answers delay crews.

What the new agents actually do

The main platform vendors shipped agents for this work during 2026. In May, Procore announced five agents built on technology from Datagrid, a company it acquired: a deep search agent, a submittal reviewer, an RFI agent, a daily log agent and a contract review agent ³. Engineering News-Record reported that the RFI agent checks requests for completeness and clarity and suggests edits ⁴. By October, Procore said its submittal agent was live and could update a submittal everywhere it is referenced in the drawings. A customer, McCownGordon Construction, said it could update up to 500 submittals at once ⁵. A procurement-log agent was still in closed beta then ⁵.

Smaller companies are moving in too. Shackleton AI, a New York start-up that came out of stealth in November 2026 with a $5 million seed round, says its agents draft RFIs, follow up on overdue submittals and flag budget variances before payment requests, with a person approving the output ⁶. Vendors demonstrate the review side with concrete examples: in one demo, an AI review of a plumbing insulation submittal flagged missing thickness details, a missing code-compliance statement and a missing manufacturer certificate ².

In short, the pattern is the same across vendors. The agent reads the documents, produces a first draft or a list of gaps, and a person decides.

How many firms are really using them

Use is growing but remains uneven. In the US, an AGC and Sage survey of 951 contractors (fielded November to December 2025) found 61% said their firms use AI or plan to increase investment in it, up from 44% a year earlier. The most common uses were office and administrative work, estimating and preconstruction ⁷. A separate survey of 235 US general and trade contractors by Dodge Construction Network, in partnership with software provider CMiC, found 85% expect AI to reduce time spent on repetitive tasks and 40% are setting aside a dedicated AI budget ⁸.

Official data shows a lower baseline. The Federal Reserve Bank of Minneapolis, reading Census Bureau data, reported that roughly 20% of US businesses use AI, but under 10% in construction (alongside agriculture, transportation and accommodation and food services) ⁹. In Canada, Statistics Canada found that 9.2% of construction businesses expected to use AI over the next 12 months in the third quarter of 2025, up from 3.2% a year earlier, against 14.5% for all businesses ¹⁰. A KPMG Canada survey of 265 construction companies, done with the Canadian Construction Association, found 90% agree tools such as AI and analytics can boost efficiency and 53% are prioritizing AI and AI-driven software ¹¹.

The gap between "interested" and "using" is the main story. Plans and budgets are ahead of daily practice.

What the evidence shows, and what it does not

The honest summary is that the evidence for time savings is mostly anecdotal and vendor-reported. Procore’s release quotes a customer, Level 10 Construction, saying the team ran 10 submittal reviews in an hour and saved about 12 hours of work ³. That is a single company’s account supplied by the vendor, with no baseline described. Shackleton’s headline claims, such as customers finishing up to 35% ahead of schedule, relate to multi-location customers such as restaurant chains, come with no disclosed baseline or sample, and the trade article reporting them says no RFI or submittal turnaround data is public ⁶. That article also notes it was drafted with AI and reviewed by an editor, which is another reason to read it as a launch summary and not as independent reporting.

Returns on AI in construction overall look modest so far. In an October 2026 interview, the chief executive of software company Kahua cited his company’s survey in which only 8% of construction respondents reported significant value from their AI investments. He said Canadian respondents were more hesitant. The survey’s sample size was not stated ¹². In an RSM survey of 80 construction firms in the US and Canada in early 2025, the top challenge was data quality (36%), and 59% said they were only somewhat or not very prepared for AI ¹³.

None of this means the tools do not work. It means no one has yet published a controlled, independent test of AI-reviewed submittals against human-only review, so your own baseline is the best evidence you will get.

Limits, risks and keeping people in charge

Accuracy. In the Dodge and CMiC survey, 57% named unreliable or inaccurate output as a chief concern, and 54% named data security and privacy ⁸. A tool that misses a missing certificate, or invents a spec requirement, can pass a bad product into the building. The cost of one wrong approval is far larger than the minutes saved.

Responsibility. An agent cannot hold a professional licence or sign off on conformance. Contracts usually name who reviews submittals, often the design team, and that person owns the judgment. Check what your contract says before an AI reviews on anyone’s behalf, and check with your insurer and counsel.

Data. RFIs and submittals contain drawings, pricing and sometimes owner-confidential details. Ask where documents are stored, whether they are used to train models, and who can see them.

Design for doubt. Large builders show what careful deployment looks like in a nearby area. Skanska’s safety assistant draws on its own safety manual and OSHA standards, is mostly used for planning, has its outputs vetted by safety staff, and was trained to say "I don’t know" rather than guess ¹⁴. Procore says its agents require human consent before actions are finalized ³. Those are good design rules for any project agent: stay inside approved documents, admit uncertainty, and never act without a named person’s approval.

Next steps for contractors, owners and design teams

1. Measure first. Pull 60 to 90 days of RFI and submittal logs: how many, how long to first response, how many returned for resubmission, how many caused schedule or cost impact. Without this baseline you cannot tell whether a tool helped.
2. Start with drafting, not deciding. Use agents first for RFI drafting, completeness checks, log updates and overdue reminders. Keep approval and rejection of submittals with the responsible reviewer.
3. Run a side-by-side pilot. On one project, have the tool and a human reviewer check the same 30 to 50 submittals and compare what each caught. Track misses, not just speed.
4. Settle the contract questions. Tell the owner and design team how you plan to use AI, confirm who remains responsible for review, and record it.
5. Clean up the inputs. Current specs, a clear drawing index and consistent naming matter, since data quality is a top barrier ¹³.
6. Set data rules. Require written terms on storage, training use and access before project documents go in.
7. Review after 90 days. Compare against your baseline and keep, narrow or drop the tool.

Sources

1. PlanGrid and FMI via Construction Dive (Kathleen Brown), Aug. 7, 2018. Industry could be overspending $177B per year, study finds. Survey of 600 professionals, commissioned by a software vendor, dated.
2. iFieldSmart (vendor blog), Nov. 5, 2026. DBIA 2026 Day 1: AI Construction Workflows. Vendor promotional post, qualitative, no sample.
3. Procore Technologies (vendor press release), May 21, 2026. New Procore AI Experience Embeds Datagrid into Procore. Vendor claims and customer quotes supplied by the vendor.
4. Engineering News-Record (Jeff Yoders), May 21, 2026. Procore Releases New AI Agents After Datagrid Integration. Trade press.
5. Engineering News-Record (Jeff Yoders), Oct. 21, 2026. Procore Unveils AI Agents for Submittals, Procurement Logs at Groundbreak Conference. Trade press, customer statement is anecdotal.
6. Construction Metrics, Nov. 5, 2026. Shackleton AI launches agents that draft RFIs and chase overdue submittals. Launch coverage, site states posts are drafted with AI and editor-reviewed, company claims unverified.
7. AGC and Sage, Jan. 8, 2026. Contractors Have "Dampened" Expectations for 2026. Survey of 951 US firms, fielded Nov. 4 to Dec. 15, 2025, Sage is a software vendor.
8. Construction Dive (Matthew Thibault), Dec. 10, 2025. AI nears "tipping point" in construction as contractors pilot tech: survey. Dodge Construction Network survey with CMiC (software vendor), n = 235 US contractors.
9. Federal Reserve Bank of Minneapolis (Erick Garcia Luna), May 29, 2026. AI adoption in business grows steadily but unevenly. Analysis of US Census Bureau Business Trends and Outlook Survey data.
10. Statistics Canada, Sept. 11, 2025. Analysis on expected use of artificial intelligence by businesses in Canada, third quarter of 2025. Canadian Survey on Business Conditions, collected July 2 to Aug. 6, 2025. Government data, expected use, not current use.
11. ReNew Canada, reporting KPMG in Canada and the Canadian Construction Association, June 19, 2025. Canada’s construction industry gets serious about investing in technology. Survey of 265 Canadian construction companies, fielded Mar. 18 to Apr. 4, 2025, consultancy research.
12. BNN Bloomberg, Oct. 23, 2026. Investor Outlook: Construction firms struggle to turn AI use into value. Interview citing a Kahua (software vendor) survey, sample size not stated.
13. RSM US (Mac Carroll and Brandon Maves), July 29, 2025. Construction’s AI maturity hinges on strategic investment. 2025 RSM AI survey, 80 construction respondents in the US and Canada, consultancy research.
14. Construction Dive (Matthew Thibault), May 6, 2026. How 3 builders are using AI for safety. Trade press, company descriptions of internal tools.