How an agent works

Every conversation an agent holds follows the same six steps: hear, understand, look up, decide within limits, act, and record. The steps are the same on a phone call, a WhatsApp thread and an email. Knowing them tells you where the facts in an answer come from, where your rules are applied, and what is written down afterwards, which is what lets your team trust an agent with its customers.

The six steps at a glance

1. Hear: the customer's words arrive

On a call, the agent listens as the caller speaks. Speech recognition (the technology that turns spoken words into text) runs continuously, so the transcript builds as the call happens. The agent recognises when the caller has finished speaking, stops if the caller interrupts, and filters out background noise from cars, streets and building sites.

On WhatsApp, SMS, email and web chat, the message arrives as text, with any attachments. Whatever the channel, the agent works out which language the customer is using from their first few words and replies in it. That includes Gulf, Egyptian, Levantine and Modern Standard Arabic, English, and many other languages, including Urdu, Russian, Chinese and Spanish, and switching between Arabic and English within a sentence.

For the technical detail on which speech providers are used and where they run, see Models and providers.

2. Understand: what the customer actually wants

Customers rarely open with a neatly stated request. The agent identifies what the customer is actually asking for, and picks out the details that matter as they are mentioned: names, dates, amounts, addresses and reference numbers. When a request is unclear, it asks a question back rather than guessing.

It understands the request in terms of your industry, using the industry model Telonic builds for your sector: the things your business runs on and the rules attached to each. A property agent knows that a unit has a payment plan and that the next instalment on a construction-linked plan (one where payments fall due as building work reaches agreed stages) depends on which milestone the project has reached. An insurance agent knows that a policy has an excess (the part of a claim the policyholder pays) and exclusions. The agent also knows where this customer is in a longer process, such as enquiry, viewing, sale and handover, and what should happen next.

It draws on what this customer has said before, on any channel, so a returning customer is not asked to start again. See Memory and context.

3. Look up: the facts come from your systems

Anything that could create an obligation for you is taken from your systems rather than generated by the model. Prices, payment plans, policy terms, entitlements, dates and commitments are retrieved from your own systems and approved documents during the conversation. A customer asking what they owe next hears their own schedule, from your system, while they are still on the line.

When the answer is not in those sources, the agent says so and offers a person. It does not fill the gap.

The technical distinction matters to anyone assessing risk. The language model (the AI component that interprets the conversation and composes the reply) decides how retrieved information is expressed: the wording, the order and whether to ask a clarifying question first. The facts in the answer come from the sources your team has connected and approved. Figures such as amounts and dates are inserted directly from the system record, not retyped by the model. Every lookup is logged with its source.

4. Decide, within your limits

The agent chooses what to do next from a set of actions and values your organisation permits. Your rules on what can be offered, what needs approval and what is never done are built into what the agent is able to do, rather than written as instructions it is asked to follow. If a payment plan is not in your permitted structures, the agent has no action that can create one.

Topics you place out of scope are declined and routed to a person, including when a question is rephrased. When a conversation reaches a trigger you have set, such as a customer in distress, a legal question or a request outside the agent's limits, the agent hands it to your team. Decisions that belong to your people, such as a claims decision, are not available to the agent at all.

5. Act: the work gets done in your systems

When the agent can complete the request, it does so during the conversation. It books the viewing into the right calendar, changes the reservation, logs the defect against the right unit, takes the first notice of loss (the first report of a claim) in the structure your claims system needs, or sends the document the customer asked for. Actions you mark as sensitive wait for a person on your team to approve them.

When the request needs a person, the agent hands the conversation over. Your colleague receives the full conversation and a written brief of what has been established and what is outstanding, so the customer does not repeat themselves. See Handover to your team.

Each connection to your systems has its own permissions, set with your IT team. See Permissions.

6. Record: every conversation leaves a complete record

When the conversation ends, the agent writes a summary, records the outcome (resolved, still open, or handed to your team), and lists what needs to happen next. The transcript, what was looked up, what was decided and why, and every action taken are logged together.

That record is added to the customer's history, so the next conversation on any channel starts with it. It is also written back to your CRM (the system that holds your customer records) and operational systems, so the record sits where your team already works. See The customer record.

How the steps differ by channel

The six steps are the same on every channel. What changes is how the agent listens and how it replies.

In practice

A motor insurer uses a Telonic agent for first notice of loss outside office hours.

1. Hear. At 11:14pm, Khalid calls the claims line after a minor road accident. The agent greets him in Arabic and English and asks for his consent to record the call. He replies in Arabic, and the agent continues in Arabic.
2. Understand. It checks that he is safe, then confirms who he is and identifies that he is reporting a new motor claim.
3. Look up. It finds his policy in the insurer's policy administration system, including his cover and excess.
4. Decide within limits. It asks the questions the insurer's motor claim process requires, in the required order. It explains what his policy says about this type of claim, from the approved wording. It does not comment on liability, because that decision is not available to it.
5. Act. It logs the claim in the claims system, issues a reference, and sends Khalid a WhatsApp message listing the documents to send and what happens next.
6. Record. At 8:30am, the claims handler opens a complete, structured file with a summary and the full conversation. Nothing needs chasing.

What your team controls

• Scope: the topics and requests the agent handles, and those it never discusses.
• Limits: the actions it can take, the values it can offer, and the actions that wait for approval.
• Handover: the triggers that send a conversation to a person, and who receives it.
• Sources: which systems and documents the agent can read, and which it can write to.
• Tone: how formal, warm or brief the agent is.
• Records: how long conversation records are kept, and who in your organisation can see them.

These are set with you during implementation and can be changed afterwards. See The controls your team holds.

Related

• Policies as limitsAgents
• Models and providersAgents
• Handover to your teamAgents
• How integrations workIntegrations
• The customer recordCustomer record

Product names and logos are trademarks of their owners. Their mention shows systems Telonic connects to and does not imply partnership or endorsement.