Anyone Can Build an Agent. Running One Is the Hard Part.

Everybody in multifamily is talking about AI agents right now. What I keep noticing is how many get built without anyone being able to see what they actually do.

Building an agent has never been easier. Running one, week after week, across a portfolio of properties, is a different job. That job is what we call the harness.

The agent isn't the product. The harness is.

The model is the part everyone talks about. The harness is everything around it: what the agent can see and do, what it costs, who can step in, and how you know it's working. For multifamily operators, here is what to look for.

• Visibility. A person should be able to read what the agent said and did, in one place, across every channel a prospect uses.
• Cost control. Different steps need different models. If everything runs on the most capable model, you pay top price for work that doesn't need it, and you need to see usage to know.
• Monitoring. Agents don't always fail loudly. Someone, or something, has to watch for errors and review quality.
• People next to the agent. Staff need a clear way to step in, and a way to stop the agent while they do.
• Change management. Models get retired. Which brings us to models.

"We'll just build it on top of our software"

I understand the appeal. If you already run a SaaS product, adding an agent on top looks like a weekend project, and with today's tools a working prototype really can come together that fast. For one property and one workflow, it can work well. I don't think the people doing it are foolish.

The trouble starts on day two. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. None of those is a model problem. They're what's left to solve once the prototype works, and they're the harness.

What happens when an AI model is retired

Providers retire models on a schedule they set, not yours. This fall, one major provider announced that a widely used model retires on November 30. Another announced that a group of its models leaves its API on January 14, 2027. Published notice policies vary: one major provider commits to at least 60 days' notice before retiring a publicly released model, and another promises at least six months for generally available models and as little as two weeks for previews.

Here is what that means in practice. Say an operator has built agents for 20 properties, with three workflows each: leasing, renewals and maintenance. That's 60 workflows. A 60-day notice means re-testing and re-approving one workflow a day, every day, with nothing else going wrong. These are illustrative numbers, so use your own. If each agent is hard-wired to one model and nobody owns the testing, a retirement date is an outage date.

Consent and fair housing questions for AI leasing agents

Leasing conversations touch pricing, availability, fair housing and consent to be contacted. In February 2024 the FCC ruled that calls using AI-generated human voices count as an "artificial or prerecorded voice" under the Telephone Consumer Protection Act, which generally requires prior express consent. Federal fair housing guidance has also moved: HUD withdrew its 2024 guidance on digital housing advertising, effective September 2025. The Fair Housing Act itself did not change. However the rules evolve, if your team can't audit what an agent said and why, you can't show anyone what happened.

Where we are at Betterbot

Here is what we do today.

• Visibility: In our Leasing Hub, each guest card shows one timeline across SMS, email, chat and voice (voice as text), with every message labeled AI, staff or prospect.
• Models: We assign a model to each type of step instead of running everything on one, and we record token usage and cost on every call.
• Monitoring: Failed jobs alert our team, integrations are monitored and paused automatically when they keep failing, and a person reviews flagged responses.
• Staff handoff: Staff get a handoff queue with email alerts that include a 24-hour AI summary. They can reply themselves or approve an AI-drafted reply over SMS or email, and they can pause the AI for one prospect, one property, or everywhere.
• Consent and fair housing: The AI doesn't text anyone who hasn't opted in. STOP, HELP and START are honored and timestamped, and the chat widget's text-us form requires a consent checkbox. Fair housing rules are written into the AI's instructions, and we keep transcripts and logs of AI activity.

At enterprise scale, this is an operations question

One agent that works is a demo. Dozens of agents across a portfolio, changing models and handing off to staff, is an operation. That's where agents tend to stall: not on day one, but on the day something changes.

Five questions to ask any agent vendor, or your own team

Whether you're evaluating an AI leasing assistant from a vendor or building one yourself, ask:

1. Can I see every conversation across every channel, and what the agent did, step by step?
2. What does each conversation cost, and can you show it by model and by workflow?
3. When a provider retires a model, who tests the replacement, how fast, and how do you roll back if quality drops?
4. When staff step in, what happens to the agent, and how do they hand the conversation back?
5. How do you handle consent and fair housing rules, and what records can you show me?

If nobody can answer the first one, the agent is running without supervision.

Bring us the agent question that worries you most. We'll show you what we do today. Book a demo at betterbot.com.

Sources

• Gartner press release, June 25, 2025
• Public model deprecation notices from major AI providers, fall 2026
• FCC-24-17
• NAA summary of the HUD withdrawal