AI for Support Coordinators — useful vs. hype

Every software vendor in the human services space is shipping "AI features" right now. Most of them are not actually useful for Support Coordinators. Some are. The difference matters because SC time is the most expensive resource in any agency — and AI that wastes it is worse than no AI at all.

Here's the version of this conversation we have with agencies, candidly.

Where AI is genuinely useful for SCs

Three categories pay off. The pattern across all three: AI does the structured part of the work that doesn't require professional judgment, leaving the SC's time for the part that does.

1. Summarization of long records.

An SC picking up a new caseload, or covering for a colleague, often needs to read through years of monitoring notes, prior plans, and incident history. The information is there; the time isn't.

A good summarization tool produces a tight, faithful summary that surfaces key flags — recent incidents, plan changes, escalations, decisions that need follow-up. The SC reads the summary, then drills into specifics where needed. Time saved: substantial. Risk: low, because the SC is still the one making decisions on the underlying record.

2. Flagging documents that look incomplete or contradictory.

Plans and monitoring notes have implicit consistency requirements. The same goal shouldn't be both "achieved" and "in progress." A note shouldn't reference a service that's not on the authorization. A new plan shouldn't drop a goal that the previous plan flagged as critical.

A flagging tool runs these consistency checks automatically and surfaces issues before they become audit findings. The SC still makes the call on what to do with each flag.

3. Drafting starting points for routine writing.

Some SC writing is routine: a monitoring-visit narrative for a stable consumer with nothing notable, an annual review summarizing the prior year, a notification letter following a specific template. The first draft for these is mostly assembly from existing data — and a model that has the source material can produce a serviceable draft in seconds.

The SC reviews, edits where the model got the tone wrong or missed something, and submits. Time saved on the routine cases freeing time for the non-routine ones.

Where AI gets in the way

The misuses share a pattern: AI being asked to make calls that should be the SC's, or producing output that the SC has to verify so carefully that there's no time saved.

Auto-classification of risk. If the system decides on its own that a consumer's risk level is "moderate" or "elevated" without an SC reviewing, you've crossed a line. Risk assessment is professional judgment work. AI flagging that risk might have changed is fine; AI deciding what the risk is is not.

"AI summaries" that are too summarized. A summary that says "Plan is on track" without any of the specifics is worse than no summary at all — the SC still has to read the underlying records to verify, but now they're also second-guessing the summary. The right summarization is detailed enough that it functionally replaces the underlying read for most purposes.

Chatbots that pretend to know things they don't. Conversational interfaces that confidently answer questions about a specific consumer without showing their source documents are dangerous. The SC needs to know what the AI is looking at; otherwise the AI's confidence is misleading.

Generated narratives that hallucinate. This is the failure mode that scares everyone, with reason. If the system invents a service that wasn't delivered, an observation that wasn't made, or a date that's wrong, the SC has to catch it. If the generated text is plausible but wrong, it can slip through — and the consequences for an audit are real. The right safeguard is to ground generation in the actual record data, with citations the SC can click to verify.

How to evaluate a vendor's "AI features"

When a software vendor pitches AI for SC workflows, three questions cut through the hype:

1. What specifically does the AI do, and what does the SC still do?

A clear answer means the vendor has thought about where AI ends and human judgment begins. A vague answer ("AI helps with everything!") means they haven't.

2. How does the SC verify the AI's output?

If the answer is "they trust it," that's a problem. If the answer is "the AI shows its sources, and the SC can click through to the source for any claim," that's a vendor who's actually thought about this.

3. What happens when the AI is wrong?

Every AI feature is wrong some of the time. A well-designed feature degrades gracefully — the SC sees what the AI got and what it missed, and can correct. A poorly-designed feature buries the failure mode.

What we built into ctAgency Suite™

AI for Support Coordination in ctAgency Suite™ focuses on the three useful categories above: summarization of consumer records, flagging of documents that look incomplete or contradictory, and drafting starting points for routine writing. Every AI output is grounded in source data the SC can verify with one click. We don't ship features that ask AI to make judgment calls SCs are responsible for.

It's a deliberately small set of capabilities, picked for measurable time savings without putting professional judgment at risk.

The honest version

AI for SC workflows is a real productivity gain when it's pointed at the right work. It's a liability when it's pointed at the wrong work. The same model that writes a great summary of a monitoring note can also invent a service that wasn't delivered.

The right posture for agency leadership is to be enthusiastic about category-1 features (summarization, flagging, drafting starting points) and skeptical of features that quietly take judgment work out of the SC's hands. Read what your vendor is shipping carefully. Ask the three questions.

If you want to talk through how we approach AI in our own products — or how we approach AI as a service for client projects — see AI & Automation services or reach out.