By Martin Banks

December 21, 2023

The world is currently full of `hot’ stories about where AI is going to take us, and do to us, and the state we will be in by the time it has finished with us. It is true that it may gain some semblance of sentience at some time in the future, but it is far from that at the moment. And the argument now is very strong that the `A’ is AI should stand for 'Augmented’ not 'Artificial’. When looked at that way the obvious question for potential users then becomes `what do you want augmented – what help do you need to augment it, where in any process do you need it, and how do you make it happen?’

This subject was addressed at a recent London conference hosted by low-code/business process management vendor, Appian. CEO Matt Calkins side-stepped the hype and used his keynote address to set about outlining where AI is going to help in that all-important area of `making businesses run better’ over the next couple of years.  It was, therefore, low on scary content, but did address the underlying reality that business managers will need to be ready for the significant changes already barrelling down on them if they are not to be caught unawares and unready: 

We need to move from hype to reality. We've talked a lot about how AI can get us to an amazing place. Now we need to figure out how we actually go there: where are we actually trying to go with this technology? AI can take us to a human centric future. AI is going to be a partner instead of a substitute for people.

Calkins also sees the coming changes being more democratically available to businesses of all types and sizes rather than exclusively for large enterprises, a view he suggested is different from current conventional wisdom. And it's different from the view of the investment community as well, which has so far bet heavily on the big technology names because, as he put it, investors believe that they're going to use AI as a tool to gain advantage over all the rest of us: 

I think investors think Big Tech is going to find a way to use AI to get our data, take our money, put us in a position where we have to rent the best technology in our entire enterprise. But we have to move our conversation away from wild predictions and toward demonstrations of effectiveness of AI in the workplace.

Don’t waste your data

What Calkins sees is a need for users to understand that AI in the workplace is not a standalone technology, rather, it is symbiotic with - even dependent on - two other areas, data and process, without which it will not have any tangible effect. He sees it as a triangle of these three primary, interdependent and synergistic areas:

AI without data would be like a card trick with no cards. It is just re-shuffling cards, just re-shuffling data; taking what it's learned, putting it into a new configuration, and putting it back out to you.

Current models have it that AI is not only dependent on data, but also dependent on two data sources: the `past’ in the form of a training data set, and the data that comprises the ‘present’. AI then relates the two of those together.  Therefore, the value of the data presented is key to the value that comes out of any AI process. In Calkins’ view most data today is waste. For example, most data that's known to organisations is not utilized properly because it's not brought to bear at the moment of decision. The moment a customer calls a call center, that's when a business needs to know about that customer in order to react properly. If that does not happen, (within 10 seconds, in his opinion) then the data is effectively wasted.

He sees AI solving much of this, allowing data to be accessed in real time…in theory. In practice, this will depend on the effectiveness of the interface between AI and the rest of the business. This is an area Appian is focussing on, but Calkins has doubts about how many users are aware of it, or the value of it. Added to that is then the business need to understand how best to get to that value:

I'd say your data is maybe double as valuable as it was a couple of years ago. And most people don't realise that they haven't really talked about it. AI isn't 100% reliable. Instead, we are entering a long term period in which AI will write, but humans will edit, AI will propose, humans will decide. It's a team effort that's going to last a long time.

One of the keys here, so far as Calkins is concerned, is Appian’s development of its Data Fabric, which has clocked up 4 billion queries despite still being rolled out to all of the users of the software. The amount of effort put into developing and implementing Data Fabric has led Calkins to suggest that Appian has got close to becoming a Data Fabric company, specialising in the development and application of the technology above and beyond everything else. 

Doing AI without training

Its importance to the application of AI by the company can now be seen as some explanation for the effort that was made. That all hinges round Calkins’ enigmatic statement to the conference delegates:

We're not going to try to take your data; we want to facilitate the enterprise that you've already got.

This has to do with the way that current AI implementations tend to work in practice, where a user’s data can be at risk as a core part of the process, while it is being used. The Appian Data Fabric is at the heart of this because it connects the entire enterprise via what is, in effect, a virtual database layered across all the data sources that exist in that enterprise. This allows users you to run rapidly-built queries and get answers based on the latest live information from any data source connected to the Data Fabric. with full row-level security. 

Calkins was keen to stress that the provision of row level security plays a vital role in the implementation of AI for practical business applications, not least because the fear of many businesses that using AI services puts their data at risk is very real for them. After all, it is the way that most AI services now operate. Most of the services are only available via the mainstream public cloud service providers, so in order to train a Large Language Model service in its particular business and requirements, a user company first has to upload its data to the service.  

We believe that it is a non-starter to load your data, especially enterprise data, onto a public AI platform, just a non-starter.

His point here is that in order to train an AI system sitting on a public cloud service, a business has to feed it with its own real data, and that makes it difficult to definitively claim that it owns that data any more. As an operating model, it does mean they can then work well - questions can be sent, the context derived from the training data and an answer delivered - but a company may have difficulty claiming in  now 'owns’ its data, and it has to keep feeding in new data to the training process. 

One way round that is for a company to install an open source AI inside its firewall, training and using it effectively 'on-premise’. The downside here is that it then requires both a high level of expertise on the part of staff, and a high level of ongoing, dedicated investment to keep the service level up to scratch. So Appian is in the throes of developing a new model for utilising AI that prioritizes data security and, as an added advantage, is expected to be faster, more accurate, and less complex in operation.

The third way to do it is to put a generic, plain vanilla AI in a private cloud space, and then inform it in real time. This is where having a Data Fabric available becomes important. Before a question is actually posed to the AI service a query is run against the entire enterprise data set, a task that requires a Data Fabric. This surfaces all the company data relevant to the query and this is sent, together with the question, to the vanilla AI service in a private cloud. Combining the question with all relevant current data extracted from the live business applications gives the AI the not only the information but also the context underpinning the question.

This will make it far more secure, for the data can be deleted as soon as the result has been accepted. It will also be more cost-effective as users will not have to pay to store large amounts of training data, ongoing data upload fees and running costs of a large cloud services site. Finally, its answers are auditable. The source data can be filtered to remove information that is not relevant for the question being asked, which can also help manage the introduction of accidental biases.

My take

If AI achieves anything over the next few years it will be an almost brutal demonstration of the famous Gartner Hype Cycle in action. At the moment, the expectations for AI are humungous – and continuing to grow – with possible results for humanity ranging from a future life of enduring, blissful indolence to total annihilation. So reaching the 'Peak of Inflated Expectations' still seems a long way off. The subsequent fall into the 'Trough of Disillusionment', when it comes, will no doubt be equally rapid – with the likelihood of many corporate fingers being badly burned.

But as and when AI moves on to the Plateau of Productivity, an operational model such set out by Calkins is likely to be playing a significant part in achieving that goal. It targets a role for AI that just about every business will need at their core, whatever their size, and offers a way of achieving it, if Calkins proves right, that offers speed, flexibility, accuracy and, above all, greater security.