The Redbo agent - how it works and what you control
An AI agent that answers whoever contacts you - a resident, a customer or an employee - on the channels they already use, only from content you approved, always with the source.
Where people meet you, and how it feels
Most bots feel like a form: a menu of buttons, a dead end, and back to the phone. Here people simply ask, on the channel already open in front of them, and get one answer that is the same everywhere.
Three channels, one answer
Website, WhatsApp and Telegram. What is said on one is said on the others, because there is one source and not three.
The answer is written in front of you
On the website the answer starts appearing straight away, instead of a blank screen that waits.
On WhatsApp it looks like WhatsApp
Reply buttons and list menus instead of a wall of text with long raw links.
Links you can actually read
The person sees the name of the page rather than an encoded address a line long, and on the website a preview of it too.
People can say the answer did not help
A thumbs-down, and if they want a short reason as well: wrong, incomplete, no link, not understood. It comes back to you as a work list rather than disappearing.
Reaching out, not only answering
You can send a message to an audience that opted in. Anyone who wants to stop writes one word, and it stops.
The result: someone who arrives with a question does not have to learn how to talk to the system.
What if the agent gets it wrong?
This is the first concern of anyone responsible for content the public relies on - rightly so. Our answer is not a promise but a structure you can inspect.
Answers only from what you approved
The agent does not draw on the open internet or general model knowledge. The sources are your documents and your site, and whatever you chose to put in.
Every answer shows its source
The person who asked sees where the answer came from and can open it. So can you - when someone complains, you can check exactly what was said and on what basis.
A link, a phone number or an address only if it exists in the source
Every link, phone number and email address in an answer is checked against the source the answer was built from. A value that is not there is simply not shown.
For common questions, you set the wording
For the questions that repeat most, you approve in advance exactly what is said. No surprises where it matters most.
When there is no answer, it says so
With no verified information the agent says it does not know and points to the contact route you configured: a phone number, a form or an email. That wording is a field you write. It does not guess to look helpful.
Answer quality is measured, not assumed
A set of questions your team approved is asked of the live agent again and again, and the score is kept over time. A drop in quality shows up before anyone complains.
The result: you can point at any answer the agent gave and show where it came from.
Any language - in wording you approved
Most platforms offer "75+ languages". It sounds like more, and it usually means machine-translating the reply in real time. For an organisation whose people speak Arabic or Russian - an authority, a utility, or a company with foreign staff - that difference is substantial.
The content itself exists in each language
Not an automatic translation of a Hebrew answer, but content that exists in Arabic and Russian and that you approved.
Wrong wording - you fix it, and the fix stays
If something reads wrong in Arabic, you correct it once and it holds, with a history of who changed what. In a system that translates on the fly there is nothing to correct - it will translate afresh next time.
The admin interface is translated too
Not only what the person asking sees. Whoever runs the system works in their own language - the admin interface exists today in Hebrew, English, Arabic, Russian, Spanish and French.
Hebrew, English, Arabic, Russian, Spanish and French
Switching on another language is your decision, not a project. The agent answers in six today - the ones our customers actually need, rather than a list designed to impress. Each one is switched on per customer, for their audience, so not every deployment shows all six.
Hebrew is not an add-on
The system is chosen and tested on the quality of its Hebrew, including spelling variants, abbreviations and sector terminology. A product built in English and translated is exactly where that breaks.
Accessibility is checked on every build
The chat window is checked automatically against WCAG 2.1 AA rules, the basis of Israeli Standard IS 5568: a name on every control, full keyboard operation, focus returned on close, and the reply announced to a screen reader. This is automated checking with partial coverage - not a full accessibility audit and not a declaration of compliance.
In practice: a language is switched on when you need it, and the content in it is approved - instead of broad coverage nobody checked.
Who controls what is said in your name
An agent answering on your behalf is a spokesperson. Control of it belongs with you, not with the vendor - and so does the ability to act, not only to ask.
Full control of content and answers
What goes in, what comes out, and what is said on sensitive topics - all yours, from the admin interface.
Permissions by role and department
Every person and every screen gets one of three levels: hidden, view, or edit. A department sees its own. Build a permission template once and apply it to whoever joins after.
Sign in with your own organisational identity
Anyone already managing users in the organisation does not need to manage another list: sign-in with the existing corporate account, plus two-step verification. A sensitive action asks for verification again, and repeated failed sign-ins lock an account progressively.
Design and branding in your hands
Colours, logo, name and avatar, with a live preview before you publish. The logo is picked up automatically from your website, and you can replace it.
Complete chat history
Every conversation is kept and searchable by date, channel, topic, and whether it was answered. When a complaint arrives, there is something to check.
Audit log
Who changed what, and when, on every action that changes something. This is what your security officer will ask to see.
The content does not go stale
The reasonable fear is an agent answering from a page replaced a year ago. So updating does not depend on anyone remembering - and neither does removal.
Scheduled sync
Your site and documents are re-read on their own, without anyone starting it. Scanned documents are read too.
Immediate update when needed
An urgent change does not wait for the next cycle - you can push it now.
Expired content takes itself out of service
An answer whose validity has passed stops being served, and is reported to you in a digest. A floor guard stops the run if it is about to remove too large a share - so a faulty detection cannot empty a knowledge base.
A page that disappears takes its answer with it
When the source page stops responding over time, the answer built on it is withdrawn rather than left pointing at a dead link. One failed check is never enough.
Removal you can come back from
If the condition that removed an answer clears - the page is back, or the date is in range again - the answer returns on its own. Nobody has to remember to undo it.
Publish without touching the website
An urgent notice does not have to wait for whoever updates the site, and nobody has to decide which page it belongs on. You write the answer and whoever asks gets it immediately, on every channel.
What your people ask - that you did not know
This is the part most organisations do not expect. The agent does not only answer; it shows what the public is actually preoccupied with, and what is missing on your side.
What gets asked most
The real picture of what your people care about, not the assumption.
Where your site lacks information
Questions that repeated and went unanswered, clustered by topic and ranked - that is your website to-do list, and you can download it as a file.
Broken links on your own site
A list of links that no longer work and content that cannot be reached, as found in the scan. A page that looks dead is re-checked from a second vantage point before it is reported as one.
A weekly answer-quality report
A weekly report you can switch on: how many asked, how many were answered, how many answers carried a source, what came back unanswered, and on which channel. Our own test traffic is not counted in it, and that is stated.
Satisfaction with answers
Not only a thumb: the reason chosen, and the answer it refers to. Every item is closed explicitly - handled, dismissed, or needs a source.
Dashboard and Excel export
One picture for whoever has to report upwards, and every table downloads to Excel in one click.
Many organisations find this worth having on its own, regardless of deflection.
What your security officer will ask
This section is written for whoever has to approve this from an information-security standpoint - and it is based on things that can be checked, not on assertions.
Customer data separated at the system level
Every query is scoped to its own organisation. One customer's content cannot be returned in an answer to another.
Your data is not used to train models
Your content is used to answer your own users. It does not go to any model provider for training.
Identifiers masked before anything is stored
ID numbers, payment details and contact details are detected and masked before storage and before anything is sent to a model, from the structure of the number itself rather than a guess. A phone number is masked on the server before it reaches a screen, an export or a report. The agent also asks people not to enter such details in the first place.
Deletion by policy, with a legal hold
The default is to delete nothing. If you set a retention period it is enforced - and a customer under legal hold or review is skipped entirely. Every deletion is written to the log.
A customer file is never deleted
When the system "removes" a document it moves it into a dated archive rather than deleting it. This is not policy alone - a check stops code that deletes a customer file from reaching the product.
The sub-processor list is visible to you
Every third party that touches your data appears in one list you can see yourself. It is exactly the list a data-protection review asks for, kept in one place.
ISMS built to ISO/IEC 27001, and A+ on two external scans
57 written policies and procedures. There is no certificate - and we are glad to hand the pack over for review. The A+ grades are from Qualys SSL Labs and Mozilla Observatory, two third-party checks you can re-run in one click.
On the shared service: data in the EU, hot site in Israel
This is how we run the SHARED service - production in the European Union, a hot standby site in Israel, and backups in a third location. Three places, not one. On a dedicated deployment the whole infrastructure can sit in Israel, in the EU, or in both, as you choose.
What happens when something fails
A question asked too little before signing and too much afterwards. Here is what actually exists.
Monitoring that checks the channels themselves
Every few minutes the channels are exercised for real, not just the server pinged. A confirmed fault raises an alert immediately - without waiting for someone to complain.
Backups every six hours
Full, off the server, and verified that the files actually arrived.
A full restore drill - every week
Once a week the system is restored from backup into a separate environment, built, booted and made to actually answer - and the result is recorded. A backup nobody restores is a hope, not a backup. A hot standby environment in Israel also exists; cutting over to it is a manual step.
A new version ships without an outage
Versions are swapped while traffic keeps flowing. If something is wrong anyway, going back to the previous version takes seconds - and it is validated before it happens, so the rollback cannot extend the fault.
The channels do not fail together
When phone lines are swamped or the website is down, WhatsApp and Telegram keep working. In an emergency that is exactly the difference.
What it takes from you to start
This is the IT manager’s question, and the answer is deliberately short.
One tag on your site
A single line in your website code. No new server, nothing to connect, no project.
Nothing for you to maintain
Updates, monitoring and backups are on us.
A screen that shows how ready you are
Setup progress is measured from real state - is there content, are there approved answers, is a channel connected - rather than from a checklist someone ticks.
The content is yours
What you put in stays yours, and can be exported.
Where the answer came from
The page the organisation published, the line the answer was drawn from, and the wire between them.
Video transcript (text)
- A resident asks how to sign up for an activity.
- The Activities page appears beside the question, and the system scans it.
- The matching line lights up, and a wire is drawn from it to the answer being written.
- A source chip names the page and the section the answer came from.
- The sequence repeats for a sibling-discount question, and for a waste-collection question from a different page.
Golden answers - when the wording has to be exact
For sensitive questions you write the answer, and the agent delivers it word for word.
Video transcript (text)
- The conversations show which questions keep coming back.
- You write the approved wording and attach its source.
- The same answer is translated into every language you run.
- And the resident gets exactly what you approved.
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.