Ask an executive director what their front-line problem is and you will usually hear a version of the same sentence: the phones never stop. Two specialists rotate onto the line, a third gets pulled in on the first of the month, and the voicemail box fills up anyway. Residents complain to the board that nobody answers. Staff complain that they cannot get through their caseload because they are on the phone.
Then the fix gets scoped as a phone problem: a new phone system, another menu layer, a callback queue, maybe a part-time person.
None of that reduces call volume. It reorganizes the wait.
Call volume is a symptom, not a workload
Here is what actually generates the calls. Pull a month of your call log and sort it, and the shape is remarkably consistent across authorities of very different sizes:
- "Did you get my paperwork?" A resident mailed or dropped off documents. Nobody confirmed receipt. They call. If the answer is unclear, they call again in a week.
- "What's the status of my recert?" The case is open, something is missing, and the notice that went out three weeks ago did not say what.
- "I got a letter, what does it mean?" Interim reexamination language, in English, at a ninth-grade reading level, to a household that speaks Spanish or Vietnamese at home.
- "My landlord says he sent it in." The RFTA packet is incomplete and both parties think the other is the holdup.
- "When can I be seen?" Waitlist and application status.
Look at that list and notice what it has in common. Almost none of these are questions about policy. They are all status checks on a case that has not moved, or attempts to fix a document gap that nobody told anyone about.
That is the important reframe: your call volume is a lagging indicator of how many cases are sitting incomplete. Every stalled recert is not one call. It is a call, then a follow-up call, then a call from the daughter who helps her mother with paperwork, then a call from the landlord.
This is why hiring for the phones feels like bailing. The cases keep generating calls faster than the callers can be handled, and the specialists you pulled onto the line are the same people who would have closed the cases.
What an IVR does and does not do
Menu trees and callback queues are not useless. They protect staff from being interrupted at random, and a good callback option is genuinely kinder than a hold queue.
But they do not touch the driver. A caller who navigates four menu levels to reach a specialist who then has to open the file, read the notes, and say "we're still waiting on a paystub" has consumed the same staff minute as before. And that caller will call again, because "we're waiting on a paystub" without a way to send the paystub right now is not a resolution.
The test for any phone investment is simple: does a call end with the thing the caller needed actually done? If not, you have bought a nicer waiting room.
Run the classification. It costs you an afternoon.
Before you buy anything, do this analysis. It is the single most useful hour of prep for either a business case or a vendor conversation, and it requires no new software.
Take one month of calls. For each one, record two fields:
- What the caller wanted — status check, document submission, policy question, complaint, landlord issue, new application, something else.
- Which open case caused it, if any — and whether that case was waiting on a document, waiting on staff review, or waiting on a third party.
Then count how many calls trace back to the same case.
What most authorities find is that a small number of stalled cases generate a disproportionate share of the phone traffic, and that the single largest bucket is document collection — a resident trying to get a piece of paper into your hands, or find out whether the one they already sent arrived.
Now you have the real number. You are not buying a phone system. You are buying a way to stop generating those calls.
What a useful AI voice agent has to do
We build one, so read this with that in mind — but the criteria are the ones we would apply as a buyer, and they rule out most of what is being sold into this market right now.
It has to read the live case file. An agent that answers from a knowledge base of policy documents can tell a caller what the recertification process is. It cannot tell Ana Ruiz that her packet is missing one recent paystub. The second is the call she actually made. If the AI is not connected to case state, it will deflect the easy 15% and escalate everything that mattered.
It has to verify the caller before it says anything. This is where a lot of vendor demos quietly cheat. Before an agent discusses a household's file, it has to know it is speaking to that household — and it has to do it with the data your files actually contain.
We learned that the expensive way. Our first verification design asked for date of birth or the last four of an SSN, with fallback questions about house number and application date. On one authority's imported applicant data, zero out of 112 applicants had a usable DOB or SSN on file. Every call bottomed out in a quiz about the month someone applied eight years ago. Callers could not answer it. Staff could not explain it. The security was theoretical; the friction was completely real.
The design we run now uses the two things a PHA reliably has and a resident reliably knows: their full name and the phone number on their file, matched server-side against the same household, with one generic failure message that never reveals which half missed. Ask any vendor what happens when the file has no DOB. The answer tells you whether they have run on real PHA data.
It has to be able to act mid-call. The call that ends "I've texted you a secure link — take a photo of the paystub and you're done" is the only version that removes the next three calls. That means the agent needs to send the link, attach the returned document to the right case, and update the case state. Answering and summarizing is not the same as finishing.
It has to hand off warmly, with the transcript. Some calls should reach a person — a complaint, a VAWA disclosure, a reasonable accommodation request, a caller who is distressed or just wants a human. When that happens, the staff member should pick up with the context already in front of them, not start from "how can I help you?"
It has to know when a call is not a resident. This one is not on anyone's RFP and it should be. This summer, an MFA robocall dialed one authority's main line and read a payroll-tax one-time passcode aloud to the AI agent. The staff member who triggered that code sat waiting by a phone that never rang. The code was in the transcript the whole time and nothing was watching for it. We now detect authentication codes in transcripts — spoken as digits, as words, as NATO phonetics — and alert staff immediately. Ask a vendor what their system does with that call. Most have never considered it.
What actually changes when this works
The phone metric everyone quotes is the deflection rate. It is the wrong headline. The number that changes an executive director's week is days-to-completion on recerts and lease-ups, and the phone volume falls out of it.
The chain is direct: documents arrive same-day instead of in a week, so cases do not sit; cases that do not sit do not generate status calls; the specialists who were on the phones go back to determinations; the backlog shrinks; fewer open cases means fewer calls still.
That is the loop worth buying. Answering more calls faster is the loop you are already in.
Where to start
- Classify one month of calls by the case that caused them. Do this first, whatever you buy.
- Fix the notice language for your top three call-driving letters, in your top two non-English languages. This is free and it moves volume.
- Make document submission one tap. No portal, no password, no app. A secure link to a phone camera. Most of your document-chasing volume lives here.
- Then evaluate a voice agent, against the five criteria above, on your own caseload rather than a demo dataset.
If you want to see what a call looks like when the agent can read the file, verify the caller, and text the link mid-call, we will run one on a real workflow of yours.