Human Fallback Design for Language Failures

The agent said, in English, that it did not support the citizen's language, and then offered a 1800 number that rang a team which also did not support the language. The ticket closed as resolved because the number had been offered. The citizen had been handed from a machine that could not hear them to a human who would not.

Fallback is not a phone icon. It is a roster that actually speaks the language you failed, a packet so the human does not ask for the Aadhaar again, a clock, and a sentence that tells the citizen what just happened. Without those, 'human in the loop' is a slide.

This guide is for CIOs who put 'escalation to helpdesk' in every RFP paragraph and never checked Tuesday night. Language failures are the escalation that most desks are least staffed for, because the English-speaking agents were cheaper to hire in the city where the tender was written.

Not legal advice. Shift-labour rules, outsourcing contracts and DPDP still govern how you record the handoff.

When to leave the model, on purpose

Leave when language-id is below a threshold you set, when the citizen asks for a human, when the agent has failed to parse a money or identity field twice, when the class is REFUSE-RIGHTS, when the citizen says the word for 'officer' in any language you know, and when the model itself is down.

Do not leave only when the citizen is angry. Anger is a late signal. Confusion is the early one. Repeated 'I do not understand' in either direction is enough.

Do not bounce back to the model after the human has taken the case unless the citizen asks. Humans who lose the caller to a cheerful bot mid-sentence will sabotage the bot, and they will be right.

The roster is the product

For every language you claim, name a desk that can finish the job in that language during the hours you claim. If Tuesday 8 pm has no speaker, your claim is smaller than your brochure.

Do not route all Indic leftovers to 'Hindi desk'. That is how you insult both the speaker and the Hindi officers.

Contracted BPOs must be tested the way the model is tested: mystery calls in the language, not a staffing Excel that lists 'Hindi, Tamil, other'. Other is not a language.

Handoff packet. If the human asks for these again, the fallback failed.

Field	Why	Must not include
Language / script the citizen used	So the human does not start in English	A joke about the accent
Cause class (LANG, UNREAD, UNSURE…)	So they do not restart the happy path	The stack trace
What the citizen already provided	Dignity and speed	Full Aadhaar unless the human job needs it
What the agent already tried	So we do not repeat a broken OCR	Chain-of-thought dumps
Reference ID	Continuity across channels	A vendor console URL the desk cannot open

Tell the citizen they have left the machine

People have a right to know whether a person is on the line. Say it. 'Aap ab karmchari se baat kar rahe hain' is a complete sentence. Do not disguise the human as a smarter bot, or the bot as a human.

If the handoff is asynchronous — a callback — say the window. A lie about 'five minutes' is a new failure message.

If the handoff is a window number in a hall, print it. Do not assume the citizen heard the IVRS over the fan.

After hours and overflow

Write the smaller claim. 'This language is staffed 10–5 on working days; after that you may leave a voice note or come to the counter.' That is a service. A 24x7 claim with a night desk that only has English is a trap.

Overflow to another district is allowed if the language matches. Overflow to a central English floor is an English gate with extra steps. Measure how often that happens. If it is the common path, you do not have fallback. You have theatre.

Two handoffs

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.

If we staff desks, why buy the agent?

The agent takes the volume it can take. The desk takes the tail and the rights. Volume without a tail is a fantasy. A tail without honesty is a locked hall.

Our BPO says they cover [all scheduled languages](/blog/do-you-need-22-languages-or-just-four).

Mystery call. Sit the same eval you sat the model. A staffing claim is a brochure.

Humans will be slower if they must read a packet.

Humans are slower if they must re-collect the packet from a tired citizen. Design the packet to be one screen.

We will add languages to the desk later.

Then remove those languages from the agent's public claim now. Claims and rosters must match on the day of go-live.

Four weeks to a fallback you can mystery-call

Do not start with the IVRS tree. Start with Tuesday 8 pm in the weakest language.

1. Week 1: for each claimed language, write hours, desk, and a named backup. Delete claims that have no row.
2. Week 2: build the one-screen packet. Remove Aadhaar unless necessary. Test with a clerk who was not in the design meeting.
3. Week 3: write the citizen sentence that announces the human. Translate it as a pack.
4. Week 4: mystery-call every language at the worst hour you still claim. File the recordings. Fix or shrink the claim.

File note you can paste

Subject: Human fallback for language and agent failures — roster and packet.

Every public language claim has a staffed desk at the attached hours. Handoff will include the attached packet. Citizens will be told they are speaking to a person. Mystery calls in each language are a quarterly control. English overflow is a recorded exception, not the design. This note is not legal advice.

Contracts and floors, not vibes

If the desk is outsourced, the SoW must name languages, hours, mystery-call rights, and the packet screen. A seat count is not a language claim. Liquidated damages for an unanswered language at a claimed hour belong in the contract, not in a steering-committee wish.

Internal desks need a floor too. A 'we have Hindi speakers in the building' sentence fails at lunch and on the fourth Saturday. Write the roster the way you write the GPU count: named, dated, fail-able.

What we will not 'contain'

Prcept AI will not sell you a containment rate that hides language failures. We will surface the handoff, fill the packet, and tell you when your roster does not match your brochure.

A sovereign stack still needs a person who can say no, and a person who can say it in the citizen's language. We will leave room for both.

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

“Human Fallback Design for Language Failures” 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 “human fallback AI design”.

A language failure without a human path is a closed door. Design who takes the call, in which language, with what context, and how the citizen knows they are no longer talking to a machine. 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 “Human Fallback Design for Language Failures” 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. “human fallback AI design” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.