Which Tenant Emails Deserve the Expensive Model? A Cold Suite Doesn't. A Roof Leak With 'Withholding' Does.

Which tenant emails deserve your best AI, and which ones deserve the cheap one that answers in a second and costs a fraction of a cent?

It is a real question with a real answer, and in commercial property management the answer is unusually clean, because the trade has spent forty years already sorting requests into "call the engineer" and "call the lawyer." What changed in 2026 is that the software can now make that first cut itself — and that the choice has a price tag attached to it.

Fourteen emails a day, and only two of them are dangerous

Take a management office running 1.1 million square feet across nine buildings, 180 tenants. The property@ inbox and the tenant service line together produce somewhere around 14 inbound requests a working day. Sort a month of them and the shape is always the same.

Roughly ten of the fourteen are logistics: it is cold in Suite 220, we need the freight elevator Saturday morning, an employee lost her access card, when is the lot being resealed, can we get a second key for the electrical closet, please add our new office manager to the after-hours contact list. These get a work order in Building Engines or Yardi Maintenance, a vendor call, and a reply.

Two or three are billing: what is this $840 after-hours HVAC charge, why did my CAM estimate jump 11% in January, can you send the January through March statement again. These need someone to look at the lease and the ledger.

And one or two are the dangerous ones. Water came through the ceiling tile in the northeast corner again. Our attorney has advised us regarding the condition of the common area restrooms. We consider this a failure of the landlord's repair obligation under Section 7.2. A customer tripped on the sidewalk expansion joint Friday. Those look like emails. They are the opening line of a claim.

Routing, in plain English, and why 2026 is when it became standard

Model routing means the cheap, fast AI handles the ordinary requests and only the genuinely hard ones get handed to the expensive, careful one — so you are not paying premium prices to reset an access card.

This became normal practice this year rather than a clever trick. Cisco built routing directly into the personal AI agent it is rolling out to roughly 90,000 employees, precisely to keep cost in line with capability at that scale. The same logic works at 180 tenants. The point is not that the cheap model is bad — a fast model in 2026 handles "it is cold in Suite 220, please send someone" perfectly. The point is that a fast model will also answer a question about your landlord repair obligations with total confidence and no idea what Section 7.2 says, and you would rather that question went somewhere careful.

flowchart TD
  A["Tenant request arrives by email, portal or phone"] --> B{"Any escalation trigger words or conditions?"}
  B -->|No| C{"Is it logistics or billing?"}
  C -->|Logistics| D["Fast model: work order in Yardi, vendor dispatched, tenant replied to"]
  C -->|Billing| E["Careful model: reads lease and ledger, drafts reply with clause cited"]
  B -->|Yes: water, injury, counsel, withholding, ADA| F["Property manager notified within 15 minutes, no auto-reply sent"]
  E --> G{"Property manager approves the draft?"}
  G -->|No| F
  G -->|Yes| H["Sent and logged against the tenant file"]
  D --> H

The line, written in this trade's language

Do not write the escalation rules in general terms. Write them the way a property manager talks. Anything that goes straight to a human, never to an automatic reply:

• Water intrusion, ceiling leak, mold, sewer backup — any of them, in any wording
• Injury, slip, trip, fall, or anything that implies an incident report and an insurance carrier
• The words attorney, counsel, withholding, offset, abate, constructive eviction, default, or notice to cure
• ADA, accessibility, or a complaint about a restroom, ramp or door hardware framed as access
• Fire marshal, code violation, elevator entrapment, life safety, or anything with a jurisdiction attached
• Any tenant currently inside a renewal negotiation, a lease amendment, or a delinquency workout
• Any request that would cost the landlord over an agreed threshold — set it at whatever your operating agreement lets a property manager approve without the asset manager, commonly $2,500 to $10,000

Everything else can start with the fast model. Note the last two triggers: they are not about the words in the email at all, they are about who sent it and what it would cost. That is exactly why the routing rules have to be maintained by someone who knows the portfolio.

Who owns the line, and how often it moves

Three people, with clear boundaries. The senior property manager owns the day-to-day list and can add a trigger any time a request gets handled badly. The asset manager owns the dollar threshold and the "tenant is in negotiation" flag, because those touch the value of the asset, not the running of the building. Your counsel owns the legal trigger words, and should review the list once a year — the cost is an hour of their time and it is the cheapest insurance in this whole exercise.

Review it monthly for the first quarter, then quarterly. The review is one question: which requests got a fast reply that should have gone to a human, and which went to a human that did not need to? The first kind of error is the one that costs you. Bias the list toward escalating too much and let it settle down.

One seasonal note that matters in this trade specifically: turn the sensitivity up in January and February. That is when CAM reconciliation statements land, when tenant lease administration departments start firing off audit letters, and when a routine billing question is much more likely to be the first step of a dispute. In July, "what is this charge" is usually just a question. In February it often is not.

What the mix actually costs

Assume 14 requests a working day, 250 working days, so 3,500 requests a year. Assume the split above: 71% logistics, 18% billing, 11% escalated. Assume a fast model costs about a third of a cent per handled request and the careful model — reading the lease, the ledger and the tenant history before drafting — costs about six cents. Costs at 2026 prices, which are roughly a tenth of what the same work cost in 2025.

The honest conclusion: on a nine-building portfolio, routing saves you about $157 a year. That is not the argument. The argument is speed and appropriateness — the cold-suite email gets a work order in under a minute at 9pm on a Sunday, and the roof-leak email gets your property manager's phone buzzing instead of a polite automatic acknowledgement. Routing matters for cost at Cisco's 90,000 people. At 180 tenants it matters because the two kinds of email should never be treated the same way.

Where this goes wrong

The failure that actually happens: a fast model answers a lease question. A tenant asks whether replacing the failed rooftop unit serving their suite is a landlord cost, the fast model gives a reasonable-sounding general answer about triple net leases, and now you have a written statement from the landlord's management office contradicting Section 7.2 of an actual lease. Never let the fast model answer anything that requires reading a lease. Route every lease-interpretation question to the careful model, and require a property manager to approve the reply before it sends.

Second failure: trigger words are easy to dodge. A tenant who writes "the ceiling tile in the corner is stained again, third time this year" has not used the word water and has just described a chronic roof leak with a potential mold claim behind it. Add pattern rules alongside word rules — a repeat request on the same suite and issue within 90 days escalates regardless of wording.

Third: keep a human on anything involving a tenant's employees or customers being hurt, full stop. That is not a routing decision, it is an incident report, a carrier notification and a photograph taken today.

Frequently asked questions

Can the fast model still create the work order, or does everything need approval?

Let it create work orders freely for the standard categories — temperature, lighting, janitorial, access cards, freight elevator bookings, lot and grounds items. Those are reversible and cheap. Require approval on anything that dispatches a vendor at a cost above your threshold, or anything after hours that carries a callout premium.

How do I know the routing is working without reading every email?

Sample. Pull twenty handled requests a week for the first month and read them cold. You are looking for one specific thing: a request that got a confident answer it should not have got. If you find none in eighty, sample monthly instead.

Does this apply to the phone as well as email?

Yes, and the phone is where it matters more, because a caller expects an answer now. The same triggers work spoken — a caller saying "there is water coming down in the stairwell" should never get a routine work-order confirmation, it should get a human on the line or a callback inside fifteen minutes.

What about requests from a tenant who is behind on rent?

Flag them at the account level, not the request level. Any tenant in a delinquency workout goes to a human by default, because in that situation even a routine maintenance reply can be quoted back at you.

A note from CallSphere

Routing is easiest to implement on the channel you probably have the least control over: the phone. CallSphere builds AI voice and chat agents that answer the management office line 24/7, capture the building, suite and problem, book the vendor visit, and escalate to a live person on the conditions you set — so the routine after-hours temperature call gets handled and the words "water" or "attorney" get your property manager on the phone instead of an automatic reply.