The problem
Not everybody is excited when they hear the word taxes … how can we make [it] easier, more welcoming, and more human?Nobody is pleased to hear the word taxes. The brief from Anguilla's Inland Revenue Department was blunter than that — when the only tax officer is busy, who picks up? The team's answer was not a better phone queue. It was a tax-time sidekick that never goes to lunch.
Anguilla's Inland Revenue Department published a brief with a staffing problem inside it. When the only tax officer is busy, the queue does not go away — it waits, and it grows, and the person at the end of it gives up and comes back tomorrow.
Island Waves was a joint team, drawn from Montserrat, Anguilla and Nevis, which meant they were building for a department most of them had never walked into. Hadassah Williams led the presentation.
What they built
A.I.D.A. runs inside a working clone of the official website of Anguilla's Inland Revenue Department (IRD), built so the team could prove embedded-widget deployment without touching the live platform. She opens by asking which language — English, Spanish or Mandarin — then answers the question everyone actually asks, in the dialect they ask it in, with a structured Universal Social Levy estimator: targeted follow-ups, an itemised estimate and a prominent legal disclaimer. She gives turn-by-turn directions to the IRD building and delivers fillable PDF returns inside the chat. Behind her sits a staff console the team nicknamed IRD's Kitchen — live monitoring, analytics, knowledge editing and human takeover.
- 3
- languages
- 70
- points, second place
- 2
- interfaces
- 24/7
- availability
English, Spanish and Mandarin
Chatbot Division
the public assistant and IRD's Kitchen staff console
against a department with one tax officer
What the judges pressed on
Dyson Lau, a Technical Officer at the Caribbean Institute for Meteorology & Hydrology (CIMH), asked two questions.
"Was the demonstrated website an existing site or built by you?" Built by them: a working clone of the IRD Anguilla website, made to prove embedded-widget deployment without touching the live platform.
"Do you see a commercial opportunity in fine-tuning regional dialect voice models?"
The turning point
They did. The team had used Gemini and Google AI Studio prompting to tune pitch and inflection until A.I.D.A. sounded authentically Anguillian — and argued that licensing Caribbean-accented voice models is a real commercial opportunity, not a nicety.
Boy, how much taxes me got pay now?Built with
- Gemini
- Google AI Studio
- RAG
The team
- Hadassah Williams — Lead Presenter
- Michaela Conner — camper
- Zenia Francis — camper
- Kai Gray — camper
- Khan Gray — camper
- Chayzan Green — camper
- Hallie-Jaiz Harris — camper
They did not fight the culture around tax. They spoke its language, down to the dialect of the question everyone was already asking.
Editorial review
Our reading of the build
An independent assessment by the editors of this publication — not the judges’ scoring, and not the team’s own account.
- Craft
- Building a clone of the client's own website to prove embedded deployment is the move of a team thinking about handover, not applause. It answers the question every client asks second — how would this actually reach our users? — before anyone has to ask it. The estimator is the other tell: it could have been a chat reply, and instead it is a structured flow with follow-ups, an itemised breakdown and a disclaimer where a disclaimer legally belongs.
- What they learnt, and where it shows
- Grounding is the thread. A tax estimate is the most consequential thing any bot at this showcase produced — get it wrong and someone underpays a government — and the team treated it accordingly, wrapping the estimator in mandatory disclaimers rather than letting a model speak freely about money owed. IRD's Kitchen is the human-handoff lesson made into furniture: the escalation path is not a promise in a slide, it is a screen somebody sits at.
- Usability
- Asking which language first, before anything else, is the right order for a trilingual population, and answering dialect in dialect is the single warmest decision in the whole field. The weakness is density: the staff console shows enquiries, answer rate, escalations and top topics on one screen, which serves an analyst better than it serves the officer with four minutes between callers. A first-run view of what needs me now would carry more of the load.
- Scalability
- Language coverage is the strength and the cost. Three languages is three knowledge bases to keep in step, and a mistranslated tax rule is worse than no translation — so the maintenance burden here is heavier than the feature list suggests. The widget architecture, though, is genuinely portable: any revenue department with a website and a document set could run this shape, which is what makes it interesting beyond Anguilla.
You've seen the customer side; now welcome to IRD's kitchen.Images captured at the 2026 ECCU Chatbot & Multimedia Showcase, 28 August 2026. Screen captures have been cropped to remove browser and streaming interface; the chatbot interfaces and the presenters are unaltered.