← Field Notes

00Executive summary

• AI on the jobsite in 2026 is a labor story, not a tech story. Read every product launch as a response to the labor math.
• Field capture (photos, video, voice, daily logs) went automatic this year, and it now ships on the phone or the truck you already have.
• Safety, asset tracking, and AI agents all sit on top of that record. Without it, they have nothing to read.
• The headline productivity numbers are real but narrow. Two of the most-quoted ones come from a single project each. Read the fine print before you plan around them.
• For a small GC: capture first, then one flagging workflow, then one agent. Keep a human on the approve button. Ninety days, in that order.

01It's a labor story

Before you read a single product announcement, get the labor math in front of you. It explains every launch of the past six months better than any technology trend does.

In 2025, 92% of firms reported having a hard time finding qualified workers.²⁰ ENR framed the question directly: supervisors and foremen are overworked, and can AI supervision help. The National Center for Construction Education and Research projects that around 41% of the US construction workforce will retire by 2031.²³ That is not a hiring problem you solve with a better job posting.

Then there is the competition for the people who are left. 58% of construction firms say data center projects are increasing competition for workers, and nearly half feel pressure to raise wages.²⁵ Data center wages run about 32% higher: roughly $30 an hour becoming $40. If you run a $2M to $20M company, you are not bidding against the data center. You are bidding against it for your own foreman.

92% of firms reported difficulty finding qualified workers in 2025 AGC, via ENR²⁰

41% of the US construction workforce projected to retire by 2031 NCCER²³

58% of firms say data center projects increase competition for workers AGC/NCCER, Skillit²⁵

32% higher wages on data center work, roughly $30/hr becoming $40 AGC/NCCER, Skillit²⁵

Four independent figures from three sources, shown separately because they measure different things. Anyone who plots them together on one axis is selling you something.

This is why the framing in the trade press misses. The contech crowd sees an AI vision play, or an agentic AI play. The owner sees something narrower and more useful: fewer calls chasing subs for status, fewer forms re-keyed at a desk at nine at night, and a way to bid a bigger job without hiring another coordinator.

Hold that translation in your head for the rest of this paper. Every product below is worth exactly what it removes from a person's week.

02The foundation: the field record went automatic

Here is the shift that matters most this season, and it got the least coverage because it is not glamorous: capture moved onto the hardware that is already in the truck.

For a decade, getting a reliable record of what happened on a jobsite meant buying something extra. A camera on a pole. A dedicated scanning rig. A tablet that lived in the trailer. Each one was a line item, a training session, and a thing to forget. This fall, four separate companies shipped the same idea from different directions: the record should be a byproduct of work that is already happening, on a device the crew is already carrying.

• OpenSpace unveiled the next generation of its Visual Intelligence Platform at Waypoint 2026, including Autolocation 2.0, Site Mode, and an agent ecosystem.⁸
• DroneDeploy's Ground Pro, now inside Procore, turns everyday captures into measurable, verifiable 3D records from an iPhone, iPad, or 360 camera.⁴ No rig.
• JCB and Command Post put the Command Post Notes app pre-installed on all new JCB Toughphones, so daily reporting works out of the box.³ The phone the crew was already issued now does the log.
• Quickbase's FastField AI Form Builder turns PDFs, photos of paper forms, and voice into mobile forms.⁶ It is offline-first, which is a product trait worth more than the AI in it.
• CompanyCam is rolling out across Infinity Home Services' family of brands to standardize documentation.⁷ That is a multi-brand operator deciding that consistent field photos are infrastructure.

The scale of what is being captured is worth sitting with. TrueLook announced it has surpassed 5.5 billion construction images captured, a milestone, in its words, 21 years in the making.⁵ Twenty-one years to get there. The next 5.5 billion will not take twenty-one years, because the capture is no longer a decision somebody has to make each morning.

Money is following the same thesis. Buildots raised $130M, bringing its total to $297M.¹ The company says its system can reduce project duration by an average of about 15%.² That is the company's own claim, not an independent study, and we flag it here rather than in a footnote because this is exactly the kind of number that ends up in a board deck stripped of its source.

03What sits on top of the record

Once the record exists, three categories of software become useful. None of them work without it.

3a · Safety and site security: AI flags, a human decides

The most mature pattern in the field right now is not automation. It is triage: the system watches continuously, flags the handful of things worth a human's attention, and leaves the call to a person.

Sentry AI deployed its 100th SolarGuard mobile security unit, and reports that sites see an average of four suspicious events per site each week.⁹ Four a week is a number a human can handle. Four hundred would not be.

Record360's 360AI flags damage on equipment walkarounds and explicitly leaves the final call with the operator.¹⁰ Arizona State's SAVE Lab is piloting AI plus wearable biosensors for heat stress with Achen-Gardner, Willmeng, and CHASSE, on the premise that construction workers are among the most vulnerable to heat-related illness.¹² Daewoo E&C is running AI real-time translation of safety briefings for foreign workers, reporting 94.6% voice-recognition accuracy in its pilot.¹¹

On the question everyone asks about cameras and wearables: a Linxup survey of 267 field-service professionals found 77% of drivers say the benefits of fleet safety tools outweigh privacy concerns.¹³ That is one survey, from a vendor in the category, and it is worth exactly that much. But it is closer to how crews actually talk about this than the assumption that monitoring is universally resented.

The economics have moved too. Cheltenham Township, Pennsylvania approved a $35,000, five-year AI camera contract protecting an $8M community center project.¹⁶ That is roughly $7,000 a year for coverage that used to require a guard. The $8M figure is the total project value, not equipment at risk. We corrected that one ourselves during the fact-check.

3b · Assets and materials: track the small stuff

Xtellio launched sensors for small tools and heaters, with what the company claims is a 10-year battery life.¹⁴ CargoSense acquired Adapt-IP to add three-dimensional positioning accuracy within one meter.¹⁵ That acquisition is aimed at data center logistics, but the use case generalizes: knowing which gate the $500,000 generator came through, and when, is worth the same to a mid-size GC as to a hyperscaler.

3c · Agents and coordination: only as good as what they read

This is where the money went this fall, and where the dependency on the record is most obvious.

PlanRadar added an AI agent that reads incoming RFIs, searches the project documents, and drafts a sourced reply, cutting the time to a first response from hours to minutes.¹⁷ (Launched August 20, not "just launched," as several September write-ups had it.) Scaffold raised a $15M seed round from Navitas, with D.R. Horton, Pulte, and Builders FirstSource participating, to coordinate homebuilders, trades, and suppliers across more than 200,000 homes in 30 states.¹⁹ Adaptive raised a $30M Series B led by Tidemark for agentic construction accounting, serving 750+ companies from $5M to $1B in revenue.¹⁸

Note who is investing in Scaffold: three of the largest builders and suppliers in residential construction. They are not buying a feature. They are buying a coordination layer over a supply chain they already own, and it only functions if the underlying field data is clean.

An RFI agent with bad drawings and no field record just gets you the wrong answer faster.

3d · Two kinds of AI, and why the distinction saves you money

Most confusion about AI in construction comes from treating it as one thing. It is two, and they fail differently.

Language models are for work whose output is words. The RFI response. The meeting minutes. The proposal. The change order narrative. You judge them on whether a person would have written something similar, and a person reviews the result.

Decision models are for work whose output is a choice. Route this to the right sub. Flag this photo for review. Score this lead. Match this delivery ticket to this PO. You judge them on whether the choice was right, and you can measure that.

In practice they compose: a decision model sorts the capture stream, a language model writes up what matters, and a person approves. Most vendors sell you one and imply the other. Ask which one you are buying, and what it does when it is unsure.

04Read the fine print on the big numbers

Two studies are driving most of the optimism in construction AI right now. Both are real. Neither says what the headline says.

Suffolk and MIT found that six AI levers can cut project costs by 17% to 20% while trimming schedules by 22% to 25%.²¹ Read the method: it was a retrospective look at one project: a $180M, 180,000-square-foot multifamily building in San Francisco, completed in 2024. That is a useful direction of travel. It is not a promise for a $5M job in Pomona, and anyone who quotes it at you as one has not read past the press release.

JPI reports that its full development and construction cycle went from about 40 months to roughly 29 months, and that one scheduling task went from three people for four weeks down to four days.²² Those numbers are striking. What made them possible came first: a standardized design library, Revit models, and Takt sequencing. The AI automated a path they had already cleared.

Solid bar is the low end of the claimed range; the lighter segment runs to the high end. Both figures come from a retrospective look at one $180M San Francisco multifamily project, not a portfolio. Source: Construction Dive²¹. The JPI figures discussed below are not charted because we have not independently verified them.

That is the lesson under both studies, and it is the least marketable sentence in this paper. Process discipline comes first, AI second. The firms getting large returns are the ones that had already standardized. AI compounded an advantage they built the hard way. It did not create one.

If that sounds familiar, it is the same failure mode we wrote about in why most contractor AI pilots die in month three. The pilots do not fail because the model is not smart enough. They fail because there was nothing underneath.

05Autonomy is real, and not for you yet

Driverless equipment stopped being a conference demo this year. Bedrock Robotics is running driverless excavators on live commercial sites in Texas and Nevada, and has raised $350M.²³ One line from that coverage is worth more than the funding number: an excavator takes about five years to learn properly. That is the real barrier, and it is not a software barrier.

SoftBank put $225M into Autonomous Solutions Inc. in Utah and formed a joint venture for autonomous construction equipment.²⁴ Caterpillar and FieldAI are working on physical AI, autonomous inspections, and digital twins on NVIDIA Omniverse.²⁶

For a $2M to $20M GC, none of this is a purchase decision. It is a capacity story, and it is how larger firms will bid work they cannot staff. Watch it. Do not buy it. When it does reach your scale, the questions will not be about the technology. They will be about insurance and liability when autonomous equipment works near people, and nobody has good answers to those yet.

06A 90-day plan for a small GC

If you take one thing from this paper, take the order. The order is the whole point.

Measure hours saved per week and calls avoided. Do not measure "AI adoption." Nobody has ever been paid for adoption.

For a five-person trade shop rather than a GC, the same logic runs through the phone instead of the jobsite: we wrote that version up as a 30-day plan for your first AI employee.

07The vendor field test

Five questions. Ask them of every vendor, including us.

08Corrections and method

This paper was built from 98 drafts and 48 research runs between September 12 and 28, 2026. Twenty-six claims were checked against the original articles: 21 verified, 5 corrected. The corrections:

1. The Suffolk/MIT figures cover one $180M San Francisco multifamily project, not a portfolio.
2. Cheltenham Township's $8M is the project value, not equipment at risk.
3. Scaffold's $15M was a seed round.
4. Quickbase's product is FastField AI Form Builder. Offline-first is a product trait, not a claim made in the launch release.
5. PlanRadar's RFI agents launched August 20, not in late September.

Left out on purpose. Data center investment and power grid stories, state politics and polling, Huawei, the OpenAI agent security story, and humanoid robots. Off-lane for this reader. Pre-construction tools (TruTec paving takeoffs, STACK estimating, Endra electrical design, BuildCheck design review) are real, but not field work. And two figures we could not stand behind: an "80 to 90 percent documentation workload" claim with no traceable source, and a garbled statistic about crypto buyers.

Still unverified at publication. The JPI 40-to-29-months figures come from a Housing Finance article that blocks automated access; we have not independently confirmed them in a browser. Treat that section accordingly until we update it. The Buildots ~15% duration reduction is the company's own claim, labeled as such in section 2.

We publish this section because the alternative, a paper with 26 confident claims and no method, is what everyone else does, and it is why nobody trusts construction software marketing. If you find something wrong here, tell us and we will correct it with your name on it.

09Sources

1. Buildots raises $130M, $297M total. Construction Dive.
2. Buildots "about 15%" duration reduction (company claim). Jerusalem Post.
3. JCB Phone and Command Post, pre-installed daily reporting. PR Newswire, Sept 24.
4. DroneDeploy Ground Pro in Procore. Procore, Sept 22.
5. TrueLook surpasses 5.5 billion images. GlobeNewswire, Sept 22.
6. Quickbase FastField AI Form Builder. GlobeNewswire, Sept 1.
7. CompanyCam and Infinity Home Services. PR Newswire, Sept 16.
8. OpenSpace next-generation platform at Waypoint 2026. PR Newswire, Sept 10.
9. Sentry AI 100th SolarGuard; four suspicious events per site per week. PR Newswire, Sept 17.
10. Record360 launches 360AI. Rental Equipment Register, Sept 21.
11. Daewoo E&C AI safety-briefing translation, 94.6%. Herald Corp (Korean), Sept 15.
12. ASU SAVE Lab heat-stress wearables. AZ Big Media, Sept 26.
13. Linxup fleet safety survey, 77% of 267 respondents. PR Newswire, Sept 1.
14. Xtellio jobsite tool telematics; 10-year battery (company claim). MarketScale, Sept 18.
15. CargoSense acquires Adapt-IP; within one meter. Pulse 2.0, Sept 20.
16. Cheltenham Township $35,000 AI camera contract. Glenside Local, Sept 20.
17. PlanRadar adds AI agents for RFIs. AEC Magazine, Aug 20.
18. Adaptive raises $30M Series B. citybiz, Sept 18.
19. Scaffold raises $15M seed. citybiz, Sept 15.
20. AGC: 92% report difficulty finding qualified workers. ENR, Sept 24.
21. Suffolk and MIT on AI cost and schedule levers (one $180M SF project). Construction Dive, Sept 16.
22. JPI, 40 months to 29 months. Housing Finance. Unverified at publication; source blocks automated access.
23. Bedrock Robotics driverless excavators; NCCER 41% retirement by 2031. Bricks & Bytes, Sept 23.
24. SoftBank invests $225M in Autonomous Solutions Inc. AI Insider, Sept 26.
25. AGC/NCCER and Skillit on data center labor competition, 58%. Scripps.
26. Caterpillar and FieldAI. PR Newswire, Sept 2.

The Jobsite Record is published annually each fall by OBZDN Group LLC, a venture studio in Los Angeles. We build and run AI systems for construction and the trades. That is our bias, stated up front. Where we cite a number we say where it came from; where it is our judgment we say so. Next edition: fall 2027.