Building a Feedback Loop From Field Staff

At 4:40 pm in a block office the power is ugly, the queue is not, and the agent has just asked an old man to 'kindly rephrase'. The clerk already knows the rephrase: the man is asking whether last year's rejected pension can be opened again with a new age certificate. She types that English sentence because that is the only sentence the agent rewards. She also takes a photo of the screen and drops it in the district WhatsApp group. Thirty-eight similar photos already sit there. Nobody has a ticket number. Nobody has closed a single one.

Field staff are not resistant to AI. They are resistant to being made unpaid translators for a system that pretends it does not need them. A feedback loop is how you stop paying that tax in silence. It is also how you find language failures, numeral failures and scheme-name hallucinations before a reporter does.

This guide is for CIOs who think the vendor's 'thumbs down' button is governance. It is not. A loop has an owner, a taxonomy, a clock, a close, and a change that staff can see. Anything less is a suggestion box painted on a wall.

Not legal advice. Labour rules, DPDP and your disciplinary manual still apply when you ask staff to record citizens.

Why staff go quiet after week two

They reported. Nothing changed. The next time they are busy. That is the whole sociology. If the first ten reports die in a vendor backlog, you will not get the eleventh. You will get workarounds: English-only typing, paper side-registers, and a quiet instruction to juniors to avoid the agent when the collector is not visiting.

Punishment kills loops. If a language-failure report is treated as the clerk's incompetence, you will only hear about successes. Write in the standing order that a tagged failure is a system event, not a CR remark.

Time kills loops. A form that takes longer than the workaround will not be used at 4:40 pm. One extra tap on the same screen the clerk already has, or a missed-call-plus-SMS pattern, beats a portal that needs a dongle.

The minimum loop that still deserves the name

A report is a structured object: who, where, channel, language or mix, what the citizen said (de-identified), what the agent said, what the clerk did instead, a tag from a short list, and whether a right or a rupee was at risk. Free text is allowed. Free text alone is a novel.

An owner is a named cell that triages three times a week, not a shared inbox called 'AI support'. Tags that stay red for more than an agreed clock escalate to the CIO, not to the vendor CSM.

A close is a note the field can read: fixed in version, workaround, or will not fix and here is the desk. If staff cannot see closes, they will assume you threw the photo away.

A tag list short enough for a busy counter. Add codes only when two weeks of free text demand it.

Tag	Meaning	Default clock
LANG	Wrong language, silent English hop, unreadable script	Triage 2 working days
NUM	Date, amount, unit or ID mangled	Same day if money or identity
HALLUC	Invented scheme, office, or rule	Same day; consider disable
FALLBACK	Human path missing, busy, or rude handoff	2 working days
UX	Could not type, font missing, kiosk timeout	Weekly batch
OTHER	Does not fit; do not let this become the majority	Weekly review of OTHER

Data, dignity, and the screenshot habit

Screenshots of Aadhaar cards in WhatsApp groups are a DPDP incident waiting for a date. The loop must make the safe path easier than the group. Crop tools, auto-redaction, and a rule that identity numbers are not required in a language-failure ticket will do more than a circular that says 'do not share Aadhaar'.

CERT-In logging duties and departmental retention still apply to the tickets. A feedback store is an ICT system. It is also often personal data. Purpose-limit it. Do not train a model on field tickets unless that purpose was written.

Citizens should not be filmed for a feedback trophy. Staff can describe. If you need a clip, you are in the citizen-testing protocol, not the daily loop.

Close the loop where staff already stand

A monthly one-pager to BDOs: top five tags, three closes, one 'will not fix'. Read it out on the existing review call. Do not invent a new conclave.

When a LANG tag produces a new eval item, say so. Staff who see their sentence become a test will send the next one. That is how you build the language pack you refused to invent in the RFP.

Two loops, one of them imaginary

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

We already have a helpdesk number.

Helpdesks close password resets. Language failures need tags and evals. If the helpdesk will take that taxonomy and the clock, use it. If they will only log 'application issue', build a thin extra form.

Field staff will game the tags to avoid using the agent.

Some will. That is a supervision issue. It is not a reason to run blind. Compare tags with sampled sessions. A clerk who never uses the agent and never tags is also a signal.

The vendor's console already has feedback.

If the department cannot export it with the vendor accounts off, it is not your loop. Consoles die with contracts.

We cannot add taps at the counter.

Then put the tag on the existing exception path the clerk already uses when they take over. One extra field on a form they already open. Not a new portal.

Four weeks to a loop that produces eval items

Steal the tags above. Delete two. Do not add ten.

1. Week 1: pick the one screen already open at the counter. Add five tags and a free-text box. Write the no-punishment sentence.
2. Week 2: name the triage cell. Run three clinics. Close at least five old WhatsApp screenshots as a gesture of good faith.
3. Week 3: publish the first one-pager. Promote two LANG items into the held-out eval. Tell those two clerks.
4. Week 4: turn off the unofficial group for official issues, or make it a pointer to the ticket form. Do not run both as truth.

File note you can paste

Subject: Field feedback loop for the citizen agent — standing arrangement.

Field staff will report language, numeral, hallucination, fallback and UX failures through the attached tagged form. Reports are system events. They will not, by themselves, be used as CR remarks. The [cell] will triage on the attached clocks. Closes will be visible to the reporting office.

Screenshots with identity numbers are discouraged; the form will not require them. Tickets are departmental records and personal data where they identify a person. They will not be used for model training unless a separate note says so. This is not legal advice.

What we will read, and what we will not scrape

Prcept AI will ingest the tags you own, on your side of the rack, and turn LANG and NUM items into eval cases with your permission. We will not scrape your WhatsApp groups. We will not ask clerks to film citizens for our roadmap.

If we stop closing tickets when the contract gets comfortable, fire us. A silent field is not satisfaction. It is surrender.

This article is informational field guidance for Indian public institutions, not legal, language-policy, procurement, finance or engineering advice. Confirm against the live Gazette, Official Languages Act and Rules, your State's official-language law, MeitY / IndiaAI notices, GFR, GeM terms, DPDP text, departmental manuals and your counsel before you file it.

How to test this with real speech, not staff English

“Building a Feedback Loop From Field Staff” fails in the field if you only tested officers. A P1 CIO/CTO should hear a first-generation student, a rural caller, or a Hinglish grievance before claiming “field feedback loop AI”.

The people who know the agent is failing are the VLW, the CSC operator and the tehsil clerk. If their only channel is a WhatsApp group the vendor does not read, you do not have a loop. You have a graveyard of screenshots. Twenty-two scheduled languages is a Constitution fact, not a model fact. Script support is not language support. Official language rules may require bilingual output even when the model prefers one script.

• Name the languages and scripts in the eval set.
• Include code-mix and scheme-name tests.
• Measure comprehension, not BLEU alone.
• Design a human fallback when language fails.

Close this loop before the next CAB

Put “Building a Feedback Loop From Field Staff” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P1 CIO/CTO, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “field feedback loop AI” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.