We taught Claude to review vendor photos so a property management team didn't have to

One of our property management clients spends most of its day keeping the homes it manages in shape. Landscapers, pool cleaners, and home cleaners all send photos after each job — proof of work, proof of standards met. That part works fine. The part that didn't: a human on the team had to open every email, look at every photo, and manually sign off before the job could be marked complete or an invoice approved.

The bottleneck: a human on every photo

For landscaping jobs, the review criteria were clear: no brown grass, no garbage cans left in frame, lawn actually mowed. The photos came in by email. Someone opened each one, scrolled through the attachments, made a call, and moved on. Not a hard task — but multiplied across dozens of vendors and properties, it added up to a significant chunk of someone's day spent doing something that was, at its core, a yes/no visual check.

That's exactly the kind of task that's easy to underestimate. Each individual review takes a minute or two. Stack enough of them together and you've turned a real person into a photo-sorting machine, clearing routine sign-offs that never actually required their judgment in the first place.

What we built

Automated photo review against client criteria

The core of the automation: incoming vendor emails get parsed, and every attached photo runs through the Claude API against the client's actual review criteria. Claude flags whether each photo passes or needs a human look. The team only sees the ones that genuinely require a decision — the ambiguous cases, the borderline fails, the situations where judgment matters.

The criteria are defined by the client and written into the system prompt. For landscaping jobs: no visible brown patches, no garbage cans in frame, clear evidence the lawn was mowed. For other vendor categories the criteria differ, but the structure is the same. Claude is working from a documented standard, not guessing.

Vendor coverage map

Once the photo review was running, we layered on two more pieces. The first: a vendor coverage map showing every contracted vendor and their service area. When a new home comes into the portfolio, the team can see at a glance who to assign rather than piecing it together from scattered spreadsheets or old email threads.

Automated invoice review

The second addition: automated invoice review. Inbound vendor invoices get checked against work records before they hit the approval queue. If an invoice doesn't match what was documented — wrong property, wrong service date, amount inconsistent with the agreed rate — it gets flagged before anyone has to approve it. The clean ones move through without touching a human queue.

What changed

Together, these three automations save the team one to two hours per day. That's not from one dramatic intervention. It's from removing a dozen small manual checks that used to require someone to stop what they were doing, open an email, scroll through photos, and move on to the next one.

The work didn't disappear — it got routed correctly. The exceptions go to a human. The routine goes to Claude. The team's attention shifted from clearing a queue to making actual decisions. That's a different kind of day.

The pattern this fits

This is a typical shape for the automation work TEG does: not a sweeping AI transformation, but a specific recurring bottleneck — a step where someone's job is basically "look at this and tell me if it's fine" — handed off to Claude so the team can focus on the cases that actually need their judgment.

Visual inspection tasks are good candidates for this because they're structured. There's a documented standard. The inputs arrive in a consistent format — photos attached to vendor emails. The output is binary: pass or needs review. Claude doesn't need to improvise. It needs to apply defined criteria at scale, which is exactly what it's built for.

The hard part isn't the Claude integration. It's identifying the criteria clearly enough to write them down. Once you can do that, the automation follows. If your team has a step in its workflow where someone's job is to look at something and decide whether it's fine, that's usually a sign there's an automation waiting to happen.

Have a similar bottleneck? We'd be happy to talk through what's possible.

Frequently asked questions

Can AI automatically review vendor photos for property management?

Yes. The Claude API can evaluate incoming photos against documented criteria — things like lawn condition, visible garbage cans, or pool clarity — and flag whether each passes or needs a human review. This works for any vendor category where you can write down a clear standard. The AI handles the routine cases so your team only sees the ones that actually require a decision.

How does automated photo review work with vendor emails?

Incoming vendor emails are parsed to extract attached photos. Each photo is sent to Claude along with the review criteria defined for that vendor category. Claude returns a pass or flag judgment for each image. Passing photos move toward approval; flagged photos land in a human review queue. The team only touches the ones that need a real decision.

What other tasks can be automated alongside photo review for property management?

The most common additions are a vendor coverage map — showing which contracted vendors serve which service areas so assignment is fast when new properties come into the portfolio — and automated invoice review, which checks inbound invoices against work records before they hit the approval queue. These three together address the main high-frequency, low-judgment tasks in a property management vendor workflow.

How long does it take to build a vendor photo review automation?

It depends on how vendor emails arrive and how clearly the review criteria are documented. If criteria are already defined and email format is consistent, the core automation can be built in days, not months. The main upfront work is writing the criteria precisely enough for Claude to apply them reliably. TEG typically starts with a scoped build on one vendor category, confirms it's working, then expands from there.