Tola AI
Tola AI answers customers from your own knowledge, drafts replies for your agents, and — when you let it — takes actions in your systems. It is not a generic chatbot bolted onto the side; it works on the same conversations, in the same inbox, under rules you set.
What Tola AI does
Three distinct jobs, and it is worth being clear about which you want:
- Name
Answer directly- Type
- autonomous
- Description
The assistant replies to the customer itself, on the channels you connect it to, and hands off to a human when it cannot help.
- Name
Assist the agent- Type
- copilot
- Description
It drafts a reply and shows the sources, and a human decides whether to send it. Nothing reaches a customer unreviewed.
- Name
Do things- Type
- tools
- Description
It calls out to your systems — look up an order, check a subscription, create a ticket — and uses the result in its answer.
Most teams should start with the second, move to the first for a narrow set of questions, and add the third once the first two are trusted.
Assistants
An assistant is a configured AI persona. You can have several — a support assistant, a sales assistant, one per language or per brand — each with its own knowledge, rules and channels.
- Name
Name and product name- Type
- identity
- Description
What it calls itself and what it calls your product. This shows up in its answers, so use the words customers use.
- Name
Temperature- Type
- how varied its wording is
- Description
Lower is more predictable and repetitive; higher is more natural and less consistent. For support, lower.
- Name
FAQ generation- Type
- toggle
- Description
Whether it learns question-and-answer pairs from resolved conversations.
- Name
Memory- Type
- toggle
- Description
Whether it remembers context about a contact across conversations.
- Name
Contact attributes- Type
- toggle
- Description
Whether it can read the customer's attributes — plan, order history — when answering.
Knowledge: documents and FAQs
An assistant is only as good as what it knows. There are two kinds of knowledge.
Documents are sources you point it at: a URL, your help center, a PDF you upload. Tola crawls and indexes the content, and the assistant answers from it. Documents can auto-sync, so when you update a help center article the assistant's answer changes with it.
Responses — question-and-answer pairs — are the precise version. Where a document says roughly the right thing, a response says exactly the right thing. Use them for the answers that must be word-perfect: refund policy, pricing, anything with a legal or financial edge.
Tola can suggest FAQs from your resolved conversations. It notices the same question being answered repeatedly and proposes a pair for you to approve, edit or reject. This is the fastest path to a useful knowledge base, because it is built from what customers actually ask rather than what you assume they ask.
Approve suggested FAQs deliberately. An approved answer becomes what the assistant tells every future customer — a rushed approval is a mistake at scale, not a mistake in one conversation.
Guidelines and guardrails
Response guidelines shape how the assistant writes: tone, length, whether to use the customer's name, when to apologise, which language to reply in. This is where you make it sound like your company rather than like a chatbot.
Guardrails define what it must not do. Topics it should never discuss, claims it must never make, things it must always escalate instead of answering. Discounts, legal advice, medical questions, anything touching money — put those behind a guardrail and let a human handle them.
Write guardrails before you turn the assistant loose on customers, not after something goes wrong. They are the cheapest safety measure available.
Scenarios
A scenario is a named situation with an instruction attached — a playbook for a specific kind of request rather than a general rule.
"When a customer asks about a delayed order: apologise, look up the order with the order lookup tool, give them the current status and the revised date, and offer to escalate if it is more than five days late."
Scenarios can be given access to specific tools, so the assistant does the lookup as part of following the playbook. They are enabled individually, so you can build one, watch it, and only then build the next.
Tools
Tools let the assistant reach outside Tola into your own systems. Out of the box it can search everything you have given it as knowledge; beyond that, your team can connect it to your order system, your CRM, your booking platform — anywhere the answer to a customer's question actually lives.
This is the difference between an assistant that explains your refund policy and one that tells a customer where their specific parcel is right now.
Connecting a tool is a short piece of setup for whoever looks after your systems: they point Tola at the right place, say how it should identify itself, and define exactly what information the assistant is allowed to ask for. No custom software is written.
Two rules are worth holding to when you plan this:
Read before write. Start with tools that only look things up. Give the assistant the ability to change things in your systems — cancel an order, issue a credit — only once the read-only ones have earned your trust.
Keep each tool narrow. "Look up the status of one order by its number" is a good tool. "Query the orders database" is not. A narrow tool can only ever do the narrow thing, which is precisely the point.
The playground
Before an assistant touches a customer, test it in the playground. Ask it your hardest questions, your most ambiguous ones, and the ones where a wrong answer would cost you money.
What to look for:
- Does it say "I don't know"? An assistant that always produces a confident answer is a liability. Confirm it escalates when it should.
- Are the sources right? A right answer from the wrong document is luck.
- What happens at the edges? Try a question just outside its knowledge, and a question that mixes two topics.
- Does it respect the guardrails? Ask for a discount. Ask something legal. It should decline.
Connecting it to a channel
Assistants are enabled per channel. That is the safety valve: put it on live chat, where the stakes are lowest and the volume of simple questions is highest, before you put it on WhatsApp or email.
When the assistant cannot answer, it hands the conversation to a human, and the customer keeps the same thread. The agent picking it up sees the whole exchange, including what the AI tried.
Copilot for agents
Copilot is the assist mode: it drafts replies inside the agent's composer, drawing on the same knowledge, and the agent edits and sends.
This is the highest-value, lowest-risk way to start. Nothing reaches a customer unreviewed, agents get faster immediately, and — most usefully — you find out where the knowledge base is wrong by watching which drafts get heavily rewritten before sending.
Rolling it out safely
A sequence that works:
- Turn on copilot for agents only. Watch what it drafts for two weeks. Fix the knowledge gaps you find.
- Approve FAQ suggestions from your real conversations until the common questions are covered precisely.
- Write guardrails and guidelines. Be specific about what it must never do.
- Test in the playground against your worst cases, not your best ones.
- Enable it on live chat only, and read its conversations daily at first.
- Add scenarios and read-only tools for the situations it handles most.
- Expand to other channels once the numbers hold — and keep watching the handoff rate. A rising handoff rate means your knowledge has drifted, not that the AI got worse.