ARTIFICIAL INTELLIGENCE · AI AGENTS
Conversational AI agents. They answer your customers, under your control.
I design conversational agents for your website and business messaging apps. They answer common questions using your information, prepare requests and hand over to a person when a request falls outside their scope, or whenever the customer asks.
01 / WHAT AN AGENT CAN DO The questions that keep coming back, handled at any hour.
A conversational agent sits on your website or on a business messaging app. It relies on what you give it, and it works within the limits we set together.
Answer common questions
Opening hours, service area, services, procedures, documents to bring: the agent answers from a knowledge base that you read and approve.
Prepare a request
For a quote or an appointment, it asks the useful questions, sums up the need and passes it on to you. You are the one who confirms the price, the time slot or the commitment.
Point people the right way
It shows the right page, the right form or the right person, instead of leaving visitors to search the site on their own.
Welcome people after hours
In the evening or at the weekend, it answers what it knows, takes down the request and tells the customer when you will get back to them.
Talk in several languages
Current models understand and write many languages, though quality varies from one language to another. I test your customers’ languages before going live.
Hand you a summary
Every conversation passed on to you comes with a summary: you see the key points at a glance, and the full exchange stays in the log if you need it.
02 / HONEST LIMITS What to know before you start.
An AI agent is useful for simple, repetitive requests. It is still a tool: it is better to know its weaknesses before trusting it with your customers.
I explain them from the start, and I build the agent to keep them in check.
It can be wrong
A language model can state a wrong answer with confidence. I reduce this risk by limiting it to your information and asking it to say when it does not know, but the risk does not go away.
It depends on the information it is given
If a price or opening hours change, the agent has to be updated. Outdated information gives outdated answers.
It does not replace your customer relationship
A complaint, an emergency or a delicate situation needs a person. The agent recognises them as best it can and passes them on.
It has a running cost
Hosted AI models are usually billed according to the volume of conversations. I estimate this cost with you before launch.
I don’t promise a set rate of correct answers or a number of extra enquiries. The result depends on your information, your customers and the tests we run together.
03 / SAFEGUARDS Rules set before the first conversation.
As a cybersecurity engineer by training, I treat an AI agent like any other access to your customers and your data: with clear, tested limits.
Human approval
You approve the information and the standard messages before going live. A quote, a price or a commitment always goes through you.
No automatic sensitive decisions
The agent does not grant refunds, turn customers away or approve payments, and it gives no medical, legal or financial advice.
As little data as possible
It only asks for what is needed to handle the request. Never a password, a card number or an identity document. The retention period is set with you.
A conversation log
Every conversation is recorded: question, answer, handover to a person. You can check it, correct the agent and find an exchange if a customer mentions it.
Transparency for your customers
From the first message, visitors know they are talking to an automated agent, and they can ask to speak to a person at any time.
Limited access
The agent only has access to the tools it strictly needs. I also test it against trick messages written to make it ignore its rules.
04 / THE METHOD From scoping to going live.
Scope
The questions that keep coming back, what the agent must never do, and where it will be installed.
Prepare
A knowledge base approved by you, the tone of the answers and the rules for handing over to a person.
Test
Real questions and trick questions, with you, before anything goes live.
Launch and follow up
Going live after your approval, then reading the log and making regular corrections.
05 / WHAT IS CHECKED The points reviewed before going live.
- Information and standard messages approved by you
- Rules for handing over to a person
- A clear notice that the agent is automated
- Real questions and trick messages tested
- Your customers’ languages tested
- Minimal data, retention period set
- Access limited to the tools needed
- Conversation log in place
GO FURTHER Expertise in detail.
AI agents: your questions.
Can an AI agent answer on my behalf on WhatsApp?
It is possible through WhatsApp’s business platform, which comes with its own conditions and usage rules. I check with you what can be done for your account before we start. On your website, installation is simpler.
Will my customers know they are talking to an AI?
Yes. The agent says so clearly in its first message, and the person can ask to speak to someone at any time.
What happens if the agent does not know the answer?
It says so instead of making something up. It offers to pass the request on and, if the person agrees, takes down the contact details you need to get back to them.
What data is kept?
Only what is needed to handle the request. The conversation log is kept for a period set with you, then deleted. I help you explain how this works to your visitors in your privacy policy.
Can I change what the agent says?
Yes. The information and the rules belong to you. I update them when your offer changes, and every change is tested before it is published.
Can the agent send quotes or confirm appointments?
It prepares the request and passes it on to you with a summary. Confirming a price, an appointment or a commitment stays in your hands.
YOUR PROJECT · ONE POINT OF CONTACT
The same questions every day? Let’s talk.
Describe the questions your customers ask most often and where you would like the agent installed. I reply with an initial assessment and the next steps.