Powering AI with High-Quality Data
By Prof. Toby Walsh, Chief Scientist at UNSW.ai, the AI Institute of UNSW Sydney .
In my previous reflections, I spoke about the extraordinary pace at which artificial intelligence has arrived in our lives. Three years on from the moment the world collectively discovered ChatGPT, we’re still grappling with the implications of that acceleration. But today, I want to look forward to the trends that are emerging now, many of them deeply connected to data, and to the opportunities they present for businesses, governments, and individuals.
One of the most surprising developments of the past few years, at least to me, is that the first “killer app” of AI turned out to be a generalist. A system that could answer questions about astronomy one moment and zoology the next. A digital polymath. It captured the public imagination precisely because it could do a bit of everything.
But here’s the irony: in most real‑world settings, that’s not what we actually need.
The Rise of the Specialist
If you’re running an insurance company, you don’t need an AI that can explain black holes or identify bird species. You need one that understands your policies, your risk models, your regulatory environment, and your customers. You need depth, not breadth.
This is why I believe the next wave of AI will be defined not by large language models, but by specialised language models. SLMs rather than LLMs. These models don’t need to be gargantuan. In fact, many will be relatively small. What matters is not their size, but the specificity and quality of the data they’re trained on.
A perfect example is BloombergGPT, a model trained by Bloomberg on decades of financial data. Only two companies in the world, Bloomberg and Reuters, were sitting on that particular treasure trove, and Bloomberg used it to create a model that speaks the language of finance fluently. Credit derivatives, bond yields, market movements – topics that would leave most of us scratching our heads – are its native tongue.
The result is transformative. Instead of memorising the arcane syntax of the Bloomberg Terminal, users can simply ask questions in plain English. The interface becomes intuitive. The barrier to insight drops dramatically. For Bloomberg, the business case is obvious: a more powerful, more user‑friendly product that subscribers are willing to pay for.
This is the pattern we’ll see everywhere. The organisations that hold valuable, domain‑specific data will be the ones best positioned to build specialised AI models – and the value of those models will come not from their ability to answer every question under the sun, but from their mastery of a narrow, high‑impact domain.
Your Data, Your Model
I’m already seeing companies emerge that offer to build these specialised models as enterprise tools. They’ll come into your business, ingest your policies, your procedures, your customer interactions, your historical records, and build an AI that knows your world inside out.
And here’s the crucial part: the model can run entirely on your own servers.
That means no data leaves the building. No sensitive information is shared with a tech giant. No risk of leakage to competitors. You can look your clients in the eye and say, with confidence, that their information remains private.
This approach solves several problems at once:
- Privacy: Your data stays under your control.
- Security: No external party has access to your proprietary information.
- Accuracy: The model is trained on your real‑world context, not generic internet text.
- Relevance: It answers the questions you need answered, not the ones a generalist model happens to know.
And importantly, it avoids the diminishing returns we’re beginning to see with ever‑larger generalist models. The step changes from GPT‑3 to GPT‑4 to GPT‑5 have been getting smaller, but the cost of training these behemoths is skyrocketing, while the engineering challenges are mounting.
But when you focus a model on a narrow domain, the gains become sharper again. Just as humans become experts by specialising, AI systems will become more capable by narrowing their scope.
The Multimodal Future
Another trend that excites me is the shift from text‑only models to multimodal models. The world is not made of text alone, and neither is our data.
We have:
- audio recordings
- video footage
- tables and spreadsheets
- geospatial coordinates
- sensor data
- genomic sequences
- images and diagrams
- structured databases
And what’s remarkable is that the same underlying architecture – the transformer model that powers today’s language systems – can be adapted to understand all of these formats.
Imagine a model that can read a geological survey, interpret satellite imagery, analyse seismic data, and produce a coherent assessment of mineral prospects. Or a model that can ingest genomic sequences, clinical notes, and medical imaging to support personalised healthcare. Or one that can combine financial statements, market data, and regulatory filings to produce real‑time risk assessments.
These are not science‑fiction scenarios. They are already emerging. The real world is multimodal. Our AI systems are finally catching up.
From Assistants to Agents
The final trend I want to highlight is perhaps the most transformative: the rise of AI agents.
For decades, we’ve had chatbots that could answer questions. They were useful, but limited. They sat passively, waiting for input. They didn’t do anything. But now we’re building systems that can take action.
Think about when you’re onboarding a new employee. There are dozens of tasks:
- creating an email account
- registering them with payroll
- entering bank details
- assigning mandatory training
- notifying managers
- issuing security credentials
Today, a human must orchestrate all of this. But an AI agent could break the process into steps, execute each one, update the relevant systems, and confirm completion. It could change files, enter information, trigger workflows – even spend money within predefined limits.
This is a profound shift. We’re moving from AI as a conversational tool to AI as an operational partner. From answering questions to taking initiative. From passive to active.
And as these agents become more capable, they will reshape workflows across every sector – finance, logistics, healthcare, government, education, and more.
A New Phase of the AI Journey
We are entering a new phase of the AI revolution—one defined not by generality, but by specificity; not by size, but by relevance; not by text alone, but by the full spectrum of human and machine data; not by passive assistance, but by active agency.
The opportunities are immense. But so are the responsibilities. As we build these specialised models, multimodal systems, and autonomous agents, we must ensure they are transparent, fair, secure, and aligned with human values.
AI is no longer a distant future. It is a present reality. And the choices we make now will shape the systems that shape our world.
This article is based on a presentation delivered by Toby Walsh at the Powering AI with High-Quality Data webinar, hosted by Intech Solutions on 27 November 2025.