It should be pretty self-evident that AI is only as good as the data it is fed. Whether it is a custom Machine Learning model driving, for example, product recommendations to customers, or an AI-as-a-Service solution from, e.g., OpenAI or Google, the data is a fundamental aspect of any data-based solution.
It starts with high quality data you can trust
With that in mind, it’s important - not just from an outcomes perspective, but also from a compliance perspective - to ensure data is of high quality, represents reality as best as possible, and is carefully considered and curated before being used in a solution.
As a simple example, when building a document-based chatbot (for example, either a customer-facing bot that relies on FAQs as an input, or a common one we see a lot of is a HR policy bot for employees), making sure the inputs to the bot are correct, unambiguous, etc, is critical.
We’ve seen organisations point RAG (retrieval augmented generation) bots at entire SharePoint sites where multiple, conflicting policies can be located and hosted, or annually updated policies are shared without removing the old policy.
Scenarios like this can make it very challenging for AI to find the correct answer. Providing incorrect solutions does nothing but break trust and further erode the probability of success.
Invest in accessible data that's readily connectable
Similarly, with machine learning solutions, we need to invest in making data accessible, high quality, and readily connectable to other datasets within the organisation. If we don't, it becomes much harder for a data scientist to work out if there are opportunities to deliver real business value and solve actual business problems out of several data sets.
In addition to this, a fairly low cost few weeks of experimentation to find out if a solution is achievable or not could rapidly become months of work on just data munging and manipulating. This is before we can spend any time working on the real business challenge and trying to add value.
Look at the wider ecosystem, playpens, sandboxes and more data!
Data accessibility and quality are, of course, just one part of the whole ecosystem. With ML solutions, it’s crucial to have playpens or sandboxes that allow organisations to stand up meaningful experiments and validate their ideas. Some places are still constrained by legacy tools and by running these experiments on laptops.
Many ML solutions inherently benefit from more data, so having to artificially truncate the quantity and volume of data we operate on degrades the quality of our models. This, in turn, means we are actively reducing the potential business impact that they may be able to deliver.
For example, product recommendation algorithms can be very RAM-intensive (and expensive). However, being able to do it at the customer/SKU level of granularity is where we drive true personalisation away from more segmentation-based recommendations in previous generations.
Fix your legacy estates and digitise your business processes
Our ability to scale AI, as well, to really deliver transformational benefits relies heavily on the level of digitisation across our technical estate.
Clunky, legacy estates with poor data flows between systems, multiple versions of the same truth, unclear definitions for key business metrics, and poor integration or a lack of interoperability between systems make it hard for any form of agentic AI to emerge. If business processes cannot be digital natively, AI solutions are unable to really deliver significant value.
Fix first, then reinvent to embed AI across your business
It’s also important to note that digitisation of what already exists is just a first step on becoming truly AI-native. To fully exploit AI to its maximum potential, we need to reinvent business processes, systems, and ways of working to embed AI at its core, rather than amending it as an afterthought.
For many organisations, however, especially those more encumbered by old school problems, there is a more appropriate path. Our advice is to take the first couple of steps more gently by digitising and modernising with AI, and then subsequently look at transforming to be AI-native.
Find answers to other common data questions
Read my take on:
How can we build more confidence in our data strategy?
Which business problems can effectively be solved by AI, ML or BI?
How can we use data to deliver personalised experiences?
How do you use AI in a regulated industry?