AI agent
What is an AI agent in customer service?
An AI agent understands a question in plain language, looks up information in your systems and carries out the action behind it. That's something other than a chatbot with choice buttons.
In short
An AI agent in customer service is software that understands a customer question in natural language, independently looks up information in connected systems such as your webshop or knowledge base, and answers or carries out an action based on that. The distinction from a chatbot lies in the looking up and the acting: a chatbot picks a pre-written answer, an agent composes one from data fetched at that moment.
What an AI agent needs to work
An AI agent is only as good as the sources beneath it. Without a connected webshop it can't look up an order and every answer stays generic. Without an up-to-date knowledge base it falls back on what's in your website copy, and that's rarely complete.
So the order is: first make sure the information is properly recorded somewhere, then put the agent on it. The other way round produces a system that answers smoothly and regularly says something untrue.
When should a human approve?
Judge per topic on three things: how often it goes wrong, what the damage is when it does, and how easily you recover. That last factor is forgotten most often. A wrongly stated opening time you fix with a second message; a wrongly promised 300 euro refund you don't.
A workable starting rule: let the agent answer on its own where the answer is fixed and costs nothing, and route anything touching money, an exception or a complaint past a person as a draft first.
How you can tell whether it works
The most usable signal is how often your team edits a draft before it goes out. If your team corrects nine out of ten drafts, the agent isn't ready to send on its own, however good the answers sound.
Measure that percentage per topic rather than overall. Usually it turns out the agent is almost never corrected on delivery questions and almost always on returns, and then you know exactly where the line should sit.
Where it goes wrong
Releasing an agent without a fallback
An agent with no way to say it doesn't know will compose something plausible. With every vendor, check what happens on a question whose answer isn't in the sources.
Overlooking a per-resolution price
With a rate per resolved AI question your bill rises as the agent works better. Calculate that on your own volume before signing.
Taking an AI agent live without incidents
The order matters. Granting rights before you've measured how often a draft gets corrected is the mistake that costs most afterwards.
Let it read along first
The agent writes drafts an agent sends. Nobody notices, and you see how often something is changed.
Measure how much gets corrected, by topic
Not an average. There are nearly always topics where it's almost always right and topics where it misses every time.
Grant autonomy by topic, not all at once
Start with the category with the lowest correction rate and the least damage when wrong. That's almost never anything involving money.
Define where it always hands over
Money, complaints, legal questions and requests for exceptions. Write that down before you go live.
Weekly, review the conversations it handed over
That list is also your list of missing knowledge base articles.
What to do this week
Take your twenty most common questions and check per question whether the answer is fully recorded somewhere. The questions where it is are the ones you start an AI agent with. The rest is a knowledge base job first and an AI job only after that.
What this looks like in a system
The AI agent in Cuego draws on your knowledge base, your policy, earlier conversations and your connected systems. If the answer isn't there, it hands the conversation over with the question and what it already checked, instead of composing something plausible. Per topic you set whether it sends itself or prepares a draft, and your team's corrections are remembered for next time.
AI agentsRelated terms
Chatbot
A chatbot follows a script. It works as long as the customer stays inside that script, and customers rarely stay inside a script.
Read onWorkflow
A workflow is a sequence of steps that runs by itself once something happens. The skill is choosing which steps deserve it.
Read onDeflection
Deflection is the share of questions that never reached an agent. It's the easiest number to overstate and the most useful to measure honestly.
Read on