What an AI agent solves for an authority - beyond "resident service"
Evidence-based research from Israel and abroad: which real problems an authority pays for today, and what already works elsewhere. Every number on this page carries its source.
Executive summary
Authorities buy "resident service". That is not what costs them money. What costs money is residents not claiming discounts they are entitled to, staff hunting for information the authority already holds, and nobody knowing what residents actually ask. You can put numbers on those three. You cannot put a number on "resident service".
The eight problems a capable agent solves
| The problem | Who suffers | The evidence |
|---|---|---|
| Benefits go unclaimed - eligible people do not receive what is theirs | Residents and the authority | 41% of public complaints about property tax to the State Comptroller in 2023 - 2025 came from eligible people who had not received discounts and exemptions (State Comptroller and Ombudsman, February 2026) |
| Contact-centre load - the same questions, around the clock | Contact-centre staff | Singapore, Ask Jamie: roughly 50% of load taken off the centres |
| The authority's own knowledge is not accessible to its own staff | Staff | UK: 20,000 civil servants, an average saving of 26 minutes per person per day |
| Licensing and permitting bottlenecks | Businesses and the city engineer | Honolulu: application review from 60 - 90 minutes to 15 - 20. LA and Austin: up to 55% reduction |
| The authority does not know what residents actually need | Management | NLP classification of 311 requests reaches over 83% accuracy |
| Wrong or outdated content on the website - unnoticed | Residents and communications | New York: the city bot advised businesses in ways that contradicted the law, because content was not controlled |
| Language and accessibility gaps | Vulnerable populations | Research: language barriers and digital literacy are the leading reported barrier in digital government services |
| Contradictory information, by definition | Residents and trust | A former office-holder appears in dozens of minutes that may not lawfully be removed; the current one appears in two |
Why a general chatbot is not enough
This is the first objection every authority raises, and it has four evidence-backed answers rather than an opinion.
1. Answer quality on government questions has been measured - and it is not sufficient
- Tow Center, Columbia University (2025): 8 AI search engines, 1,600 queries - over 60% incorrect answers. ChatGPT Search: 67% errors
- EBU and BBC (October 2025): 22 public broadcasters, 18 countries, over 3,000 answers - 45% contained at least one significant problem; 20% serious factual errors, including outdated information
- ODI, CitizenQuery-UK (2026): over 22,000 citizen queries against official answers - broadly reasonable accuracy, but a long tail of failures: wrong benefit eligibility, a court order claimed to be required when it was not
2. The most troubling finding: they answer even when they cannot
ODI researchers called this absence of refusal "a dangerous property"; the Tow Center highlighted how confident the tools are when wrong. A resident asking "am I entitled to a discount?" gets a confident, well-written answer that is sometimes wrong. The error is attributed to the authority. An agent that can say "I do not have that, contact department X" is worth more than one that answers everything.
3. A general chatbot cannot see the authority's documents
- Scanned PDFs - minutes, by-laws, plans - are usually not read at all
- Documents behind forms, archives and internal systems are out of reach entirely
- Scanned Hebrew documents require OCR, a step a general chatbot does not perform
- Internal procedures will never be accessible to a public chatbot - and must not be
4. An authority holds contradictory information by definition
The phenomenon has an academic name - knowledge conflict and temporal validity. Successive office-holders appear across dozens of historical documents that may not lawfully be removed, while the current incumbent appears in few. A general chatbot answers according to what is common in the documents, which means a wrong answer given with full confidence. A dedicated agent needs a layer that decides what is valid today.
What already works elsewhere
At first, bots only answered questions. Today they complete processes, and help the authority own staff as well. Anyone still offering only an FAQ is offering a 2019 product.
| System | Who | What it does beyond FAQ | Reported result |
|---|---|---|---|
| Boti | Buenos Aires, since 2019, WhatsApp | Completes processes, medical triage, events, parking | About 2 million contacts a month, 82% resolved automatically |
| Ask Jamie | Singapore GovTech, 80 government sites | Routing and cross-agency answering | Over 15 million contacts, about 50% reduction in centre load |
| Bürokratt | Estonia | A network of cross-ministry agents - coordinates with another body instead of referring you to it | The world's most advanced pilot in inter-agency coordination |
| Tel-Ma | Tel Aviv-Yafo Municipality | Website, "My Digital", WhatsApp 106, and agent assist | Part of 15+ municipal AI applications |
| HerzlAI | Herzliya Municipality | Municipal information, education, transport, payments | Presented as a digital gateway, not a chatbot |
| Extract | MHCLG, planning authorities in England | Turns planning documents into data | From two hours to about two minutes |
| CivCheck / Archistar | Honolulu, Los Angeles, Austin | Pre-submission compliance checking | 60 - 90 minutes to 15 - 20; up to 55% reduction |
| M365 Copilot | UK government, San Francisco | Internal assistant for staff | 26 minutes per person per day, across 20,000 participants |
Standards and regulation already in force
| Standard | What it requires | Relevance to an authority |
|---|---|---|
| EU AI Act, Article 50 | From 2 August 2026: people must be told they are talking to an AI; generated content must be machine-readably marked | Even outside the EU, this is already the expected bar in tenders |
| ISO/IEC 42001 (December 2023) | An AI management system - certifiable | The strongest tender differentiator; connects to an ISO/IEC 27001 pack |
| NIST AI RMF | A voluntary AI risk management framework | The language in which control documents are written |
| UK ATRS | Transparent recording of algorithmic tools; mandatory in government departments from 2025 | A transparency card for an authority is a proactive move ahead of regulation |
| Accessibility - WCAG and Israeli standard 5568 | Digital accessibility is mandatory in Israel | An inaccessible widget is the authority's legal exposure |
Documented mistakes not to repeat
- New York, MyCity: a city bot advised businesses in ways that contradicted the law, because content and restraint were not controlled
- Rotterdam and the Netherlands: an algorithm that scored people for welfare enforcement. Never score or filter people
- Measuring deflection instead of value: fewer contacts is not necessarily an improvement - a resident may simply have given up
- Inferring policy from logs: propensity to contact differs between groups, so logs are a biased sample
- Digital-only service: research shows it widens the gap. A human channel must remain visible
What this means in practice
- Measure benefits take-up, not call volume. Easy, cheap outreach delivers almost all of the effect
- Do not measure conversion in the first week - a Spanish study across 400 neighbourhoods found a delayed digital effect
- Keep a visible human channel alongside the agent, always
- Decide what is valid today. In authority documents, what is common is not necessarily what is correct
- Explicit disclosure that this is an AI, on every channel
About this research
This document summarises research conducted through three lenses - leading products and authorities, academic literature, and standards and regulation - drawing on roughly 45 public sources. The figures cited belong to the bodies that published them. No customer data of ours was used, and no customer is named.
2026-08-17
Talk to Us
Schedule a live product demo, ask a general question, or set up a business consultation. Pick what fits and we'll come back to you shortly.