A voice agent for the 106 municipal hotline: how to test if it fits

Note: general recommendations only, not professional advice. You are responsible for checking before use. The full text is at the end.

Who this is for

For the manager of a 106 hotline (the municipal call center residents call about local services), a head of resident services, a director of operations, and whoever coordinates the topic for the authority's management. You don't need a technical background. You need four things: two months of call data from the phone system or the ticketing system (or, if you don't have it, a manual sample from one busy week), one person to lead the process, management approval to start testing, and two to three hours a week for six weeks. In step 4 you'll also need a meeting with the authority's legal counsel and its information security officer.

The problem in three lines

A large share of calls to the hotline repeat: green waste pickup, a property tax question, what's happening with a ticket I opened. At peak times residents wait, call again with the same question, and the call-center staff are left with no time for the calls that really need a person. On the other side there's a fair worry: who listens to the calls, what's recorded, and what happens when an agent gives a wrong answer about a dangerous hazard or a sensitive call.

The method in 6 steps (six weeks)

Step 1 (week 1): Map the calls into three groups

What to do: Take the call and ticket report for the last two months and split the calls into categories. If you have no categorized report, sample a few hundred calls from one busy week and categorize them by hand. The goal is to know what people actually call about, not what we think they call about.

How: For each category, write down how many calls there were, how long such a call takes on average, and at what hours it comes in. Then assign each category to one of three groups:

• The agent answers on its own: information questions that have a written, approved answer. Opening hours, department addresses, pickup days, a general explanation of how to apply for a property tax discount.
• The agent collects details and hands off: opening a ticket about a non-urgent hazard, checking status by ticket number, a request for a callback. The agent asks, records and passes it to the ticketing system or a staff member, and decides nothing itself.
• Straight to a person: emergencies and dangerous hazards, welfare calls, calls involving minors or violence, complaints about staff, and anything that needs judgment.

You can use an AI tool to categorize the report (prompt 1). Before you paste it in, delete names, phone numbers, addresses and ID numbers. The category, the time and the call length are enough.

Common mistake: Starting with the category everyone talks about in meetings rather than the one that actually repeats most. The ten most common questions are usually where an agent saves the most staff time.

Step 2 (week 2): Write a one-page boundaries document

What to do: Before you meet vendors, write a short document that defines exactly what the agent is allowed to do. The hotline manager signs it, and legal counsel reviews it. This is the document you'll check every vendor and every answer in the pilot against.

How: The document is a closed list. Anything not on it goes to a person.

Rules worth starting with:

• Emergencies first: When a resident describes danger (fire, a burst pipe, a fallen power pole, a threat to life), the agent stops asking questions. It transfers to a staff member immediately and, if needed, points to the national emergency numbers.
• Sensitive calls go to a person: welfare, minors, violence, bereavement. The agent doesn't handle them. It says one polite sentence and transfers.
• Not sure? Transfer: If the agent doesn't recognize the topic or isn't sure of the answer, it transfers. An unnecessary transfer is better than a wrong answer.
• No promises and no decisions: The agent doesn't promise a handling date, doesn't interpret municipal bylaws, doesn't approve discounts and doesn't decide on behalf of the authority.
• No reading out personal data: The agent doesn't say a resident's personal details out loud, such as a debt or an address, even if it has access to them. The resident may be talking from a public place.
• A person is always available: At any point the resident can say "agent" or press a key and reach a staff member.

A template for the document is in the templates section below, and prompt 2 helps write a first draft from the call map.

Common mistake: Writing "the agent handles everything except a few cases." A list of what's forbidden will always have gaps. Write a list of what's allowed, and everything else goes to a person.

Step 3 (week 2): Scenarios and approved wording

What to do: For each category in the "the agent answers on its own" group, write the answer in advance and get it approved by the relevant department and by the spokesperson's office. The agent only says wording from this list. For each category in the "collects and hands off" group, write which details it asks for, in what order, and where they go.

How:

1. Choose the six to eight top categories from the call map. No more at this stage.
2. For each one, write an answer of three to five sentences, in residents' words rather than the authority's internal terms.
3. Add a follow-up line to each answer: what to tell the resident if the question wasn't resolved, and which channel to use.
4. Get the wording approved by the department and the spokesperson's office, and record who approved it and when. Every future change goes through the same route.
5. Decide on languages. Hebrew is the base. Other languages, such as Arabic, Russian or Amharic, come in only after you've checked yourselves, with native speakers, how the agent understands and speaks them, in real calls and not in a demo.

Common mistake: Letting the agent write answers on its own from the authority's website. Website pages go out of date, and a resident will get old information in a confident voice. Only approved wording, with an approval date.

Step 4 (week 3): Privacy, recordings and questions for vendors

What to do: Before a single resident talks to the agent, settle in writing what is recorded, what is kept, for how long, where, and who may access it. In the same week, send vendors one standard list of questions, so you compare them on the same basis. Recorded calls and transcripts are personal data. What the law requires of you, including what changed with Amendment 13 to the Privacy Protection Law, you check with the authority's legal counsel and information security officer. What you'll find here are questions, not answers.

How: Questions to bring to the meeting with legal counsel and the information security officer:

• What recording notice wording do we need at the start of the call, and what happens if the resident doesn't agree to be recorded?
• For what purposes may we use recordings and transcripts: quality control, improving the wording, anything else?
• How long will we keep recordings and transcripts, and who deletes them at the end of that period?
• Who at the authority may listen to calls and read transcripts, and is every access logged?
• May the data be stored with an outside vendor, and if so, under what conditions and in what location?
• May the vendor use our calls to train or improve its models?
• Which of the authority's internal procedures apply here: procurement, information security, databases, accessibility?

The vendor questions are in the template in the templates section. Pay particular attention to five topics: where the data is stored, what's recorded and for how long, Hebrew quality (including accents, background noise and fast speech), exactly how handoff to a person works and what the staff member sees at the moment of transfer, and what is logged for each call. Ask every vendor for written answers, not only answers in a meeting. Prompt 6 helps compare the answers.

One more thing that's easy to forget: accessibility. Some residents can't or don't want to talk to an agent: older people, residents with hearing or speech disabilities, people who speak a language the agent doesn't understand well. Make sure there is always a quick way to reach a person, and that the other channels (a staff member, a written message, a web form) keep working as usual. Check the binding accessibility requirements with whoever is responsible for accessibility at the authority.

Common mistake: Putting privacy off until after launch. Once recordings sit with a vendor, it's very hard to change where they're stored and who accesses them. Settle it before the first call.

Step 5 (weeks 4 to 5): A closed, monitored pilot

What to do: Run the agent first on an internal test line, and then, if all is well, on a defined window of hours with residents, with a person monitoring every day. Don't open it to all calls at once.

How:

1. Week 4, internal line: Hotline staff and volunteer employees from other departments call in and run test scenarios written in advance (prompt 5). Include hard ones: someone speaking fast, background noise, a confused question, and emergency and sensitive scenarios.
2. Week 5, limited window: If the internal week had no failure on an emergency or a sensitive call, open a small window of hours to residents, for example low-traffic hours. Always with an immediate option to reach a staff member.
3. Daily review: One person listens to a sample of calls every day and checks every answer against the approved wording and the boundaries document.
4. Failure log: Every error is recorded: wrong categorization, an incorrect answer, an unnecessary transfer, a transfer that didn't happen. Date, scenario, what happened and what should have happened.
5. Immediate stop: One emergency call that wasn't transferred immediately, or one sensitive call the agent tried to handle, stops the pilot until you understand what happened and fix it.
6. End-of-week fixes: Fix wording and boundaries based on the log. Wording whose content changed goes back for approval.

Common mistake: Comparing the pilot to the vendor's demo. The demo runs on questions prepared in advance. The pilot tests what happens when a real resident talks the way they talk.

Step 6 (week 6): Measure and decide

What to do: At the end of the pilot, give management one page with the measurements, what went wrong and what was fixed, and one recommendation: widen gradually, fix and run another pilot, or stop.

How: The measurements. Measure all of them yourselves and compare them to the situation before:

• Resolved without transfer: how many calls in the scenarios assigned to the agent ended without a staff member.
• Categorization accuracy: how many calls were assigned to the right group, including those correctly transferred.
• Wrong transfers: calls transferred even though the agent was supposed to handle them, and calls not transferred even though they should have been.
• Content errors: answers that didn't match the approved wording. The goal here is zero.
• Emergencies and sensitive calls: how many came in, and how many were transferred immediately.
• Wait for a staff member: how long residents who reach a person wait, before the pilot and during it.
• Repeat calls: how many residents called again about the same topic within a short time.

Set the thresholds before the pilot, not after. Write down in advance what has to happen to widen, and what stops you. Widening, if you decide on it, goes in stages: more hours, then more categories, and only at the end more languages. Prompt 7 helps summarize the pilot log.

Common mistake: Deciding based on how the first week felt, good or bad. In the first week the wording is still settling and the team is still learning. Decide by the thresholds you set in advance.

Templates

Replace whatever is in square brackets.

Boundaries document: voice agent at the 106 hotline of [authority name]
Version: [number] · Date: [date] · Approved by: [hotline manager], [legal counsel]

1. The agent answers on its own, only with the approved wording:
   [category] · [category] · [category]
2. The agent collects details and hands off, without deciding:
   [category]: asks for [details], hands off to [ticketing system / staff member]
3. Straight to a person, with no further questions:
   emergency and dangerous hazard · welfare · minors · violence · complaint about staff · any topic not listed in sections 1 and 2
4. Always forbidden:
   promising a date or outcome · interpreting bylaws · approving a discount or exemption · reading out personal data · giving an answer that isn't on the list
5. Not sure: transfers to a person.
6. The resident can ask for a staff member at any point: [word / key].
7. Recording notice at the start of every call: [wording approved by legal counsel]
8. Changes to this document: only with approval from [role], with a new version and date.

Questions for a voice agent vendor: [vendor name] · Date: [date]
Please answer each question in writing.

Data and privacy
1. Where are recordings, transcripts and call data stored? In which country and with whom?
2. What exactly is recorded, and for how long is it kept? Do we set the retention period?
3. Do you or your subcontractors use our calls to train or improve models?
4. Who on your side can access calls, and is every access logged?
5. How is data deleted on request, and how do we get confirmation that it was deleted?

Quality
6. How does the agent handle spoken Hebrew, accents, background noise and fast speech? Can we test it on our own recordings?
7. Which other languages are supported, and how do we test quality in each one?
8. How do you make sure the agent only says approved wording and doesn't improvise?

Handoff to a person
9. How does a resident reach a staff member, and how long does it take?
10. What does the staff member see at the moment of transfer: transcript, summary, topic?
11. What happens when no staff member is available?
12. How is an emergency detected, and what happens the moment it is?

Logging and review
13. What is logged for each call, and how do we receive it?
14. What reports will we get, and how often?
15. How is wording updated, and who confirms the update went live?

Authority and accessibility
16. Have you worked with local authorities' procurement and information security procedures, and what documents will you provide for our review?
17. How does a resident who can't use the agent (hearing or speech disability, language, age) reach a person quickly?
18. What happens when your service is unavailable? Where do calls go?
19. How does the engagement end, and what happens to our data the day after?

Ready-made prompts

Replace whatever is in square brackets. Before you paste in reports or transcripts, delete names, phone numbers, addresses and ID numbers.

I manage a 106 hotline at a local authority. Here is a two-month call report, with no identifying details: [paste: category or short description, time, length].
Assign each row a category and one of three groups: "the agent answers on its own" (general information with a written answer), "the agent collects and hands off" (non-urgent hazard ticket, status, callback request), "straight to a person" (emergency, welfare, minors, complaints, judgment).
At the end, return a table: category, number of calls, average length, peak hours, group. If there isn't enough information to assign a row, write "unknown" and don't guess.

Here is the hotline's call map table: [paste].
Write a draft one-page boundaries document for a voice agent at a 106 hotline: what the agent answers on its own, what it collects and hands off, what goes straight to a person, and what is always forbidden.
Write it as a closed list of what's allowed. Anything not on the list goes to a person. Include rules for emergencies, sensitive calls, uncertainty, personal data and requests for a staff member. Short sentences. At the end, list what you didn't have enough information to decide.

Write an answer of three to five sentences for the scenario [scenario name] at a local authority's 106 hotline.
Use only this information from the department: [paste].
Plain words, no internal terms, no promise of a date or outcome. At the end, a follow-up line: what to tell the resident if the question wasn't resolved, and which channel to use.
If the information I gave isn't enough for a full answer, don't fill it in from your head. Write what's missing.

Write three versions of an opening message for a voice agent at the 106 hotline of [authority name].
The message needs to say: that the caller is talking to an automated agent, that the call is recorded, for what purpose [purpose], and that they can ask for a staff member at any point by saying [word] or pressing [key].
Up to three short sentences in each version. This is a draft for legal counsel to review, not final wording.

Here is the scenario list and the boundaries document: [paste].
Write 20 test scenarios for a voice agent pilot: 8 ordinary questions, 4 with missing or confused information, 3 that must go to a person, 3 emergency scenarios and 2 where the resident asks for a staff member midway.
For each scenario: what the resident says (in everyday spoken language, the way people talk on the phone), what the correct outcome is, and what counts as a failure.

Here is the list of questions we sent vendors, and written answers from two or more vendors: [paste].
Build a comparison table by question. For each question: what each vendor answered, in one line.
Mark questions a vendor didn't answer or answered too generally. At the end, a list of follow-up questions for each vendor. Don't rank and don't recommend a vendor, only compare.

Here is the pilot log, with no identifying details: [paste rows: date, scenario, outcome, transferred to a person yes/no, answer matched the wording yes/no, note].
Baseline before the pilot: [wait time for a staff member, repeat calls].
The thresholds we set in advance: [paste].
Summarize by measurement: resolved without transfer, categorization accuracy, wrong transfers, content errors, emergencies and sensitive calls. For each one: the number, against the threshold. At the end, one recommendation: widen, fix and run again, or stop, with a short reason. Don't round numbers in your favor.

One-page checklist

Week 1: Mapping

• Two months of call data, or a one-week sample, with no identifying details
• Every category assigned to one of the three groups
• Six to eight top categories chosen

Week 2: Boundaries and wording

• Boundaries document written as a closed list, signed by the hotline manager
• Emergencies and sensitive calls go to a person immediately
• A staff member can always be requested
• Approved wording for each category, with a follow-up line, who approved it and when

Week 3: Privacy and vendors

• Meeting with legal counsel and the information security officer, with the question list
• Recording notice wording approved
• Retention period and list of people with access set
• Every vendor answered the same question list in writing
• A quick route to a person for residents who can't use the agent

Weeks 4 to 5: Pilot

• An internal week with test scenarios, including emergencies
• A limited window of hours for residents, only after a clean internal week
• Daily review and a failure log
• An immediate stop rule, written down and known to the team

Week 6: Decision

• Thresholds set before the pilot
• One page of measurements for management
• A written decision: widen gradually, fix and run again, or stop

How you know it's working

Record the baseline before the pilot, and compare against it:

• Emergencies: every emergency call and sensitive call was transferred immediately. One case that wasn't transferred halts widening.
• Content errors: no answers that depart from the approved wording. If there are, understand why before you continue.
• Resolved without transfer: within the categories assigned to the agent, the share resolved without a staff member rises week by week.
• Wait time: residents who need a person reach one faster than before, not slower.
• Repeat calls: fewer residents call again with the same question.
• The staff: they report getting more calls that really need them, and fewer questions about hours and addresses.
• Decision: at the end of six weeks there's a decision based on numbers. "We stop" is also a good outcome, because it saves the authority an engagement that wouldn't have worked.

Disclaimer

This content is general recommendations only, not professional, legal, accounting or tax advice. Every authority is different. You are responsible for checking that it fits your situation, and for checking the templates, prompts and outputs before use. The questions about privacy, recordings and data retention are points to check, not legal statements. For questions about call recording, personal data and privacy protection, accessibility and resident service, consult a lawyer. Before approaching or engaging any vendor, check the procurement and tender rules and the internal procedures that apply to you with the authority's legal counsel. Outside tools, prices, privacy settings and terms of service change, so check them at the source.