I came home from Blueprint this year pretty energized. A lot of companies are getting clearer about where AI fits into their businesses, which workflows they want to support, and where they need help.

And a lot of those conversations came back to the same thing: data.

For the last couple of years, the question has been, “What can this AI agent do?” Can it answer questions? Write an email? Build a report? Recommend an action?

Increasingly, the answer is yes. To all of it. The demos are getting good.

The next, better question is: “What information is this agent actually working from?”

As basic capabilities become table stakes, I expect the AI arms race to look a lot more like a data arms race. That should change how operators evaluate tools.

Because an AI answer can sound perfectly reasonable while missing a pretty important piece of the story. Anyone who’s ever made a decision based on a spreadsheet called “FINAL_v3_ACTUAL_FINAL” knows the feeling.

Here’s how I’d evaluate the multifamily data behind an AI solution, and how your team can try a small version of the exercise yourselves.

Give it a question worth answering

“Help us operate more efficiently” is a reasonable goal. Measuring whether AI delivers is harder, especially when operators don’t have a reliable baseline to compare results against.

It also leaves a lot of room for a tool to declare victory. An email went out faster. A report appeared. Great. But did it actually help anyone?

Start with a specific problem someone on your team needs to solve. Name the property or group of properties, the time period, and the outcome you’re trying to understand.

For example:

  • Leasing: “Why did completed-tour-to-application conversion fall at this community over the last two months?”
  • Maintenance: “Which repair categories account for the increase in spending this quarter, compared with last quarter?”
  • Unit readiness: “Which steps are taking the most time between move-out and an apartment being ready?”
  • Portfolio performance: “How do these stabilized communities compare on vacancy loss over the same period?”
  • Resident experience: “Which types of requests are most often reopened after being marked resolved?”

Then explain what you want back: a comparison, possible explanations, a list of cases to review, or a recommendation someone can act on.

Try to leave your preferred explanation out of the question. “Why are our rents driving prospects away?” has already decided what happened. “What changed in leasing conversion, and what should we investigate?” gives the analysis some room to work.

Finally, define what improvement would look like. Fewer hours checking reports? Faster resolution of repeat requests? Shorter delays getting apartments ready?

Pick something you can measure. “We used more AI this month” probably won’t make the owner meeting. The point is to define what a useful answer would require before getting impressed by the answer itself.

What can your AI actually see?

“We integrate with your PMS” is a starting point. You still need to understand what comes through that connection and whether it’s enough for the job.

A leasing assistant might know which apartments are available but have outdated pricing. A maintenance tool might see an open work order without knowing the same issue has been reported three times. A portfolio reporting tool might compare communities without accounting for renovations or different operating conditions.

Each could produce a plausible answer. Each could also miss something that changes what your team should do.

Ask the vendor:

  • Which systems and sources inform the answer? Your operating data, external market information, uploaded documents, or some combination?
  • How much detail comes through? Can it examine underlying activity, or does it only receive summary metrics?
  • Can it see the current status? Does it know a unit’s been leased, a balance has been paid, or a repair’s been completed?
  • Does it have relevant history? Can it recognize a recurring issue or repeated inquiry?
  • What’s outside its view? Could staffing changes, vendor delays, or property-specific policies affect its recommendation?
  • How much of the portfolio is represented? Are all communities connected and providing comparable information?
  • Will it tell you what’s missing?

The information required depends on the question. Lead volume alone won’t explain why leasing conversion fell. Total maintenance spend won’t tell you whether repeat repairs are contributing to the increase.

Have the vendor walk through a real example and show where its visibility ends. An impressive collection of integration logos still leaves plenty to discuss.

How current is current?

A report on last quarter’s expenses and an answer about apartment availability today have very different freshness requirements.

Historical information may be exactly right for the first. For the second, a lease signed this morning should change the answer.

Same goes for a paid balance, a completed repair, or a vendor who’s just pushed delivery back another week. Your team’s day keeps moving. The AI agent’s information needs to keep up with the work you’re asking it to do.

Also, “uploaded today” and “current today” are two different things. Yesterday’s report can make a very punctual arrival.

Ask:

  • When did each source last update successfully?
  • How quickly do changes become available to the AI?
  • Can we see the date its answer is current through?
  • What happens if one source updates and another falls behind?
  • Does it recognize corrected or replaced information?
  • Will it flag stale data before recommending an action?

If a vendor says “real time,” ask them to demonstrate a changed record and show when it reaches the tool.

Agree on freshness requirements for each task. A daily refresh might work for one analysis and create headaches for another. The useful question is whether the information is current enough for the decision in front of you.

Can it put the pieces together?

Multifamily data has a habit of living in several places. The explanation you need may involve more than one of them.

Imagine an apartment’s been vacant longer than expected. The availability report shows the vacancy. Work orders show outstanding repairs. The vendor schedule shows a delayed delivery. Leasing records show prospects are still being quoted the original ready date.

Put those together, and the team has something useful to investigate and correct.

Of course, those records need to refer to the same apartment and the same turn. Otherwise, you’ve assembled a very detailed story about the wrong thing.

Ask:

  • How does the tool connect records from different sources?
  • Can it distinguish separate events involving the same unit, vendor, or prospect?
  • How does it handle duplicates and inconsistent naming?
  • Does it preserve enough history to explain what changed?
  • What happens when sources disagree?
  • Can it show the sequence behind its explanation?

Pay attention to definitions, too. One property’s “completed tour” might include virtual tours; another’s might exclude them. Expense categories may vary between reports.

Ask the tool to explain what it’s comparing before you let it rank properties. A performance gap can turn out to be a reporting gap, which makes for an awkward follow-up conversation.

Can you check the answer and use it?

“Leasing conversion fell because rents are too high” sounds decisive.

But did the data establish that? Or did asking rents rise around the same time conversion fell?

Maybe response times increased. Maybe tour coverage changed. Maybe fewer available apartments matched what prospects were looking for.

A useful answer should show the pattern, identify its supporting information, and explain which possibilities it can reasonably distinguish.

Then it should help someone take a next step.

Ask:

  • Can I trace this statement back to its supporting records?
  • Does it separate documented facts from possible explanations?
  • Will it say when it doesn’t have enough information?
  • What additional information could change the recommendation?
  • Does the action account for staffing, availability, and approvals?
  • Who can act on it, and how will we know whether it helped?

Recommendations need to survive contact with the property. Expanding tour hours requires someone to staff them. Replacing equipment may require an inspection and approval. Those constraints belong in the conversation.

And “I don’t have enough information” can be a useful answer. It tells your team what to check next. Confident guessing gives them a different kind of assignment.

Try a small context check yourself

You can test this approach before committing to another tool.

Say a regional manager wants to understand why fewer completed tours are turning into applications at one community. Here’s a manageable way to prepare the information for an AI analysis.

1. Define the comparison

Choose the community and time periods. Agree on what counts as a completed tour and an application.

Give prospects enough time to apply before comparing outcomes. A group that toured yesterday is still making decisions. Comparing it with last month’s group can create a problem that hasn’t actually happened.

Write down the definitions and the question you want answered.

2. Gather relevant information

Start with tour and application records. Add response times, follow-up activity, relevant pricing and concessions, and availability during the periods being compared.

Then ask the onsite team what changed. Were tour hours reduced? Was the model unavailable? Did construction affect access?

Include those observations as dated context and distinguish them from verified system records. Someone’s recollection may help explain the numbers; it still needs to be treated as a recollection.

Keep the scope manageable. You’re investigating a specific question. The entire shared drive doesn’t need an invitation.

3. Check the records

Look for duplicate prospects, missing outcomes, and inconsistent categories. Use reliable identifiers to connect records where possible.

Check a small sample against the source systems. Do the dates and statuses agree? Does “no application” mean no recorded application, or did the matching process miss it?

Document unreliable fields and gaps before asking the AI to analyze them. A blank cell shouldn’t quietly become a conclusion.

4. Make the dates and limitations visible

Record the source, extraction date, and coverage period for each dataset.

If pricing changed, use what was available when the prospect toured. Today’s asking rent can’t explain every decision made two months ago.

Give someone responsibility for keeping this information current. Otherwise, a perfectly good analysis can become outdated while sitting in a folder.

5. Ask for work you can verify

Use your company’s approved AI environment and include only the information needed. Provide your definitions and known limitations with the records.

You could ask:

Using the supplied records, compare completed-tour-to-application conversion across these periods. Show the calculation and supporting records. Identify patterns worth investigating, separate observations from possible explanations, and flag missing or conflicting information. Recommend practical follow-up checks before proposing operational changes.

Check the calculations and review the explanation with someone who knows the property. Ask what would make the conclusion change.

A well-written paragraph is easy to enjoy. The numbers still need to add up.

6. Refresh before acting and track the result

If a recommendation depends on current pricing, availability, or staffing, confirm those details first.

Choose a specific intervention and track what happens. If your team changes tour coverage or follow-up practices, measure both the change and its outcome. Keep other changes in view so you don’t give one intervention credit for everything.

This is where the effort becomes clear. Gathering files is one step. Keeping them accurate, connected, and useful takes ongoing attention.

Make the demo a little uncomfortable

In a useful way. Everybody can keep their shoes on.

Bring a work order that was reopened after being marked complete. Ask about an apartment whose ready date changed. Request a portfolio comparison where one community is missing information.

See whether the tool flags the issue, asks for a check, or fills in the gap with a confident answer.

Measure how much work it takes to use the result. Does your team spend less time validating reports? Can someone follow through on the recommendation? Does it prevent an unnecessary call or help resolve a delay?

Can it show what it knows, acknowledge what’s missing, and help someone make a better decision?

That’s the test I’d use. And it applies across leasing, maintenance, financial reporting, resident experience, and retention.

Why we care about this at Renew

At Renew, we’ve spent years working through what useful context looks like for resident retention. That work became the focus of our Resident Retention Intelligence platform: giving operators a fuller picture of resident decisions and helping teams understand where to act.

Renewals are part of that picture. So are the questions around them: What’s changing? Who needs attention? What’s a useful next step? We built Renew to help teams answer those questions without having to assemble the story manually every time.

So, yes, I’m enthusiastic about what AI can do. I’m especially interested in what happens when it has the information to be helpful. Whether you’re evaluating Renew or another tool, ask to see what it’s working from. We’re happy to have that conversation.

‍