AI in the Construction Field: Photos, Voice Notes and Punch Lists

Ask crews for two habits only: take the photo, then say one sentence about it. AI turns those captions and voice notes into daily logs and a punch list afterwards, and the foreman reviews the result at the desk. If an AI step needs more than a few seconds on site, move it to the office.

A punch list is the list of small items still to fix before a job can be signed off; some contractors call it a snag list. The tools below do not find defects. They record what a person saw and said, fast, and tidy it. The foreman's eyes decide what counts as a defect, and the client or inspector decides what counts as complete.

Follow me on Instagram@sagnikteaches

Friday afternoon, final walkthrough of a two-bathroom refit with the client. Twelve items are called out in six minutes while walking, and nobody types. By Monday morning each trade should hold only its own items, each with a photo and a location, and the client should have a one-page summary. Getting there depends almost entirely on how the items were spoken on Friday.

Connect on LinkedInSagnik Bhattacharya

CompanyCam and Jobber appear below as examples, not requirements. The capture habit is what matters, and the tools only decide where the captions end up. Whatever you use, the assistant's output is a draft that the foreman reads, fixes and owns.

Subscribe on YouTube@sagnikteaches

Two-second capture rules for crews

Teach one spoken pattern: location, item, trade, action. It is short enough to say with gloves on and complete enough that an assistant can sort it. Start every area by naming it, say "next room" when you move, and take one item per photo. CompanyCam's own guidance for its walkthrough feature says the same thing: name the area you are in and say clearly when you move to a new one.

A few more habits matter. Take a wide shot and a close one, in that order, so the item can be found in the room. Keep faces, paperwork and alarm panels out of frame. Say the word "variation" aloud if a client asks for something extra, so it does not land on the defect list. And if the speaker corrects themselves, the assistant should flag it rather than choose.

The time sum is worth doing before you roll this out. Assume typing one item on a phone takes 40 seconds with a signal and clean hands, and speaking it takes 10. On a 42-item walkthrough, typing costs 28 minutes of someone's time on site (42 x 40 seconds) and speaking costs 7 minutes. Both timings are assumptions; time your own crew on a real job and use that.

What CompanyCam and Jobber actually do with voice and photos

Two tools come up most for crews who work from a phone. Their features are real, but plan rules and vendor pages do not always agree, so treat this table as what each vendor says in October 2026 and confirm it for your own plan.

On price, CompanyCam's pricing page shows three plans at $63 (Core), $129 (Crew) and $199 (Scale) a month billed annually, while its help-centre plans article lists Crew at $119 and Scale at $249 billed annually, with $199 as a legacy Scale price, so confirm the figure at checkout. The pricing page's plan table gives Core, Crew and Scale 10, 100 and unlimited AI credits, yet the same page's FAQ says AI features are included in Crew, Scale and Enterprise, and the help articles describe the checklist and document features as available on select plans. Those do not line up exactly, so ask CompanyCam which AI features your plan includes and what happens when the credits run out; I could not find the second point stated. CompanyCam's best-practices help article warns that "AI is not perfect and can make mistakes" and says to review and correct what it generates, which is the rule behind every step here.

Jobber describes its field feature in plain terms: crews can "log notes, update job progress, and document work hands-free just by speaking to Jobber in any language", and its field documentation page says crews can add notes from the field using text, photos, videos or Jobber Voice. Jobber's help article on its AI Voice and Chat beta, though, says actions and voice tasks are available to admin users, which does not match those pages, so check the current article and test with a crew login before you build a routine around it.

If you use neither, nothing here depends on them. A phone camera, built-in dictation or voice memos, and an overnight pass through a business-plan assistant produces the same lists. For a broader look at the app choices, see what AI adds to a construction photo documentation app, and which plan, and for turning spoken job notes into paperwork, turning on-site voice notes into job sheets and quotes.

A room-by-room checklist to carry on the walkthrough

A walkthrough goes faster and misses less with a checklist in hand, and an assistant can draft one from the job's scope in a minute. Your specification decides what is acceptable; the assistant only turns it into prompts to look for. A bathroom example, to adapt:

• Walls and ceilings: finish even, no runs or missed patches, edges cut in cleanly.
• Tiling: grout complete and even; sealant continuous at the bath and shower junctions; no cracked or loose tiles.
• Fixtures: taps run and shut off cleanly with no drip; waste runs freely; basin, toilet and towel rail firm.
• Electrics: extractor fan runs quietly; lights and switches work.
• Doors and joinery: doors open and close without rubbing; skirting tight to wall and floor.
• Finish: surfaces clean, protection removed, site tidy.

Draft a room-by-room walkthrough checklist for the job described below, with no more than 15 items per room, each phrased as something to look at or test.
Use only items implied by the scope. If you add an item from general practice, mark it GENERAL so the foreman can delete it.
SCOPE: [paste the scope or specification summary]

If your app can build a checklist from a spoken description, as CompanyCam's can on select plans, the same list can be created at the start of the walkthrough and ticked off with photos as you go. Either way, the foreman reads the draft once, deletes what does not apply, and keeps the final version as the firm's standard for that type of room.

Friday's walkthrough: twelve spoken items become a punch list

Here is the raw dictation from the final walkthrough, as a foreman might mutter it into a phone (illustrative):

Walkthrough, two bathroom refit, Friday. Ensuite first. Paint run under the window on the east wall, painter. Basin tap drips from the spout when it's off, plumber. Grout missing at the corner of the shower tray, tiler. Extractor fan rattles, electrician or whoever fitted it. Next room, main bathroom. Paint run, east wall again, no, the north wall by the door, painter. Towel rail loose on the wall, plumber. Silicone not finished along the left side of the bath, tiler. Mirror chipped at the corner, that's delivery damage, client wants it replaced. Bedroom one door rubs at the top corner, carpenter. Skirting gap at the stairs, carpenter. Client asked about a second shelf in the airing cupboard, that's extra, not a snag. Last one, kitchen bin still full, tidy up, labourer.

Turn the walkthrough notes below into a punch list table with columns: No | Location | Item | Trade | Type | Suggested priority | Flag.
Types: Defect (our work), Supplier damage, Variation request, Housekeeping.
One row per item. Never merge two items, even if they sound alike; if the same defect appears in two rooms, make two rows.
Use ONLY what is in the notes. If the speaker corrects themselves, keep the corrected version and write CHECK in the Flag column. If the trade is unclear, write the likely trade and CHECK. Priorities are suggestions only; the foreman decides.
NOTES: [paste]

A reliable answer is a table the same shape as this one:

The count works out as nine defects, one supplier-damage item, one variation request and one housekeeping note. That sort is the real value: a plain list of twelve items would have sent the shelf to a carpenter as a defect and the bin to the client as a failure.

The foreman's edits take about two minutes. The tap drip stays at high and goes to the plumber first, because a leak can damage finishes. The mirror becomes a supplier claim, not a workmanship defect, and the client is told so. The shelf leaves the punch list and goes onto the quotes list. The bin comes off the client-facing version. And the foreman checks the first draft's habit: without the "never merge" instruction an assistant will often fold items 1 and 5 into a single "paint runs, painter" line, because the words match. They are in different rooms and need separate photos. A merged line is how a painter fixes one run and the re-inspection finds the other.

Photos that make each item checkable

A punch list item without a photo invites an argument. Give each item a wide shot to place it in the room and a close shot to show it, and take the after photo from the same spot. Number the items on the list in the order you walk, and write the number in the photo description so the file and the line match. If your app lets you add a voice caption at the moment of capture, use it, because CompanyCam's own documentation says output quality depends on the photo descriptions: the more detail at capture, the better the log or summary that comes out later.

It is tempting to ask a chat assistant to "find the defects" in a folder of photos. Resist it for anything that matters. An assistant can describe a photo, and it may call a shadow a crack, or miss a finish that fails your specification because it has never read it. Use its descriptions to draft captions you then correct, and keep the foreman as the one who decides what is a defect. For a report layout that puts the items in severity order, with a caption written for each photo, see the snagging report template with severity order and AI captions.

When the client asks for something extra on site

Item 11 deserves a habit of its own. A client who says "while you are here, could we add..." has made a variation request, and a variation is a priced change, not a defect. Teach the crew to say "variation" aloud and capture it with the same pattern: location, request, who asked, trade. An assistant then drafts the note.

A crew lead might dictate: "Kitchen, client wants two extra sockets either side of the hob, electrician, ask for a price." A good draft reads: "Variation request. Location: kitchen, either side of the hob. Request: two additional sockets. Requested by: the client, on site. Trade: electrician. Status: awaiting price. Not yet approved; no work to start." The foreman gets the price from the electrician, the office prices it, and nothing happens on site until it is approved. The assistant never quotes a price, and it never writes "approved".

Closing items: trade messages, re-inspection and sign-off

Trades need their own items, not the whole list. Ask the assistant to split the table by trade and write one short message for each, with photos attached from the app. For the painter, in this example:

Hello [first name], punch list items for you at [address] from Friday's walkthrough:
1. Ensuite, east wall: paint run below the window. Photo attached.
5. Main bathroom, north wall by the door: paint run. Photo attached.
Please fix both by [date] and send an after photo for each. Reply if access is a problem.
[your name]

Keep the original item numbers, so a trade never sees a renumbered list that no longer matches the master. Then give each item one of four states: open, fixed (with an after photo), verified (by the foreman on site) or disputed. Ask the assistant each morning to list items that are marked fixed with no after photo, and open items past their date. Both are mechanical checks, and they stop closure by assumption.

For the client, the assistant can draft a one-page completion summary from the master table. By the end of the week it might read:

Completion summary, [address]. Of nine workmanship items from Friday's walkthrough, seven are fixed and verified. Two remain: the ensuite basin tap and the extractor fan, both booked for [date]. The chipped mirror is being replaced by the supplier, with a date to follow. The extra shelf you asked about is outside the original scope, and a quote will follow separately.

Check each sentence against the table before it goes, particularly dates and the words "fixed and verified". Only the foreman says a thing is verified. For the client-update side, writing client updates and snag lists with AI for builders has more wording to borrow.

The daily log from the same captions

The same captions can feed a daily log, since the photos and notes already describe the day. CompanyCam's Daily Log document turns the day's photos into an update with a to-do list drawn from what still needs doing, and photos without descriptions leave it with nothing to write from. With any assistant, the prompt is just as plain. Four captions from one day might be "Kitchen, first-fix electrics complete to the island wall", "Bathroom two, floor screed laid, no traffic until Thursday", "Skip changed, delivery note attached" and "Plasterer delayed to Wednesday, waiting on materials". The log built from them might read:

Work done. Kitchen first-fix electrics complete to the island wall. Bathroom two floor screed laid. Delays. Plasterer moved to Wednesday, waiting on materials. Deliveries. Skip changed; delivery note filed. Next day. No traffic on the bathroom two floor until Thursday.

The foreman checks it against what happened, adds anything missing, and files it. The rest of the technique, including daily reports from voice notes, sits in whether AI can write a construction daily report from voice notes.

What slows crews down, and how to design it out

• Noise and dust. CompanyCam's advice for voice features is to speak clearly, minimise background noise and not rush. If the room is loud, take the photo now and add the voice caption in the van.
• Signal. In CompanyCam, walkthroughs can be recorded offline and are written up once the phone reconnects, Quick Caption text appears only when you are back online, and the summary, daily log and recap documents do not work offline. Capture on site, generate later.
• Language. Jobber says crews can record notes in any language and translate them instantly. For mixed-language crews, see translation apps for briefing mixed-language crews as well.
• Too many places. One app, one folder, one naming rule. A photo in a personal camera roll is a photo nobody can find.
• Review nobody owns. The foreman reviews the list on the day. A list that waits a week is a list nobody trusts.

Keep the privacy rule from the FAQ in the toolbox talk: frame the work, not the people, and never photograph codes, keys or documents. If you paste captions or lists into a general assistant, use a business plan.

A one-week rollout for a small crew

1. Monday, ten minutes at the toolbox talk. Show the pattern, then read three weak captions aloud and fix them as a group.
2. Tuesday, a practice walkthrough in a finished room or the yard. The foreman reviews the output with the crew lead the same afternoon.
3. Wednesday and Thursday, a real job. The crew uses the pattern, the foreman reviews every list the same day, and every correction gets written down.
4. Friday, a twenty-minute review. The three most common corrections become new lines in the prompt or the next toolbox talk.
5. The following week, add a second crew and repeat the Friday review.

Two rules make the week work. Nobody types on site, so the habit is the same in the rain as in the dry. And the person who reviews the output is named, so that "the office will look at it" never becomes "nobody looked at it".

Four numbers that show whether the capture habit has stuck

• Walkthrough to issued list: the hours between the walkthrough and the list reaching the trades. The aim is the same day.
• Items with a photo and a location: the share of items that have both. This is the quality of the capture.
• Re-inspection failures: items marked fixed that are not, as a share of items. A rising figure means the closing loop needs an after-photo rule.
• Defects found after handover: the number the client finds that your walkthrough missed. Each one is a prompt to add a line to the next walkthrough's checklist.

Write the four numbers down for the next three jobs. Without a baseline, "it feels quicker" is the only evidence you have, and that is not enough to justify a subscription or to hold a crew to a new habit.

Field-capture questions foremen ask before rolling this out

Do crews need to speak in a special way?

Only a little. Teach one pattern, location then item then trade then action, and the habit of naming the room before the first item and saying 'next room' when moving. Speak clearly, pause between items and keep background noise down where you can. A sentence that follows the pattern gives an assistant everything it needs.

Can AI spot defects in site photos?

Chat assistants can describe what is in a photo, but they can misjudge a shadow, a reflection or a paint finish, and they do not know your specification. Use photo descriptions as drafts, never as findings. The person on site decides what is a defect, and the photo is the evidence of what they saw.

Which app should a small firm pick?

Start with the tools your firm already has and test them with a crew login, because plan rules about who can use AI features differ and some vendor pages disagree. If you have no app, a phone camera, built-in dictation and an overnight pass through a business-plan assistant gets you most of the benefit for no new subscription.

What about privacy in site photos?

Treat photos as records that may show occupants, children, documents on tables, alarm panels and keys. Tell crews to frame shots on the work, avoid faces, and never photograph paperwork or codes. Keep photos in the job app, not personal camera rolls, and use a business plan if you paste captions into a general assistant.

Further reads

• AI for a Painting Business: Snag Lists and Final Walkthrough Sign-Off — Snag lists and final walkthrough sign-off for painters.
• Five AI Apps for Construction Site Work: Photos, Logs, Checks — Five AI apps for site photos, logs and checks.
• AI for Electrical Project Management: Diaries, RFIs, Snag Lists — Site diaries, RFIs and snag lists for electrical work.
• Best AI App for General Contractors: Site and Office Tools Compared — Site and office AI tools priced per month for a six-person firm, each with a worked example, plus four stacks from $189 a month.
• Home Builders AI: Tools and Cost for Weekly Buyer Progress Updates — The tools and monthly cost behind weekly AI-written buyer updates for a small home builder, from a $20 assistant to a full portal, with the routine and checks.
• AI Marketing for a Construction Company: Case Studies, Bids and LinkedIn — A commercial builder's workflow: project fact sheet and photos to a case study, a bid-ready sheet and a month of LinkedIn posts, with client consent first.
• AI Tools for Construction Project Management: RFIs, Logs, Reports — One eight-week fit-out followed through its daily logs, RFIs and weekly client reports: which tools do each job, what they cost and what you must still check.
• AV Installation Checklist: Commissioning Tests and Punch List With AI — A grouped AV commissioning checklist with the tests to record, then AI steps for a room test sheet, a punch list from walkthrough notes and the sign-off page.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: CompanyCam pricing page, plans article and help articles on AI features, Quick Caption, offline mode, checklists and documents (October 2026); Jobber's pricing page, field documentation page and AI Voice and Chat help article (October 2026). Plan rules and AI limits change, so confirm what your own plan includes before relying on a feature.