Support · retrieval · escalation · deflection measurement
Charles Trinh — Support Assistants

Support assistants that reduce load rather than move it: retrieval quality, escalation design, and the measurement that proves deflection instead of claiming it.
What this is
Most support assistants in healthcare fail for a boring reason: the documents they search are wrong, stale or contradictory, and the model is blamed for repeating them accurately.
The second failure is measurement. A bot that closes a ticket by routing it to a human has moved the work, not removed it, and a dashboard that counts closures will happily report success while the queue grows.
The work is therefore two-sided: build retrieval you can trust, then instrument the outcome so the claim of deflection survives someone checking it.
Deflection versus movement
A bot that closes a ticket by routing it elsewhere has moved the load, not reduced it, and the metric should say so.
Track repeat contact within thirty days by the same member on the same topic. That number is close to immune to gaming, and it is the one that decides whether the programme continues.
Retrieval quality first
Answer quality is bounded by retrieval; a model cannot be more accurate than the documents it is given.
Build a small labelled set of real questions with the passage that answers each, and measure retrieval separately from generation. When the two are measured together, every failure looks like a model problem.
Escalation design
Define the conditions for escalation explicitly, and make escalation fast, because a stuck member repeats not resolves.
Escalation should carry the transcript, the member context and what was already tried. A handoff that starts with the bot could not help pushes the whole conversation back onto the person.
What must never be automated
Anything clinical, anything binding, and anything where being wrong is irreversible stays with a human.
That list belongs in the routing rules, not in a policy appendix. If the assistant can be asked a clinical question, the refusal path has to be as well designed as the answer path.
Measurement
Resolution rate, repeat contact rate and time-to-answer; only the second one tells you whether the answer worked.
Report all three together, split by topic and by channel, because an average hides the topic where the assistant is actively making things worse.
Content as the product
The knowledge base is the product; the assistant is a query interface over it.
Whoever owns the content owns the quality. A weekly review of the questions with the worst retrieval results is more valuable than another prompt iteration.
Privacy in transcripts
Transcripts are records; retention, access and redaction need deciding before the first conversation.
Decide what is stored, who can read it, and how long it lives, then enforce it in the store. Retrofitting deletion onto a transcript archive is an expensive way to learn this.
Feedback loops
Every escalation is a labelled example; the ones nobody mined are the cheapest training data available.
Review escalations weekly, tag the cause, and feed the fix back into the documents rather than the prompt. That loop is the only one that compounds.
How I do it
The first two weeks are spent on the knowledge, not the assistant: sampling real questions, finding the passage that should answer each, and removing or correcting the documents that cannot. Most of the quality change happens here and it costs nothing per token.
Then the assistant is built over that corpus with retrieval measured separately from generation, and a routing table that names what escalates and what refuses. Escalation is built before the answer path, so there is somewhere honest to send the work from the first day.
The last step is instrumentation. Resolution, repeat contact and time-to-answer are reported weekly by topic and channel, with a standing review of the escalations. A programme that cannot show its repeat-contact trend is not ready to claim deflection.
Artifacts you keep
- A labelled retrieval set with questions and their answering passage
- Routing rules that name what must escalate and what must refuse
- A handoff payload carrying transcript, context and attempted steps
- A report of resolution, repeat contact and time-to-answer by topic
- A weekly escalation review feeding fixes into the knowledge base
What support automation may do
A support assistant may retrieve, draft and route. It may not decide entitlement, disclose a record or promise an outcome. Those three exclusions define the boundary, and they are the ones an organisation will be tempted to cross when volume rises.
Encode the exclusions as refusals in the assistant rather than as instructions in a prompt. A refusal that is a code path is a control; a refusal that is a sentence in a system prompt is a suggestion.
Retrieval quality is the product
The assistant is only as useful as the documents it retrieves. Chunking, freshness and permissions on the index decide the answer more than the model does, and the index inherits the access model of the source system or it leaks.
Test retrieval separately from generation, because a wrong answer with the right document retrieved is a generation problem and the reverse is an index problem. Conflating them is why support automation pilots stall.
Escalation with context
The measure of a support assistant is the quality of the handover when it hands over. A transcript, the retrieved sources and the customer's stated intent should arrive with the escalation, so the human starts where the machine stopped rather than at the beginning.
Track the handover rate per category. A category that always escalates is a category where the assistant should not be in the path.
Measuring it honestly
Deflection rate alone rewards the assistant for refusing to help. Pair it with customer satisfaction on deflected conversations and with the re-contact rate within seven days, because a deflection that produces a second contact has moved work rather than removed it.
Publish the three numbers monthly with the same definitions. The combination is defensible; any one of them alone can be gamed by an organisation that is not trying to game it.
T
he knowledge base behind it
A
n assistant cannot be better than the documents it reads. Ownership, freshness and the review cadence of those documents are the real project, and a stale policy document produces confident wrong answers that nobody can trace.
A
ssign an owner per document class and a review date, and surface the last-reviewed date in the retrieved source list. A reader who sees the date makes a different judgement about the answer.
Where the queue goes
Every assistant produces an exception set, and the set needs a queue with an owner and an age metric. An unowned queue is a backlog that grows quietly until it is a complaint.
Measure the age of the oldest item weekly. The number is unglamorous and it is the one that predicts whether the assistant is helping or hiding work.
The assistant's own documentation is a control too. If the team cannot describe what the assistant will and will not answer, the boundary exists only in the prompt, and the prompt is the least reviewable artefact in the system.
Proof
- ONC (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — public suite URL
- 67 public hostnames, 51 serving, one AWS account
- 104 clearinghouse enrolments across 53 jurisdictions
- USCDI v7 proposal — Patient Digital Identity Credential — listed at isp.healthit.gov
- SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564
- AWS Business Associate Addendum active since 18 August 2026
Where this work is delivered
Engagements run from Las Vegas and reach across the valley: clients in Las Vegas, NV; Las Vegas, Nevada; Henderson; Clark County, Nevada; and North Las Vegas, Nevada, plus remote delivery nationwide.
How an engagement runs
1. Trace one real case
Follow a single patient or claim end to end with a timestamp at every hop, before any document is written.
2. Written findings
A findings list naming the defect, the system, the owner and the date.
3. Monthly review
Review the list monthly and check whether the dates held.
Frequently asked
What is this service?
This page describes charles trinh — support assistants — the scope, the failure modes it prevents and how the work is measured. It sits in the AI and Automation category at trinh.io.
Where is the work delivered?
Engagements run from Las Vegas, Nevada across the valley — Las Vegas, Henderson, North Las Vegas, Spring Valley, Paradise, Winchester, Sunrise Manor, Enterprise, Summerlin, Clark County — and remotely anywhere in the United States. On-site work in the Las Vegas valley, remote delivery elsewhere.
What does it cost?
Standard rate $195 per hour, committed $165 per hour at twenty hours or more per week for three months or more; $1,600 standard day, $1,350 committed day. Scoped in writing before work starts, fixed fee where the deliverable is defined.
Who is Charles Trinh?
A technology architect in Las Vegas, Nevada, working in healthcare identity, consent and interoperability. Not a physician, not a medical practice, and not affiliated with any other Charles Trinh.
How does an engagement start?
With one trace of a real case and a written findings list naming the defect, the system, the owner and the date. The first call is twenty minutes and free: 702-559-3489.
Engagement
Scoped in writing before work starts; fixed fee where the deliverable is defined. First call is 20 minutes and free. 702-559-3489 · charles@trinh.io
Talk to Charles
- Phone: 702-559-3489
- Email: charles@trinh.io
- Based in Las Vegas, Nevada · working remotely across the United States