A regulator's complaints channel only works if every conversation ends up in the right hands. A citizen names a bank, an insurer, a mobile money provider – often by nickname, abbreviation, or however they've always referred to it – and somewhere behind the scenes, that has to become a case sitting in the correct institution's queue. When the regulator oversees one institution, this is trivial. When it oversees hundreds or thousands, it's the difference between a system that scales and one that quietly falls over.
This is a problem we've solved for a number of regulators now, and it's solved the same way each time: the AI agent listens for what the citizen actually says, matches it against the institution they mean, and routes the case automatically – no live agent required to work out where it belongs. Neither citizen needs to know the internal structure of who handles what. They describe their problem in their own words, and the right team ends up with the case.
What makes hundreds of institutions manageable
Underneath deployments like these is a simple idea: every institution a regulator supervises is set up as its own team, and each team carries a list of the names people actually use for it, rather than only its formal registered name:
- Common nicknames and abbreviations
- Regional or colloquial spellings
- Frequent misspellings and typos
The AI agent checks what the citizen says against that list, and if it finds a match, the case is routed straight to that institution's queue. If it doesn't, the case is flagged rather than guessed at.
That works well for ten institutions. The real test is whether it still works cleanly at thousands – and it's exactly what this is built for. A regulator's list of institutions can run into the thousands without the underlying routing logic changing at all. Bringing a new institution onto the platform, or adding it to an existing regulator's coverage, is a matter of adding one more entry to the list – not rebuilding how complaints get matched and routed.
One inbox, full visibility
Routing a complaint correctly is only half the job – a regulator also has to be able to see what happens to it afterwards. The same inbox gives regulators:
- Every chat and ticket fully logged and available for review at any time, with nothing routed into a black box on the institution's side
- The ability to engage the citizen and the responsible institution directly from within the same conversation
- Visibility of the full exchange as it happens, rather than having to reconstruct it afterwards from separate records
No repeating yourself, even across institutions
Complaints don't always start with the regulator. Increasingly, a citizen's first contact is with their own bank's or provider's AI agent – and only escalates to the regulator once that conversation stalls.
We've built this handoff directly into a number of our live regulator deployments. When a citizen using an institution's own AI agent indicates dissatisfaction or wants to file a formal complaint:
- The conversation transfers straight into the regulator's system
- The full context of the exchange transfers with it
- The case enters the regulator's official workflow immediately, with nothing lost in translation between systems
The citizen never has to explain their problem twice, and the regulator never receives a case stripped of the detail that led to it – which matters when that detail is what determines whether a case needs simple routing or something closer to mediation.
When a case needs mediation, not just routing
Not every complaint resolves once it reaches the right institution. Some are disputed – the institution's account of events doesn't match the consumer's, or only partial resolution has been reached – and these need a third party to weigh both sides fairly rather than simply route the case and wait.
This is the mediation capacity built into the inbox. Within the same conversation, the AI agent:
- Reviews the full case record from both the consumer and the institution
- Surfaces the relevant regulatory obligations and precedent
- Proposes a resolution for the human reviewer to consider
A live person still decides what actually gets sent. It's designed to stop disputes stalling indefinitely – historically one of the weakest points in first-generation complaints systems – and it sits in the same inbox as everything else described here, rather than a separate tool the regulator has to learn.
Deadlines that enforce themselves
Routing a complaint correctly and giving the regulator visibility only carries weight if institutions are actually held to a timeline. A resolution window can be built directly into the platform and applied automatically across every institution a regulator supervises, giving the regulator:
- Automatic escalation the moment an institution misses its deadline, with no citizen having to chase it
- Every response logged and every breach surfaced without manual tracking
- A permanent, auditable record of how each institution actually handles consumer redress – not just how it says it does
From individual cases to sector-wide oversight
Because every complaint is captured as structured data – the institution, the issue type, the product, the outcome – rather than free text, regulators can look across cases rather than only at them one at a time. Regulators use this to:
- Analyse complaints by product, issue type, value band, age, and sex, to see where consumer harm is concentrated and how it's shifting over time
- Identify emerging patterns across institutions
- Compare issue trends market-wide
- Prioritise supervisory attention where the evidence points
For a regulator, this turns complaint handling from a purely reactive function into something closer to an early-warning system – the same data that resolves an individual citizen's case is what tells the regulator which institutions, products, or regions need closer supervisory attention next.
Meeting citizens in their own language
None of the above matters if citizens can't get their complaint into the system in the first place.
Our AI agents are built to operate in the local languages citizens actually use – languages like Kinyarwanda, Tagalog, and Taglish, alongside English and French – with models trained for the specific linguistic and financial context of each market rather than a translated, off-the-shelf system.
This isn't a nice-to-have layered on top. Accurate institution matching, structured data capture, and enforced timelines only reach the citizens they're meant to protect if the complaint can be filed accurately in the language and phrasing people actually use, including the rural and lower-literacy users who are often least served by formal, English-only channels.
Why this matters specifically for regulators
A bank or insurer managing its own AI agent only ever has one "institution" to worry about. A regulator is different – its list of teams is the entire sector it supervises, and that list only grows as more institutions come online.
Getting a complaint to the right institution the first time, without a human triaging it, is what lets a national complaints system take on its thousandth institution as easily as its first – and having the entire lifecycle of a case, from intake through resolution, mediation, and escalation, sit in one inbox with structured, sector-wide data behind it, is what lets the regulator actually supervise that system rather than just receive reports from it.
This same approach underpins every regulator and supervisory body we work with, and it's the model we bring to new ones as they come on board.
Where we've deployed this
- National Bank of Rwanda – automated consumer protection across more than 600 financial institutions, with a 15-business-day resolution window enforced automatically.
- Bangko Sentral ng Pilipinas – a single channel covering roughly 2,500 supervised institutions, handling 14.4 million interactions a year and resolving 89% without human intervention.
- Bank of Mozambique – a structured complaint intake flow that identifies the responsible financial institution automatically and generates a trackable case, replacing a paper-based process that previously took months.