Preparing small and medium 
businesses for AI success

In today’s data-driven world, Small and Medium Enterprises (SMEs) must harness 
the power of Artificial Intelligence (AI) to remain competitive. Before they can 
fully embrace AI technologies, it’s essential for SMEs to ensure that their data is 
AI-ready. This involves understanding the various types of data they possess and 
preparing it to train their AI or machine learning models.

Understanding data types

Typically three primary types of data is collected:

1. Structured data: This type of data is highly organised 
and easily interpretable by both business users and 
machine learning algorithms. It includes quantitative 
data such as dates, names, addresses, and other 
customer information.
2. Unstructured data: Unstructured data is qualitative and 
includes raw audio, emails, and Internet of Things (IoT) 
sensor data. While it provides deeper insights, it can be 
challenging to process and annotate.
3. Semi-structured data: Falling between structured and 
unstructured data, this type includes formats like HTML 
and JSON, often used for data transmission between 
servers and web pages.


Without a clear understanding of data and strong data governance, you won’t be able 
to effectively feed the AI systems. Data quality and accuracy are paramount, whether 
you are building or procuring AI solutions.

Aurelie Jaquet, CSIRO Australia’s National Science Agency, emphasises the importance of data in AI.

The AI data lifecycle 

To embark on a successful AI project, businesses should 
follow a four-stage AI data lifecycle:

1. Data sourcing: Good quality data may come 
from various sources, including publicly available 
open‑source data, purchased commercial data, or data 
collected in-house.
2. Data preparation: This phase involves structuring and 
labelling raw data, annotating it, and adding metadata 
to inform and train machine learning models.
3. Model training and deployment: AI models and 
machine learning algorithms are trained using rich, 
quality data that are sourced and prepared.
4. Model evaluation by humans: Data patterns can change 
over time, so ongoing human monitoring, retraining, 
and refinement with new data is crucial for improving 
AI models.


NAIC SME BRIEFS



Overcoming barriers to AI 

SMEs often face significant barriers when implementing 
AI, including the high cost and time investment required 
to start from scratch. To address these challenges, SMEs 
have three paths to consider: building AI capabilities 
in-house, purchasing AI solutions, or partnering with 
third‑party experts.

Common obstacles for SMEs include sourcing and preparing 
high-quality data, which can be slow and expensive. 
Access to data annotators can also be a challenge. 
SMEs can tap into the expertise of larger companies or 
explore emerging technologies like self-supervised or 
weakly supervised learning to overcome these barriers.

Key lessons for SMEs

SMEs must focus on the following key aspects:

Benchmarking and human evaluation: 
Good governance involves ensuring 
the accuracy of AI modeling through 
benchmarking and human evaluation. 
Blindly adapting models is unlikely to 
produce accurate results.

Data handling: SMEs should handle 
data containing personally identifiable 
information with care, adhering to data 
retention and privacy regulations. Safe 
and secure data storage is essential, and 
transparency with customers or clients 
about data usage is a must.

Achieving data culture and maturity

Building a strong data culture within your organisation 
begins with assessing existing skill capabilities and 
investing in digital literacy training. A data-mature 
organization leverages data effectively to achieve its goals, 
fostering a data-first culture that empowers data use in 
discussions and decision-making.

Case study: Whisper

Unstructured Data and Weakly Supervised Learning 
Whisper, a remarkable open-source neural net 
AI model, is being trained to transcribe and 
translate speech audio from 97 different languages. 
This automated speech recognition (ASR) software 
was trained on 680,000 hours of unstructured 
audio data collected from the web, utilising weakly 
supervised learning. The approach enabled Whisper 
to achieve high accuracy, even for audio it was not 
initially trained on.

It’s crucial to ensure that AI models work effectively 
for all users, as highlighted by MingKuan Liu from 
Appen, given the challenges faced by speech 
recognition engines in real-world scenarios.

Next steps

When looking to harness the power of AI, understanding 
the importance of data readiness, data governance, and AI 
data lifecycle stages is the first step towards a successful AI 
journey. Collaboration, smart data handling, and ongoing 
evaluation are key to overcoming barriers and building a 
data-centric culture.

The National AI Centre is at the forefront of building 
Australia’s responsible and inclusive AI future: csiro.au/naic