Getting personal: Introducing the AND Digital ‘Ask ANDi’ chatbot

At AND Digital, we believe the future of the web isn't found in a menu—it's found in a conversation.

The Strategic Engagement Shift

The digital landscape has pivoted toward conversational and interactive content that understands unique customer context. Websites are no longer "click-and-browse" repositories of information; they are becoming agentic environments where the interface itself adjusts in real-time to the visitor's intent.

Websites moving to operate 24/7 specialised agents will provide a fundamentally stronger customer experience by meeting customers’ increasingly baseline expectations for instant, context-aware responses rather than searching for answers.

Ask ANDi is our proprietary AI chatbot designed to meet this shift toward conversational and interactive content. It’s a first step into an era where the best interface might be no interface at all—where digital touchpoints mimic human-to-human interaction through natural language and proactive assistance.

Embracing Conversational Design

The Challenge: Beyond Deterministic Logic

Our vision for Ask ANDi is fueled by the value of human-AI partnership.

Traditional web applications follow deterministic ("if this, then that") rules. Modern AI requires a probabilistic approach—defining behavior, personality and nuance required for genuine human engagement, rather than just function.

We designed the chatbot to:

• Represent Our Culture & Brand: Return accurate, holistic information about AND’s culture, capabilities, and unique style of working, balancing confidence and curiosity in its replies.
• Eliminate "Dead Ends": Answer complex queries without frustrating conversation loops, ensuring a seamless hand-off to a relevant AND expert.
• Contextualise Advice: Use deep knowledge of our partnership model to relate previous AND work directly to a specific user's challenge.

Ask ANDi helps clients - current, past and prospective - to understand AND’s current services and capabilities in a self-serve manner. Ask ANDi needed to return correct, accurate and holistic information about our culture, company, and credentials. It also had to function efficiently - to surface accurate results, Ask ANDi needed the ability to understand in-depth the client’s help or information request. It needed to relate previous AND work to the specific queries of the chatbot user.

It was important that ANDi connects visitors with a relevant AND expert without frustrating conversation loops or dead ends - that can lead to friction and flow drop-off. We implemented a simple bespoke Contact Us form in the first instance, while we explore the appropriate context for the chatbot to understand when to hand over to the right AND expert.

A defining challenge was chatbot governance. We wanted to keep Ask ANDi inside a strict context - the generic chatbot system of assumption or outright hallucinations were not acceptable, as they would be irrelevant to the customers needs. We could not accept a return of creative responses with fabricated links.

How we did it: Defining Probabilistic Requirements

Designing a probabilistic chatbot like Ask ANDi requires a shift from "coding" to "curating" behavior.

1. Defining Personality as a Guardrail

In a probabilistic system, "tone" is a functional requirement. Humans must define a precise system prompt that acts as the bot's behavioral compass.

• The Human Role: Designers curate the "personality"—for Ask ANDi, this means being expert, confident, and clear while avoiding heavy jargon.
• The AI Role: The model uses this persona to navigate unstructured data, ensuring its "style" remains consistent even when responding to queries it hasn't encountered before.
  

2. Probabilistic Intent vs. Deterministic Results

The most complex interplay occurs when translating messy human language into accurate corporate data.

• Intent Extraction: The AI handles the probabilistic task of understanding a client’s unique context—mapping a vague question to a likely need.
• Fact Anchoring: Humans provide the deterministic foundations to prevent hallucinations - in Ask ANDis case that’s our indexed website. This ensures the "creativity" of the AI is strictly bound to "correct, accurate, and holistic" information.

3. Stress-Testing for Accountability

Because AI augments capability without replacing accountability, the design process requires several "Human-in-the-Loop" cycles.

• Role-Playing: Senior experts regularly "stress-test" responses against real-world scenarios.
• Refining the Loop: Humans identify "dead-end loops" where the AI fails to satisfy a query, then train the model to ask relevant follow-up questions to gauge intent more accurately or intelligently handover to a relevant human expert.

As we explored this, it brought home the quote "Any fool can make something complicated; it takes a genius to make it simple."

Technical Architecture

With over 80% of enterprises now deploying generative AI applications, the focus has shifted from "can we build this?" to "how quickly can we scale it?". Leading cloud platforms now provide "out-of-the-box" tools that allow even non-technical teams to deploy context-aware agents in minutes, making intelligent digital assistance a baseline expectation for any modern brand.

Our architecture follows a modular, serverless pattern designed for independent scaling, security, and real-time knowledge integration.

1. The Core: Intelligence & Data

At the heart of the system is the Vertex AI App, which acts as the primary conversation engine. To ensure accuracy, it draws from two live data streams:

• Unstructured Knowledge: A Document Data Store that automatically indexes PDFs, such as case studies and manuals, as they are uploaded.
• Live Web Data: A continuous web crawler that synchronizes the bot’s knowledge with our latest public-facing digital content.

2. Middleware: The Control Layer

Rather than exposing the Al engine directly, we use a Serverless TypeScript API hosted on GCP Cloud Run. This layer provides:

• Governance: It houses the System Prompt and app configurations that define the bot’s personality and safety guardrails.
• Security: Acts as a secure proxy to manage access keys and protect the AI endpoints.
• Efficiency: Scales instantly to meet demand without the need for dedicated server management.
  

3. Front End: Seamless Integration

The user interface is a React application hosted natively within our HubSpot marketing ecosystem. This allows the bot to feel like a natural extension of our brand while utilizing the power of a modern JavaScript framework for complex conversational features.

Rapid Continuous Deployment

To maintain agility, we utilise GitHub Actions for automated builds and deployments. Any change pushed to our repository—whether a personality tweak in the prompt or an update to the API—is automatically validated and deployed to production, ensuring Ask ANDi is always operating at its peak.

The Outcome: An Intelligent, Connective MVP

The development of Ask ANDi has successfully transformed our digital presence from a static repository into an agentic, conversational environment. This initiative provides a robust MVP foundation for the next phase of conversational and interactive digital engagement.

• Advanced Interaction: This framework paves the way for introducing voice and avatar technology to further humanize the digital experience.
• Intelligent Automation: We are progressing from a bespoke "Contact Us" handover to intelligent, in-conversation automation available on every webpage.
• Continuous Enhancement: Refining the bot is an ongoing activity; we are committed to adding more content and context while monitoring and tuning the human-AI partnership.

Ultimately, Ask ANDi serves as a powerful first step into a new era of digital touchpoints—one where proactive assistance and natural language replace traditional navigation to meet the baseline expectations of the modern brand. If you are interested in exploring this too, speak with us about our Chatbot accelerator.