AI for citizen grievance redressal and complaint management
Most grievance systems in India are not failing at resolution. They are failing at intake, where a complaint arrives too vague to act on, and at closure, where a case is marked disposed without anyone asking the citizen whether the problem stopped. Those are the two ends a voice agent can hold, and they are the two nobody is measured on.
- Actionable intake
- Complete enough that no officer calls back for clarification
- Reference on call
- Issued and read back before the citizen hangs up
- Verified closure
- The complainant confirms it, not only the department
- Repeat clusters
- A fix that did not hold, flagged instead of reopened clean
An AI grievance redressal system, in the sense that matters operationally, is a voice and messaging layer over the grievance platform you already run — CPGRAMS, a state portal, a departmental CRM. It answers the complaint line at any hour, takes the grievance in the citizen's own language, writes it into the system of record with a reference number issued on the call, chases the owning officer on a schedule, and calls the citizen back when the case is marked resolved to establish whether it actually was.
The gap it addresses is well documented and structural. Indian grievance systems report disposal rates far higher than citizen satisfaction, and the arithmetic of that gap is not mysterious: disposal is recorded by the department that owns the case, and satisfaction is reported by the person who raised it. A system in which only one of those two is instrumented will always look healthier than it is.
AiSewak builds this layer for grievance cells, department nodal officers and district administrations. The agent does not decide, close or sanction anything — it makes intake usable and closure verifiable, which is where the measurable improvement sits.
The lifecycle, and the two ends that leak
- 1
Intake an officer can act on
The difference between water problem in Ward 7 and no supply on the third-floor line of Block C since Tuesday morning, reported by four households is the difference between a case that gets worked and one that comes back for clarification, restarting the SLA clock. The agent probes until the description is specific — location, duration, what was tried, who else is affected — then stops. An interrogation loses the citizen as surely as a vague note loses the officer.
- 2
Identity, location and reference
Name, a callback number confirmed digit by digit, and a location resolved to the unit the system routes on — ward, block, revenue village, consumer number, application number. The reference is issued during the call and read back, because a citizen without a reference has no way to follow up and will file again as a new case.
- 3
Deduplication before creation
Check for an open case at the same location in the same category before creating another. Several complainants on one case is information; several cases for one problem inflates the pending count, splits the assignment and makes a ward's real performance unreadable.
- 4
Classification and routing
Category, sub-category and the owning officer — derived from jurisdiction first, then subject. A grievance correctly described and wrongly addressed is, operationally, a grievance that was not filed. Where the taxonomy has drifted from what citizens actually report, the intake data shows it within a month.
- 5
Scheduled chasing against the SLA clock
Automated follow-up on open references before the SLA expires rather than after, with each stated status recorded and dated. A cell that can show it asked on the 4th, the 11th and the 18th is in a materially different position at review than one that can show it forwarded the case.
- 6
Closure verification
When a case is marked resolved, the agent calls the complainant and asks whether it was. Confirmed closures are recorded as such; disputed ones reopen with the citizen's own words attached. This is the step that turns a disposal number into a resolution number, and almost no system performs it at scale because it is pure call volume.
Disposal is not resolution, and the gap is measurable
Every grievance platform reports disposal: a case was closed by the officer who held it. Very few report whether the citizen agrees. The two diverge for ordinary reasons — a case closed with an action-taken report that answers a different question, a fix that did not hold, a closure recorded to stop an SLA breach. None of these involve bad faith, and none are visible without asking the complainant.
A verification callback makes the divergence a number. Run it identically across every district and department and you get the first comparable quality measure a grievance cell has ever had: not how fast cases close, but what share of closures the citizen accepts. That figure is uncomfortable in month one and is the single most useful thing the system will produce.
The second measure is repeat clustering. A complaint that is the fourth about the same street or the same connection in two months is not a new complaint; it is evidence that the earlier closure was wrong. Systems that create a fresh case each time look compliant while the underlying problem persists, so the agent flags the cluster rather than letting the clock reset.
Both measures are reported with the transcript and timestamp behind every figure, because a grievance cell should not publish a resolution rate it cannot show the working for.
Sitting on top of the platform you already run
The agent is a channel and a follow-up engine. The system of record does not change.
Cases are created in CPGRAMS, the state grievance portal or the departmental CRM over its own interface, and the reference that system returns is what the citizen is told. Nothing is stored in a parallel register that later has to be reconciled — two pending counts that disagree is the failure mode every cell has seen.
Where an interface exists, intake is fully automated. Where it does not, the agent produces a complete, correctly structured case file that a clerk submits in one action rather than reconstructing from a phone note, and the reference is read back to the citizen on a short callback. Say which of these two you are building before the pilot, not during it.
Status queries read from the same system, so a citizen calling to ask what happened gets the real state of their case rather than an acknowledgement. Outbound follow-up, SMS or WhatsApp confirmation of the reference and the next step, and the closure callback all attach to the same case ID. The department keeps one lifecycle, with a channel in front of it that never sleeps.
Statutory grievance cell or a representative's office — they are not the same build
A department grievance cell or a district administration has authority over the case: it can assign, escalate within a hierarchy, commit to a service level and close. An agent built for that operator may quote the SLA, name the owning officer's post and state the appeal route, because those are facts about a system the operator controls.
An elected representative's office has standing but no statutory power. It cannot direct an officer, cannot commit to a timeline and should never imply either. An agent built for that operator has to be configured to promise nothing and to add value through persistence and follow-through instead. We build both, and the constraints are written into the prompt rather than left to tone — that page is linked below.
The practical consequence for a department is that the two systems should meet rather than compete. When a representative's office files on behalf of a citizen, it should land in the same platform with the same reference, not as a parallel escalation that arrives by letter and is counted separately.
Safety, privacy and the cases that must leave the system immediately
A grievance line is not an emergency service and must never behave as one. Fire, medical emergency, violence, a child in danger, a woman in immediate danger, an active financial fraud: the agent breaks out of the intake flow, gives 112, 1098, 181 or 1930 as applicable, confirms the citizen has the number, and raises a flagged alert rather than filing an ordinary ticket. This is a hard constraint, not a fallback.
Abuse, threats and acute distress route to a person with the transcript attached. A line that traps an upset citizen in a loop produces exactly the incident a public body cannot afford, and it is entirely avoidable by recognising the signal early and transferring.
Grievance records contain personal data and frequently sensitive personal data. Processing runs under a named purpose with consent captured in-call, a defined retention window and an executed deletion path under the DPDP Act 2023. Access is role-based and logged, Indian data residency is available, and grievance data is never reused for any other purpose — a citizen who complains about a department must not find that complaint has become a contact record for something else.
Call a live agent before you decide
These are running agents, not recordings. Open one, press call and speak to it in Hindi or English — the same stack that runs the deployments described above.
Yojana Didi — Hindi intake
A citizen question answered and the matter recorded in the same Hindi conversation.
Open the demo →Yojana Tai — Marathi
Marathi intake with Maharashtra department and scheme routing behind it.
Open the demo →Yojana Baisa — Rajasthan
Marwari and Mewari handling against Rajasthan's state grievance structure.
Open the demo →Go deeper
AI for public grievance redressal
The lifecycle in detail — where intake, routing and closure break across Indian public bodies.
Inside a state grievance engine
How 1076 handles volume, how tickets escalate, and what the disposal figures do and do not mean.
Municipal and Swachh grievance handling
Civic complaint flows, ward-level routing and the citizen-feedback windows that depend on them.
KPIs for government voice AI
Which measures survive scrutiny, and why disposal rate on its own never does.
Related pages: The main government page · Automating the helpline itself · The same lifecycle without statutory power · Municipal civic complaints · Languages and dialects
Frequently asked questions
What is an AI grievance redressal system?
A voice and messaging layer over the grievance platform a public body already runs. It answers the complaint line at any hour, takes the grievance in the citizen's own language, writes it into the system of record — CPGRAMS, a state portal or a departmental CRM — with a reference issued on the call, chases the owning officer against the SLA clock, and calls the complainant back when the case is marked resolved to confirm that it was. It does not decide, assign priority or close cases; those stay with named officers.
Does it replace CPGRAMS or our state grievance portal?
No. The platform stays the system of record and the agent files into it, reading back the reference that platform returns. Creating a second register beside the official one is how bodies end up with two pending counts that disagree. Where a platform exposes no interface, the agent produces a complete structured case file that a clerk submits in one action, and the citizen gets the reference on a short callback.
How does it improve resolution rather than just disposal?
By instrumenting the end nobody measures. Disposal is recorded by the officer who closed the case; the agent adds a callback that asks the complainant whether the problem stopped. Confirmed closures are recorded as confirmed, disputed ones reopen with the citizen's words attached, and repeat complaints about the same location and category are flagged as a cluster rather than reset as new cases. Run identically across districts, that produces the first comparable closure-quality measure a grievance cell has.
How is this different from a grievance line run by an MLA or MP's office?
Authority. A department cell or district administration owns the case — it can assign, escalate within a hierarchy, commit to a service level and close, so its agent may quote the SLA and name the owning post. A representative's office has standing but no statutory power over any officer, so its agent must promise nothing and add value through persistence instead. We build both; the representative-office version is linked in the related pages above.
What happens if someone reports an emergency on the complaint line?
The agent leaves the intake flow immediately and gives the correct number — 112 for police, fire and ambulance, 1098 for a child in danger, 181 for a woman in distress, 1930 for financial fraud, 14416 for mental health — confirms the citizen has it, and raises a flagged alert to a human rather than filing an ordinary ticket. Abuse, threats and acute distress are routed to a person with the transcript attached. A grievance line is explicitly not an emergency service.
Which Indian languages does the voice agent speak?
Hindi plus the 22 scheduled languages, and the regional variants that decide whether a call actually lands — Bhojpuri, Awadhi, Marwari, Mewari, Magahi, Santhali and others. Agents run on a mix of Indic-first models and the MeitY Bhashini stack, and a single agent can switch language mid-call when the person on the other end does.
How is grievance data protected?
Grievance records routinely contain personal and sometimes sensitive personal data. Processing runs under a named purpose with consent captured in the call, a defined retention window and an executed deletion path under the DPDP Act 2023. Access is role-based and logged, Indian data residency is available, and the data is never reused for another purpose — a citizen complaining about a department must not find that complaint has become a contact record for something else.
Can it work over WhatsApp and SMS as well as voice?
Yes, and the combination is usually what works. Voice is the channel that reaches citizens with no smartphone and no ability to navigate a form, so it carries intake. A message after the call gives the citizen the reference number, the next step and the expected timeline in writing, which measurably reduces the follow-up call asking what the reference was. Status queries and closure confirmations work on either channel.
Run the closure callback on cases you have already disposed
The fastest way to see whether this is worth doing is to take one month of closed cases in one district and let the agent ask those citizens whether the problem stopped. That single number decides the rest of the conversation.
Every AiSewak agent identifies itself as an AI at the start of the call, never asks for an OTP or a payment, and honours DND. Election deployments require ECI / state CEO registration as a political advertiser.