Lessons from 2 years of building Generative AI experiences.

Over the past two years, Time Under Tension has been at the forefront of building generative AI experiences for Australian businesses. When we founded the agency, we saw a massive opportunity at the intersection of generative AI, design, and user experience.

But what exactly does a "generative AI experience agency" do?

Rather than explain the theoretical space we occupy, let me share the real-world challenges we've been solving - and the exciting opportunities we've unlocked - for our clients.

  1. Conversational Search

  2. Writing with AI

  3. From RAG to Riches

  4. AI workflows

  5. Agentic solutions

  6. One-stop-AI-shop approach

  7. Right model for right job

  8. R&D as a competitive edge

💬 Conversational search is effective (and fun!)

Transforming product search into a new age conversational experience

Traditionally, when searching online for products, consumers are faced with menus, clicks, filters, keywords, and sliders to locate what they want. Effective, sure, but not exactly fast, or fun.

We saw an opportunity to shake the product search experience up, using conversational AI.

Through R&D, we proved that conversational search, especially when powered by a helpful AI, can cut product search times by (at least!) half. Searching in natural language isn't just quicker; it's far more satisfying for the user too.

Building on these insights, we launched a conversational AI product search capability, transforming the search experience completely - faster, easier, and more personal.

Conversational search, coming to a retailer near you soon!

Back to 🔝


🤝 Humans want AI to write with them, not for them

Creating AI writing assistants that people want to use

For many of our clients - especially larger brands and agencies - writing documents, like briefs - is a constant struggle. Information gaps, lack of detail, and unclear requirements often create confusion for teams, whether they're writing or receiving briefs.

Initially, we considered a purely AI-driven solution. But it quickly became clear that people don't want AI to write for them; they wanted AI to write with them.

Our solution is a hybrid UX approach: humans drafting and editing on one side, supported by a context-aware AI assistant on the other. This hybrid experience feels intuitive and genuinely collaborative. Instead of replacing the human touch, AI enhances it, offering relevant suggestions, filling gaps, and sparking ideas precisely when needed.

Brief writing is a common challenge for businesses. Getting the UX right can make all the difference.

Back to 🔝


🧠 Reasoning is a game changer for corporate knowledge

Building next generation corporate knowledge experiences with deep research

Imagine you're asked to find ten specific needles hidden within a barn-sized haystack filled with thousands of needles. That’s exactly the challenge businesses face when trying to use AI to extract valuable information from massive corporate knowledge repositories.

Recently, a client came to us with over 15 years of accumulated documents stored in a Box folder, needing quick and accurate answers from this vast information pool.

Retrieval Augmented Generation (RAG) alone simply doesn't cut it—particularly when tackling complex, multi-faceted topics requiring human-level precision. This is where Reasoning becomes a game changer.

Reasoning allows AI to break down complex queries into logical, manageable steps—creating a structured research plan before diving into document retrieval. Think of it as giving AI a precise map to locate exactly what it’s looking for inside that massive haystack.

Basic Retrieval Augmented Generation (RAG) typically delivers average results - acceptable for simple queries but underwhelming when handling complex questions. However, when enhanced with Reasoning, the AI can produce structured research plans, carefully navigate complex queries, and assemble results that are genuinely awe-inspiring.

Our AI knowledge solutions are able to assemble detailed responses and reports, by building structured plans before delving in and searching knowledge bases

Back to 🔝


🔗 Think beyond tasks, think workflows

Disrupting processes using AI powered workflows (including our own!)

True story: At Time Under Tension we run AI Design Workshops to help clients tease out and prioritise generative AI use cases. Even with AI assistance (such as transcription and research), it still took about a week to process the workshop notes, research, and draft the comprehensive report.

That was until Jackson - one of our developers - joined a workshop with us. Surprised by our week-long report turnaround, he figured there had to be a better way. Fast forward to today: Jackson built WorkshopGPT, an AI-powered tool that reduces our report turnaround from days to just hours.

WorkshopGPT ingests workshop assets (like photos of post-it notes), and then get’s to work, it;

  • transcribes all the workshop materials, using vision

  • fills information gaps, flagging to us where it’s augmenting

  • researches similar case studies, from a knowledge base

  • deep researches for potential solutions (online)

  • drafts detailed recommendations

  • and… packages it all into a detailed PowerPoint template, with space for us to add our human magic

Thats weeks worth of work, completed in a few hours hours.

We can now deliver workshop recommendations in hours, not days. Our focus can shift from the mundane layout and research, to the valuable thinking parts.

But Jackson still isn’t satisfied - his next goal is real-time delivery, right there in the workshop!

WorkshopGPT is just one of several AI tools we now use in our business, rethinking traditional workflows.

Back to 🔝


🕵️‍♀️ After workflows, come agents!

AI agents can do the work we don’t want to do

As a team immersed in AI every day, we have a lot of debate about what exactly defines an "AI agent". Whilst the debate is ongoing, what is clear to us is that a lot of our work is agentic in nature.

In this example, we built an AI which integrates with an eCommerce PIM, of thousands of products, and autonomously run a set of tasks against each one. Checking legal and compliance of product descriptions, reviewing images for hazardous or misleading imagery, writing drafts as needed and SEO and SEM keyword / search tasks.

Autonomous. Intelligent. Effective. Doing stuff we don't want to do! These are some of the terms we've used when trying to define agents, all of which has been achieved for this client, in a task that would otherwise be impossible (or take months) for humans to do.

AI agents are at their best when doing the stuff humans don’t want to do.

Back to 🔝


🔐 One-stop AI shops make AI accessible to all (and secure too!)

A secure approach to enabling an entire business with AI capability, overnight

AI adoption can be messy.

Some employees are expert prompters, chaining workflows like pros. Others? Not so much.

Then there’s security. Most off-the-shelf AI tools default to training on your data, creating serious risks for sensitive corporate information.

The solution? A one-stop AI interface built securely on enterprise infrastructure, can offer:

Single sign-on & controlled permissions – keep access streamlined and secure
“Safe AI” protocols – no risk of your data training external LLMs
Democratised best-in-class prompts – ensuring everyone benefits from top-tier prompting workflows
Fine-tuned on your organisation’s knowledge – making AI more relevant and powerful

With this approach, entire teams - regardless of prompting skills - can harness AI securely and effectively.

One-stop AI shops are become a more common way to roll-out AI across organisations.

Back to 🔝


💸 Control AI usage costs by selecting the right model for the right job

Saving money by hammering the right nail with the right hammer

Have you considered how much your choice of AI model impacts cost? For instance, summarising an annual report using Gemini 1.5 Flash-8B costs just $0.01. Doing the same with Claude 3 Opus? A whopping $5.62 - that’s a staggering 500x difference!

Choosing the right model isn't just about capability; it’s also about cost efficiency.

Think of it this way: If you’re just heading down the street, walking makes sense. Taking a helicopter might get you there faster, but is it efficient? The same logic applies when selecting an LLM. Just because the most powerful model is available doesn’t mean it's the right choice for every task.

That’s why in our AI SaaS tool, Peak, we match tasks with the most suitable models. With over 40 curated Skills, Peak ensures users always get to use the right model for the right job - saving time, money, and resources.

Each of Peak’s 40 Skills are designed with model switching built in

Back to 🔝


🧪 Always on R&D pays (essential) dividends

“Even if AI development completely stopped, we would still have 5-10 years of rapid change absorbing the capabilities of current models and integrating them into organisations and social systems.” - Ethan Mollick

With new generative AI models emerging weekly (including LLMs, image, video, music, and more), staying across their capabilities is crucial for maintaining a competitive advantage.

Recently, we explored Google's new Gemini Live Stream capability. By combining Gemini with existing product documentation—in this case, a coffee machine manual—we created a virtual troubleshooting AI that visually diagnoses problems in real-time.

This visual diagnosis opens enormous possibilities for customer support, training, and troubleshooting across almost any product category.

But technical know-how alone isn’t enough. Continuous exploration and ongoing R&D are critical—not just to keep pace, but to uncover novel applications before your competitors even realise they exist.

The most innovative AI applications often come from experimental combinations of multiple emerging capabilities.

Back to 🔝


The lessons all point to one thing: experience matters

Looking back over some of the lessons from our first two years as a generative AI experience agency, a clear theme emerges. The difference between an AI application that gets used, and one that transforms a business doesn't come down to which model you use or how many parameters it has – it comes down to experience design.

The most successful AI experiences we've delivered share common traits: they're contextual, intuitive, and focused on solving real business problems.

As we head further into 2025, we're seeing the AI landscape mature well beyond the initial hype. Our clients are no longer asking "should we use AI?" but rather "how can we implement AI in ways that actually stick?" The answer often lies in the experience around the technology – and that's exactly where we'll continue to focus.

If you'd like to chat about how we're applying these lessons to create generative AI experiences that deliver real business impact, reach out to us at hello@timeundertension.ai

Previous
Previous

Are you vibe coding yet?

Next
Next

Browser Agents: The Marketing Superpower Your Competitors Don't Know About Yet