AI demos usually go pretty well. Most demos do (apart from the odd screen freeze here and there). It’s usually because the questions are tidy and the responses in your chat interface are polite. Why wouldn’t it go well? You designed it for a best-case scenario.
Real life and real customers are a completely different story.
They misspell product names, ask for two things at once, and come back to a conversation a week later because they lost connectivity or got distracted by their kids. To a good digital agent, though, that shouldn’t matter at all, and it would know all of that information from the get-go, because you designed it that way.
At Helm, we don’t think of an AI-driven customer experience as a chatbot. Sure, sometimes they’re involved, but as one part of a much bigger picture. AI-powered CX is a complete customer journey: channels, language, technology, trusted knowledge, business rules, integrations, controls, human support and ongoing management – all working together.
Why good demos fail in production
Take an insurance claim for example. In the demo, the assistant explains how to submit one, and it does it pretty well. In production, it also has to identify the customer where required, collect the right information, accept documents, create or update the claim, confirm what happened, protect sensitive data and hand the case to a person the moment things fall outside its authority.
Start with the outcome, not the chatbot
We hear “let’s implement AI” a lot. That’s not exactly a brief. It needs to start with a defined customer need, pick a journey with real demand, a clear outcome and enough operational readiness to support automation.
“The bank needs a chatbot.” is not a very good problem statement, but this one is more like it:
“Customers who lose their card need a secure way to block it and to understand what happens next.”
The first one picks a channel before anyone understands the journey. The second gives your product, service, risk and technology teams something they can design and measure.
So map the journey before you build anything. Walk it the way your customer does: on their phone, on patchy data, in their own language, halfway through their day. Note every place they have to repeat themselves, wait, switch channels or give up. That’s where automation can help, and where it can make things worse.
Then write down where you are today: contact volumes, completion rates, repeat contacts, handling time, abandonment, complaints and customer feedback. Without a baseline, you can’t tell whether the new service is actually better.
Seven parts of an AI customer experience
1. Channels
Your customers might start on WhatsApp, move to your website, and end up phoning the call centre. Some of them may even end up in a physical store. Choose channels based on how customers behave and what they’re trying to do, not on whichever interface is quickest to launch. And plan for continuity: if someone starts in one channel and moves to another, decide which context follows them and how it stays protected.
2. Language and intent
Natural Language Understanding (NLU) helps a system work out what someone means, instead of relying on exact keywords. A good digital agent picks up the intent, the relevant details and the context across several turns of conversation.
In South Africa, that includes how people really write: switching languages mid-sentence, using slang and shortcuts. Language choice and code-switching need deliberate design, local testing and realistic training data.
The goal is understanding the request well enough to move the customer forward.
3. Trusted knowledge
Your assistant needs an approved source of truth: current policies, product information, fees, service rules and process guidance that someone owns and that’s structured well enough to retrieve. Retrieval-Augmented Generation (RAG) helps a language model ground its answers in selected business content, rather than relying only on its general training.
Your team has to decide which sources are authoritative, who updates them, what the assistant may say, and what happens when information is uncertain or contradicts itself.
4. Business systems and actions
A digital agent becomes far more useful when it can complete an authorised task, usually by integrating with your CRM, billing platform, claims system, case-management tool, knowledge base, order system, or core banking environment.
Every integration needs defined permissions and failure paths. What can the agent read? What can it change? Does a transaction need confirmation? What happens when an API is down, or a record doesn’t match? Answer those questions before a customer finds the exception for you.
5. Identity, privacy and security
Privacy belongs in the design from the start. It cannot – repeat – cannot just be a sign-off at the end. Decide what personal info the service collects, where it goes, who can see it and how long it’s kept, and work through it with your security and compliance teams with POPIA in mind.
6. Human handover
Sensitive, unusual, disputed or high-risk matters should move to a human agent, with the context attached. A handover is only complete when the right team has enough information to carry on without making the customer start again. Remember, they need your help most at this point.
7. Measurement and ongoing management
Products change, policies change, and customers keep finding new ways to ask the same question. You’ll need analytics, conversation review, content ownership, model and prompt testing, incident processes and a regular optimisation cycle. Launch day is where the real work begins.
