Solutions / Front desk
Customers get answers from your own material, without waiting for your team.
A chatbot that answers from your own manuals, policies and product pages, shows where each answer came from, and hands over to a person when it should.
What outside research says
14% more customer issues were resolved per hour when 5,179 support agents were given an AI assistant. Newer agents gained the most.
Source: NBER, 2023 (opens in a new tab)
Findings from the research linked under each figure. They are not Yielde results.
The loop
How it works
From a question to an answer with its source
Six steps, from the first event to the result. Choose a step to read what happens in plain words, how it is built, and what keeps it right.
Step 1 of 6
A customer asks
In plain words
Someone wants to know your returns policy, how to reset a device or whether you deliver to their town, and asks at nine at night instead of waiting for Monday.
Safeguard The customer is told at the start that it is an AI assistant, and can ask for a person at any point.
Under the hood
Each channel is built to open a conversation with an id that ties every message together, so the bot reads the whole exchange, not only the last line. The customer is told at the start that they are chatting with an AI assistant and that they can ask for a person. Each conversation has a rate limit, a cap on how many messages it takes in a short time, so a script hammering the chat cannot run up costs or flood your team.
Works with
- Your website
It checks what is asked
In plain words
Before it answers, it decides whether this is something it should answer at all. A question about your products is fine. A request for legal advice, or for another customer’s details, is not.
Safeguard Out-of-scope and unsafe requests are declined with a short reason and an offer of a person. It does not improvise outside its brief.
Under the hood
A first pass sorts each message into one of a few routes agreed with you: answer from your material, hand to a person, or decline politely. The routes are fixed in code, and the language model only picks one. An attempt to talk it out of its brief, such as "ignore your instructions", is read as an ordinary message and declined. A question that needs to know who the customer is goes to a person, unless identity checks are built in.
Needs a person: goes straight to step 05, Your team steps in.
It finds the right passages
In plain words
It looks through your own material, the way you would flip to the right page of the manual, and pulls out the few paragraphs that answer the question.
Safeguard It searches only the material you approved for it. A folder you left out is never in the index, so it cannot be quoted.
Under the hood
Your documents are split into passages and indexed two ways: by meaning, so "my parcel never came" finds the delivery policy, and by exact words, so a model number or product code is found too. Each passage keeps its document, section and version date. Only the folders, pages and entries you approve go into the index. When a document changes, its passages are replaced and the old version stops being used.
Works with
- Help centre
- Google Drive
- Your website
It answers with a source
In plain words
The customer gets a short, plain answer with a link to the page it came from, so they can read the full policy themselves. If your material does not say, it says so.
Safeguard No source, no answer. If your material does not cover it, the bot says so and offers a person.
Under the hood
The language model writes only from the passages it was given, and each claim carries a citation to its passage. A second check compares the answer with those passages before it is shown, and a sentence they do not support is removed. When nothing relevant is found, or two passages disagree, it says it does not know and offers a person, rather than filling the gap from general knowledge.
Works with
- Your website
Answered from a source: goes straight to step 06, You see what customers ask.
Your team steps in
In plain words
When it cannot help, or the customer asks for a person, your team gets the chat with a short summary on top, so nobody asks the customer to start again.
Safeguard Once a person has the conversation, the bot does not answer in it again until your team hands it back.
Under the hood
The hand-over carries the full thread, a short summary, the passages it found and why it stopped: no source, a request for a person, an upset customer or a sensitive topic. Outside your hours the customer is told when a person will reply instead of being left waiting. Your team answers on the same channel, and the bot stays out of that conversation until it is handed back.
Works with
- Help centre
You see what customers ask
In plain words
On the schedule you choose, you see what people asked most, which questions it could not answer and which pages left them confused. That tells you what to write next.
Safeguard Nothing joins the material it answers from until a person approves it.
Under the hood
Every conversation is logged with the passages used and how it ended: answered, handed over or declined. Questions are grouped by meaning, so the many ways of asking about delivery count as one topic. A question your team answered after a hand-over is drafted as a new help entry, and it joins the index only after a person approves it.
Works with
- Help centre
- Google Drive
What it knows
Only your own material: the help centre, the website pages and the Drive folders you choose, plus past tickets turned into help entries you approve, so no customer details carry over. Each answer is built from passages of that material with the source shown, and the index is updated when a document changes.
What it may do on its own
Answer from that material with a link to the source, ask a clarifying question, decline out-of-scope or unsafe requests, hand a conversation to your team, and group questions for the report.
What waits for your yes
Anything it has no source for, any request for a person, refunds, complaints, account changes and anything that needs to know who the customer is. Every new or changed answer, before it joins the material it answers from.
How we test it
Before launch it is run against a test set of real and awkward questions: the question your material does not cover, two pages that disagree, a request to ignore its instructions, an angry customer. Each answer is checked against its source, and the set is replayed after every change to the material or the brief.
What we measure
Once it is live, these are tracked and reviewed with you.
- Questions answered from your own material
- Hand-overs, and the reason for each
- Questions it could not answer, by topic
What you get
- A retrieval system built on your own documents
- Website widget and WhatsApp channel
- Hand-over to your team with the full thread
- A report of the questions customers ask most
Works with
- Your website
- Help centre
- Google Drive
Who it is for
Businesses whose team answers the same questions every day.
Built to order. We scope it with you on a call, price it before any work starts, and run it in draft mode until you trust it.
Related
- AI receptionist
Never miss a lead again.
Never Miss a Lead is one agent built to answer your calls, WhatsApp and email while you are on the job, take the details and line up the booking for your yes.
Waitlist open - Booking agent
Bookings made and kept, without the phone tag.
An agent that offers customers real open slots, books the one they pick, sends the reminders and fills cancellations from a waiting list.
Built to order - Email response agent
An inbox sorted and replies drafted by the time you sit down.
An agent that reads every incoming email, sorts it, and drafts the reply in your voice using your prices, policies and past answers.
Built to order