What is the difference between an AI call center and workflow automation for a PHA?

An AI call center answers, understands, and routes contacts — it improves how the wait feels. Workflow completion software advances the underlying case: collecting the missing document, verifying income, reconciling EIV, and preparing a packet for staff approval. The test is whether the case changed state by the end of the interaction. Only completion permanently removes the calls that case would have generated.

Why is deflection rate a poor metric for housing authority AI?

Because deflection measures absorbed contacts, not removed ones. An incomplete case keeps producing calls until someone closes it, so a system that handles those calls gracefully without advancing the case leaves the underlying volume intact. Days from recertification notice to signed 50058, and days from RFTA received to HAP executed, are the metrics that reflect real change.

Can an AI voice agent resolve a resident's case status question?

Only if it reads live case state. An agent grounded in a knowledge base of policy documents can explain what the recertification process is, but cannot tell a specific household that their packet is missing one recent paystub. Resolving status requires read and write access to the case file, not a nightly sync.

Should a housing authority buy an AI phone system or workflow software first?

Classify a month of calls by the stalled case that caused each one, fix the largest completion bottleneck — usually document collection on recertifications or blocker triage on RFTA packets — and then put the voice layer on top of a workflow that can finish things. In the reverse order, the voice agent mostly explains that someone will get back to the caller.

What should I ask in a demo to tell the two categories apart?

Ask to see a call where the case changes state on screen during the call; what the agent says when a resident asks why their recertification has not moved; the resident's document attached to the case with the requirement marked satisfied; whether the integration is read-only sync or write-back and on what interval; and the days-to-completion improvement at a reference authority you can call.

All articles

StrategyJuly 31, 2026 · 7 min read

An AI Call Center Is Not a Workflow. Know Which One You're Buying.

Both categories are being sold with the same words. They produce very different numbers at the end of a quarter.

Corina Alparaque

Public Housing Center — writing from the work we do with housing authorities every day.

Key takeaways

Answer-and-route AI reduces how long residents wait. Workflow completion reduces how many times they have to call at all.
The distinguishing test: at the end of the interaction, has the case moved? If the outcome is a summary and a routing decision, the case has not moved.
Deflection rate is the wrong headline metric. Days-to-completion on recerts and lease-ups is the one an executive director can defend to a board.
A voice agent grounded only in policy documents can explain a process; it cannot resolve a status call, because status lives in the case file.
If you buy answer-and-route alone, the same cases keep generating calls — you have made the queue faster, not shorter.
The two are complementary. The mistake is buying the phone layer expecting the workflow outcome.

There are two distinct products being sold into public housing right now, and they are being described with almost identical language: AI, resident communication, automation, integrations with your PMS, deployed in days.

They are not the same thing, and the difference shows up in exactly the place that matters — what your numbers look like in six months.

We build one of them and have a clear commercial interest here. So rather than argue, here is the test that distinguishes them, and an honest account of what each is good for.

The test: did the case move?

At the end of the interaction, ask one question. Has the case changed state?

A resident called, the AI identified the intent, summarized it, and routed the call to a specialist's queue. The case has not moved.
A resident called, the AI verified them, read their open recertification, identified the one missing document, texted a secure upload link, received the photo, attached it to the case, and marked the requirement satisfied. The case moved.

Both of those interactions might be described in a sales deck as "resolved without human intervention." Only one of them removes future work.

This is not a criticism of the first category. Answering every call on the first ring, in the caller's language, at 9pm, is genuinely valuable — it is a real improvement over a full voicemail box and a resident complaining to a board member. But it is an experience improvement. It changes how the wait feels. It does not change how long the case takes.

Why the distinction compounds

Here is the mechanism that makes this more than a semantic argument.

An incomplete case generates repeat contact. The household calls to ask what is needed. They call again to check whether what they sent arrived. Their adult child calls. The landlord calls with the other half of the story. This continues, reliably, until someone closes the case.

If your AI answers all of those calls beautifully and none of them advance the case, you have not reduced volume — you have absorbed it. The same underlying backlog keeps producing contacts, indefinitely. Your call metrics look excellent. Your days-to-completion is unchanged. And in year two you are paying per minute for a call pattern you never removed.

If instead the interaction completes the missing step, that case stops generating contacts. The volume reduction is permanent rather than absorbed, and it compounds, because a shorter backlog means fewer open cases means fewer status calls.

That is the whole argument. It is not about voice quality or language coverage or how fast it answers. It is about whether the loop closes.

What answer-and-route requires versus what completion requires

The reason most products land in the first category is not laziness. The second is substantially harder to build, and the requirements are structural.

To answer and route, a system needs: good speech handling, your policy documents in a knowledge base, an intent taxonomy, and a way to transfer a call. All of that is genuinely achievable in a matter of days, which is why fast deployment timelines are a common selling point for this category. Those timelines are usually honest.

To complete work, a system additionally needs:

Read and write access to live case state, not just a nightly sync. It must know that Ana Ruiz's annual is open and waiting on one paystub, right now.
Identity verification that works on real PHA data. This one is genuinely hard. Our first design asked for DOB or SSN last-4 and fell over: on one authority's imported applicant file, zero of 112 records had either. The workable design verifies against what files reliably contain and residents reliably know — their name and the phone number on their file — matched server-side, with a single generic failure that reveals nothing.
The ability to take actions with consequences — send a secure link, attach a document to the right case, satisfy a requirement, open a ticket, notify a specialist — with every one of those actions logged.
A model of the actual HUD workflow, including what "complete" means for a recertification packet or an RFTA, and which exceptions require a human.
A hard decision boundary. Everything consequential — eligibility, income, rent, HAP, inspection outcomes — stays with staff. The software assembles and prepares; a person approves.

That list is why completion products take longer to stand up than a phone layer. It is also why they change the number that goes in front of your board.

They are complementary, and the sequencing matters

None of this means you should not have a good AI phone layer. You should. Most PHA call volume arrives by phone, in the evening, in several languages, from people who will not use a portal.

The mistake is buying the phone layer expecting the workflow outcome, then concluding two quarters later that AI did not work for your authority.

The sequencing we would recommend:

Find out what your calls are actually about. Classify a month of them by the stalled case that caused each one. Almost always, the largest bucket is a status check or a document that needs to get to you.
Fix the biggest completion bottleneck, which is usually document collection on recerts or blocker triage on RFTA packets.
Put the voice layer on top of a workflow that can actually finish things, so that answering a call and resolving it are the same event.

Do it in the other order and the voice agent spends its life politely explaining that someone will get back to you.

The questions that tell you which one you are looking at

In a demo, ask for these specifically:

"Show me a call where the resident's case changes state during the call." Watch whether the case file updates on screen.
"What does the agent say when a resident asks why their recert has not moved?" If the answer is generic process language, it is reading a knowledge base.
"Show me the document the resident sent, attached to the case, and the requirement marked satisfied."
"What is your integration — read-only sync, or write-back? On what interval?"
"What is your days-to-completion improvement at a reference authority, and can I talk to them?"

Products in the first category answer the first four questions with a version of "we route that to staff." That is a legitimate product. Just know that you are buying a faster front door, and that the room behind it is the same size.

What we do

Public Housing Center is a workflow completion platform first. The Recertification Desk runs annuals end to end — outreach, document collection, AI income verification, EIV reconciliation — to one 50058-ready packet per case. The Lease-Up Desk triages RFTA packets into a routed blocker list and prepares the inspection, rent, and affordability work in parallel. The AI call center sits on top of both, which is the only reason it can resolve a status call: it is reading the same live case the specialist is.

Staff approve every determination. Housing authorities run this daily.

If you want to see the difference concretely, ask us to run a live call against a real open case of yours and watch the case update while you are on the phone.

Frequently asked questions

What is the difference between an AI call center and workflow automation for a PHA?: An AI call center answers, understands, and routes contacts — it improves how the wait feels. Workflow completion software advances the underlying case: collecting the missing document, verifying income, reconciling EIV, and preparing a packet for staff approval. The test is whether the case changed state by the end of the interaction. Only completion permanently removes the calls that case would have generated.
Why is deflection rate a poor metric for housing authority AI?: Because deflection measures absorbed contacts, not removed ones. An incomplete case keeps producing calls until someone closes it, so a system that handles those calls gracefully without advancing the case leaves the underlying volume intact. Days from recertification notice to signed 50058, and days from RFTA received to HAP executed, are the metrics that reflect real change.
Can an AI voice agent resolve a resident's case status question?: Only if it reads live case state. An agent grounded in a knowledge base of policy documents can explain what the recertification process is, but cannot tell a specific household that their packet is missing one recent paystub. Resolving status requires read and write access to the case file, not a nightly sync.
Should a housing authority buy an AI phone system or workflow software first?: Classify a month of calls by the stalled case that caused each one, fix the largest completion bottleneck — usually document collection on recertifications or blocker triage on RFTA packets — and then put the voice layer on top of a workflow that can finish things. In the reverse order, the voice agent mostly explains that someone will get back to the caller.
What should I ask in a demo to tell the two categories apart?: Ask to see a call where the case changes state on screen during the call; what the agent says when a resident asks why their recertification has not moved; the resident's document attached to the case with the requirement marked satisfied; whether the integration is read-only sync or write-back and on what interval; and the days-to-completion improvement at a reference authority you can call.

Keep reading

Operations