A self-storage operator running 29 stores connected Anthropic's Claude to his Hummingbird property management system and reported cutting 250 man-hours a month, with more than $1 million a year in savings. He built the whole deployment, by his own count, in about a week.
AJ Osborne, who has been one of the loudest voices pushing independent owners toward AI adoption, framed the stakes plainly: a single owner today can have the analytical power of a $200,000-a-year hire doing portfolio work around the clock. That is the pitch. The useful part of this case study is that the operator named the specific functions he runs in production, which lets us pressure-test the claim instead of nodding at it.
This is a teardown, not a highlight reel. Below are the 11 AI functions in that deployment, what each one actually requires to work, and the point where most owners will get stuck.
What Are the 11 AI Functions in Production?
At the core of this deployment, the operator listed his live use cases: call handling, customer service, lead management, online rental strategies, website merchandising, Google Business Profile reputation management, customer communications, operational reporting, revenue management, marketing content creation, and executive summaries and decisions.
The caveat that matters most: each category has roughly six sub-workflows beneath it. So the 11 are categories, not tasks. The real footprint is closer to 60 sub-workflows.
Here is the full list, the underlying requirement for each, and the common failure point:
| # | AI function | What it actually requires | Where owners get stuck |
|---|---|---|---|
| 1 | Call handling | Voice/telephony layer plus PMS write access to complete rentals | PMS with no open API; number porting; escalation rules |
| 2 | Customer service | Documented policies and live tenant records | Undocumented policy produces confident wrong answers |
| 3 | Lead management | Web and phone lead capture tied to a source | Attribution and de-duping leads across channels |
| 4 | Online rental strategies | Rental funnel data plus rate and availability | Testing price changes without hurting conversion |
| 5 | Website merchandising | CMS access and analytics with brand guardrails | Giving an agent write access to the live site safely |
| 6 | Google Business Profile reputation | GBP access and review data | Verification limits; review-response compliance |
| 7 | Customer communications | Message templates and an SMS/email send channel | Opt-out handling and message compliance (TCPA) |
| 8 | Operational reporting | Management summary endpoint and 12-month history | Verifying numbers match the PMS, not hallucination |
| 9 | Revenue management | Lease-level data, a market feed, and rate write-back | Governance over what the AI is allowed to change |
| 10 | Marketing content creation | Brand voice, offers, and asset library | Human review before anything publishes |
| 11 | Executive summaries and decisions | Aggregated portfolio data in a fixed format | Trust, data breadth, and machine isolation |
What Does Each Function Actually Require Under the Hood?
Strip away the labels and every function on that list depends on one thing: structured access to live data. The operator created a developer account on Tenant Inc.'s Nectar portal, requested API keys for his stores, and told Claude to read the documentation and connect.
Within about 40 minutes, I was connected to Claude and the first endpoint we hit was the management summary report, and it had visibility into every data point on the management summary report.
- Operator, 29-store self-storage portfolio
The management summary report is the keystone. It carries occupancy, move-ins, move-outs, revenue versus budget, web leads, and phone leads. Once the model can read it and 12 months of history, functions 8 through 11 (reporting, revenue management, marketing measurement, executive summaries) fall out of the same data. The operator had Claude build a knowledge base so it remembered the baseline rather than re-reading it every session.
The customer-facing functions (1 through 7) need more than read access. They need write access and a channel. Call handling has to complete a rental in the PMS. Communications need an SMS or email pipe with opt-out logic. Reputation management needs a Google Business Profile connection. Those are integrations, not prompts, which is exactly why they are harder than the reporting layer.
Where Do Owners Actually Get Stuck?
The bottleneck is never the model. It is data access and governance.
First: the PMS. Most systems are not ready for this level of integration. If your software has no open API, the model can only read a PDF you export. That still helps with reporting, but it kills call handling, live rate changes, and two-way communications.
Second: documentation. Functions 2, 6, and 9 fail quietly when the underlying policy is vague. As one operator maxim in the sector goes, garbage policies produce confident wrong answers faster. An agent that answers policy questions is only as good as the policy file behind it.
Third: governance. The model can push rate changes back into the PMS, which is powerful and dangerous. The sharpest security advice from practitioners in this space: run the agent on an isolated, wiped machine with only the logins and data you want it to touch. Put into that computer only the logins, access, and data the agent should have. That maps directly to how the independent operator technology gap is closing without a development team.
Is the $1 Million and 250 Hours Believable?
The number is plausible but unaudited, and it scales with portfolio size. This operator runs 29 stores. The savings stack from three sources: eliminated virtual-assistant and part-time labor, faster reporting, and revenue optimization that lifts net operating income.
The labor line is the most concrete. The work AI absorbed includes data crunching, customer service, and the monthly end-of-month report that had been a manual task for 15 years. One communications workflow alone reportedly saves 50 hours a month minimum.
The revenue-optimization line is fuzzier and where skepticism belongs. Attributing NOI gains to AI versus market conditions is hard, and no isolated measurement has been published. Osborne himself has noted hiring a former Marriott revenue-management specialist to formalize this work, which suggests the quantification is still being built. Treat the $1M as a directional claim from an aggressive early adopter, not a benchmark. The more defensible parallel is the broader full-stack AI revenue management trend now hitting the sector.
What Should a One- or Two-Store Owner Do First?
Do not start with call handling. Start with the management summary report, because it needs read access only and carries the highest-value data. The on-ramp is deliberately low: download your summary report, attach it, and start there.
The practical sequence for a small operator:
- Confirm whether your PMS exposes an open API. If yes, that unlocks the write-access functions later.
- Feed the model your monthly summary report and last year of history; have it rebuild your monthly report the way you actually read it.
- Add return-on-ad-spend math from the leads and revenue already in that report before touching Google Analytics.
- Only then layer in a customer-facing function with tight escalation rules and machine isolation.
The entry skill is language, not code. You need to be able to communicate in plain English what you want the system to do. For deeper mechanics, see our step-by-step guide to connecting Claude to Hummingbird.
The Numbers Worth Writing Down
- Portfolio: 29 self-storage stores
- Claimed labor reduction: 250 man-hours per month
- Claimed savings: more than $1 million per year
- Build time: about a week of self-taught use
- First live connection: about 40 minutes, starting with the management summary report endpoint
- Named AI functions: 11 categories, each with roughly six sub-workflows
- Communications workflow alone: 50-plus hours saved per month
The Model Is Not the Hard Part
The lesson here is not that AI is magic. It is that the value lives in the data plumbing, and self-storage finally has open plumbing. Everything in this deployment ran on read and write access to a PMS through an open API. Owners on closed systems can copy the prompts and still get none of the results.
The 11 functions are a shopping list, not a miracle. Reporting and revenue analysis are safe, high-value first steps because they need read access only. Customer-facing automation is where the real risk and the real integration work sit. Start where the data already lives, isolate the machine, and decide up front what the AI is allowed to change. That is the difference between a $2,400-a-year employee that never sleeps and a confident tool making expensive mistakes on your live ledger.
Sources
- Claude Cowork Product Guide, Anthropic
- Nectar Data Connector, Tenant Inc.
- Tenant Inc. and StoragePilot Announce Integration Inside Hummingbird PMS, Tenant Inc.