Industries / Construction and property / Commercial and residential construction / ConTech
Project records
Project records is the software builders use to keep a job's paperwork in one shared place, on the web and on tablets and phones in the field: the drawings and each revision of them, RFIs (requests for information, the formal questions a builder sends the designer), submittals (the product data and shop drawings a subcontractor sends for approval), daily logs and punch lists (lists of unfinished or faulty work).
7
Milestones achieved, from 2000 to 2025.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2000
Project files move online
Aconex, founded in Melbourne in 2000, gave the firms on a building project one shared website for exchanging and storing their documents. Oracle bought it for about $1.2 billion, closing on 28 March 2018, and called it "a provider of cloud-based collaboration software for construction projects." At the time of the deal Aconex said it had 5.5 million project users; that is the company's own figure.
Credit
Leigh Jasper and Rob Phillpot, the founders; Oracle
2012
Drawings go onto the iPad
Apple launched the iPad in January 2010. In March 2012 PlanGrid began sending construction drawings to tablets as PDFs, so crews could read and mark up the current sheet on site in place of a paper set. Autodesk bought PlanGrid for $875 million, closing on 19 December 2018, when it was used on more than one million projects in 90 countries. Autodesk later shifted its work to Autodesk Build, which it folded into its Autodesk Forma platform from late 2025. In October 2024 its own FAQ said a tool to move PlanGrid projects across was "not currently available."
Credit
PlanGrid's founders Ryan Sutton-Gee, Tracy Young, Ralph Gootee and Kenny Stone; Apple; Autodesk
2019
Buzzsaw is switched off
Autodesk shut down Buzzsaw, an online document service used by design and construction firms, on 31 January 2019 and moved its users to BIM 360 Docs. Data was kept on Autodesk's servers for up to 90 days after a contract ended or the data was moved. Autodesk did the same with BIM 360 Glue, a service for sharing and checking project models. Its help page says the data "will not be accessible after July 31st, 2026."
Lessonproject records often have to be kept for years after a job ends, longer than one software product may last, so the people who own them need a tested way to get them out.
Credit
Autodesk
2019
Inspection records go missing
On Hong Kong's Shatin to Central Link rail project, work at set stages, called hold points, must stop until inspectors check it and sign a RISC form (Request for Inspection, Survey and Check). At Hung Hom station many of these forms turned out to be missing. In January 2019 officials said about 40% were missing for one tunnel. The government's expert team later found that for some hold points as many as 78% were unavailable. In February 2019 the rail company, MTR Corporation, began using a digital RISC form system, iSuper, at another site on the line. The expert team found that iSuper improved the availability of forms but warned that "the change from paper form to digital form alone may not be a panacea."
Lessona digital form helps only if the inspection itself is done and written down at the time..pdf).
Credit
Expert Adviser Team, Transport and Housing Bureau, Hong Kong (final report, December 2020); MTR Corporation; main contractor Leighton
2020
Hong Kong requires digital records
Hong Kong's Development Bureau required public works contracts with an estimated value above HK$300 million, tendered on or after 1 April 2020, to use a Digital Works Supervision System. Such a system has five required parts, including the RISC inspection form and the site diary. Each submission and approval is time-stamped. A record can be changed only by rejecting it and submitting a new one, which supersedes the earlier record. The Urban Renewal Authority tendered for its own system in 2020. In January 2026 the bureau listed the system among the technologies "adopted in public works sites in recent years."
Credit
Development Bureau, Hong Kong SAR Government (Technical Circular (Works) No. 3/2020, 27 March 2020)
2021
Procore goes public
Procore, a California company first incorporated as Butterfly Lane, Inc. in January 2002, sells cloud software that builders use to run projects. In May 2021 it sold shares to the public at $67 each and listed on the New York Stock Exchange. It had 10,166 customers at the end of 2020, and they had created more than 1 million projects on it since 2014. By the end of 2025 it had 17,850 customers by its own count and yearly revenue of $1,323 million.
Credit
Tooey Courtemanche, founder; Procore Technologies
2025
AI pre-checks submittals
Builders began using AI agents in daily work to compare a subcontractor's submittal against the specification before it goes to the designer. Gilbane Building Company used Trunk Tools' TrunkSubmittal on 30 active projects in summer 2025 and reviewed more than 2,000 submittals in three months. It then signed up to use it on more than 200 projects over two years. Software that reads specifications was already in use: in 2020 Autodesk bought Pype, whose AutoSpecs drafted submittal logs (the list of every submittal a job needs) from the specification book for more than 325 customers. Nobody outside Trunk Tools has published how often the agent's verdicts are right.
Credit
Trunk Tools; Gilbane Building Company; Pype's founders Sunil Dorairajan and Karuna Ammireddy
Watch for
Agents that draft RFIs
Procore says its RFI Creation Agent "instantly generates RFI content and searches project documents for answers," cutting the time to get information "from days to seconds." It also offers Agent Builder, which lets customers make their own agents from plain-language instructions, for example for daily logs. Agent Builder was in open beta, meaning a public test version. In January 2026 Procore bought Datagrid, whose agents are described as doing "RFI validation" and "daily reports."
Achieved whenan outside party, such as a contractor or researcher, publishes how often the agent's answers match the designer's on real RFIs, or the agents leave beta and named customers report routine use.
Claimed by
Procore, October 2025
Spoken updates become daily logs
Autodesk says its Daily Log Agent runs on a phone and "can turn a superintendent's spoken update into a structured daily log," replacing typed entries. No accuracy figures were given.
Achieved whenthe agent is generally released and a contractor or independent tester reports routine use or measured accuracy.
Claimed by
Autodesk, at Autodesk University, September 2026
Most submittals fail, says agent
Trunk Tools says its agent found 72% of submittals on Gilbane projects non-compliant and 25% partly compliant. Later it said that across nearly 11,000 submittals it found 64.9% non-compliant and 30.6% only partly compliant. It also reports a 34% cut in review errors. These are the company's own figures.
Achieved whenan independent reviewer re-checks a random sample of the agent's verdicts against experienced human reviewers and publishes how often they agree.
Claimed by
Trunk Tools, September to December 2025
Open
01
Score RFI answers against designers
a test set of 100 closed RFIs from finished projects. Each comes with the drawings and specifications that were current when the question was asked, and the designer's official answer. An AI agent gets each question and the documents, and drafts an answer that names the sheet or specification section it relied on. Two experienced reviewers then mark each draft as agreeing, partly agreeing or disagreeing with the designer, and check that the cited sheet or section exists and supports the answer. A first pass sorts out questions that need a design decision rather than a lookup, since no document can answer those.
The briefCredit · unclaimed02
Test the submittal checkers
a test set of 50 submittals for common products, such as doors, paint and roofing membranes, each paired with the specification section it must meet. Start from public specification sections and manufacturers' public product data sheets. Plant known faults in some submittals, such as a wrong fire rating, a wrong finish or a missing test report, and leave others clean. Run each checking tool on every submittal and count what it catches and what it flags wrongly. Publish links to the source files and a list of the planted faults rather than copies of manufacturers' documents.
The briefCredit · unclaimed03
Test voice daily logs
a test of tools that turn a superintendent's spoken end-of-day update into a daily log, under the conditions they will meet: site noise, accents, trade words and more than one language. Record updates in a quiet room and mix in recorded site noise afterwards, so no one has to work on a live site. Each update is spoken from a written set of facts (crew counts by trade, weather, deliveries, inspections, delays, incidents), so the right answer is known in advance.
The briefCredit · unclaimed
Open 01 · the brief
Score RFI answers against designers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a test set of 100 closed RFIs from finished projects. Each comes with the drawings and specifications that were current when the question was asked, and the designer's official answer. An AI agent gets each question and the documents, and drafts an answer that names the sheet or specification section it relied on. Two experienced reviewers then mark each draft as agreeing, partly agreeing or disagreeing with the designer, and check that the cited sheet or section exists and supports the answer. A first pass sorts out questions that need a design decision rather than a lookup, since no document can answer those.
Why it is still open
Procore says its RFI agent searches project documents for answers and cuts the wait "from days to seconds," but the claim is its own. ENR's report on the launch carried no customer or independent test results.
Built with what exists
a long-context model that reads PDFs and drawing images together, such as Claude Fable 5.2 (see the Note "Three working devices"); a classifier such as Gemini Decision 1 to sort questions into "answerable from the documents" and "needs a design decision" (see the Note "A thousand questions, one pass"); each question run twenty times to see whether the answer holds steady (see the Note "All twenty times"); Procore's RFI agent, if the project owner gives access.
First step
ask one general contractor or public owner for 30 closed RFIs from a finished job, with the matching drawing and specification set, under a written data-use letter, and remove names and prices.
Done looks like
100 RFIs from at least three projects, each scored by two reviewers, with a published tally of agree, partly agree and disagree, the share of answers whose citation is real and supporting, and the share that changed between runs.
Open 02 · the brief
Test the submittal checkers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a test set of 50 submittals for common products, such as doors, paint and roofing membranes, each paired with the specification section it must meet. Start from public specification sections and manufacturers' public product data sheets. Plant known faults in some submittals, such as a wrong fire rating, a wrong finish or a missing test report, and leave others clean. Run each checking tool on every submittal and count what it catches and what it flags wrongly. Publish links to the source files and a list of the planted faults rather than copies of manufacturers' documents.
Why it is still open
Trunk Tools reports that its agent has reviewed nearly 11,000 submittals and found most of them non-compliant, and Gilbane uses it at scale. But the figures are the company's own, and ENR's report on Gilbane's rollout gave no measure of accuracy.
Built with what exists
Unified Facilities Guide Specifications, the US Department of Defense's public specification sections; manufacturers' product data sheets; TrunkSubmittal and Autodesk's submittal agent, where trial access is given; a general long-document reviewer like the one Enveda built on Claude (see the Note "Seven more flagged").
First step
pick five specification sections, download product data for two manufacturers' products against each, and write every requirement in each section as a numbered checklist.
Done looks like
50 submittals with 100 planted faults plus 20 clean ones, and a published table of each tool's catch rate (faults found out of 100) and false-alarm rate (clean items flagged out of 20).
Open 03 · the brief
Test voice daily logs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a test of tools that turn a superintendent's spoken end-of-day update into a daily log, under the conditions they will meet: site noise, accents, trade words and more than one language. Record updates in a quiet room and mix in recorded site noise afterwards, so no one has to work on a live site. Each update is spoken from a written set of facts (crew counts by trade, weather, deliveries, inspections, delays, incidents), so the right answer is known in advance.
Why it is still open
Autodesk announced in September 2026 a Daily Log Agent that turns spoken updates into logs, and Procore announced a Daily Log Agent in October 2025. Neither published accuracy figures.
Built with what exists
ElevenLabs Scribe v3 speech-to-text (see the Note "Fewer word errors"); a small extraction model such as Meta Muse Ember to fill in the log fields (see the Note "The jobs agents repeat"); a field list modelled on a published site diary form, such as the one Hong Kong's Digital Works Supervision System requires; the vendors' own agents, where access is given.
First step
write a 12-field daily log form and record five practice updates, then mix them with a free recording of construction noise.
Done looks like
40 recorded updates, 20 in English and 20 in one other language spoken on local sites, each at two noise levels. A published table gives the word error rate and the share of log fields filled in correctly for each tool.