A recent Wavestone study found that only 5% of companies have a GenAI based solution in production, however, McKinsey reports that nearly 70% of all companies regularly use some form of GenAI. What does this mean? To me, this suggests 65% of all companies are integrating commercially available chatbots and copilots into individual workflows on the desktop. Nearly two years since the launch of ChatGPT, only a tiny fraction of data and analytics teams are building customized GenAI solutions. Does this surprise you? It doesn't surprise me at all. Since the LLM hype started, I've been consistently saying that the overwhelming majority of companies would operationalize AI at scale through chatbots and copilots. Why? Because 'training' an LLM is an extremely costly, time-consuming, and resource-intensive process that's impractical for 95% of companies. In the immediate term, the biggest opportunity for companies to use technology to influence the behavior of off-the-shelf solutions is through more complex RAG patterns, or through fine tuning open-source LLMs. Building the former requires a partnership between data teams and engineering teams, and the latter will require data science expertise. Longer team, opportunities also exist to build more 'small' language models, but I suspect these will be built primarily by vendors and integrated through business applications. To me, the biggest takeaway from the current state of GenAI is that most data teams are far less involved in the deployment of GenAI solutions than many CDOs originally anticipated. Instead of data teams driving the GenAI agenda, it appears it's increasingly being driven by vendors and software engineering teams. Do you agree? #genai #ai #CDO
Eddie Short
3mo
I think the two statements can be true Malcolm Hawker When McKinsey talk about GenAI they bundle in everything and the kitchen sink, and the reality is most organisations are now doing some form of CoPilot - either officially or just using Bing or Native ChatGPT off the side of a desk As you say, integrating your own LLM - that’s a serious piece of work. The biggest use case, for me, is Contact Centres, many of which have been dehumanised - so a LLM replacement offers better service at lower cost, what’s not to like. But it takes time and a lot of testing to get there
Dan Everett
3mo
I imagine RAG is the faster and easier path right now. And in that same Wavestone survey https://wwa.wavestone.com/app/uploads/2023/12/DataAI-ExecutiveLeadershipSurveyFinalAsset.pdf Only 37% of CDAOs say their efforts to improve data quality have been successful So better be careful what you augment with.
I would love to change it to #REgenerativeAI and stop the mindless roll-out
Larry Burns
3mo
It doesn’t surprise me at all. I’ve been saying from the beginning that: a) there are less expensive, more reliable, and less intrusive solutions to most business problems that are readily available, and b) most organizations (and their data teams) haven’t even mastered these solutions, and are thus not really positioned to take proper advantage of these newer technologies.
Dharma Kuthanur
3mo
Malcolm Hawker Agreed...not surprising at all. Most of the GenAI activity within organizations is around developing apps and customizing/automating workflows which are often led by developers, within LOB functions or CIO office/central IT or a combination of both. Data teams will be more involved as RAG adoption picks up. But even then I wonder if the center of gravity for AI will continue to lean more towards CIO and LOB functions, due to the app development focus.
Beth Bauer
3mo
Great post, Malcolm Hawker. Sometimes they are buying it and don't realize / think of it this way. One simple example... 4 yrs ago, we implemented a method to build and evaluate legal contracts (in the confusing world of data purchases!) using GenAI software we purchased and then trained specifically for this purpose. Given how similar legal language is to writing code, at the time, I could not believe this was not already happening. We also built in a mechanism to track the contract to both procurement and the data value proposition. This was apparently revolutionary. Since it was not home grown or internally "marketed and sold" by a mgmt company, do you think anyone even thinks of this as GenAI in production? Oftentimes, the result is in the eye of the beholder.
David Skrobela
3mo
I'd agree. I hear a lot of "we're rolling out co-pilot" or "we're starting to beta test Einstein"..and the next sentence is our software vendor teams are helping with the roll out. Implementing an AI tool used by most of the corporate world does not constitute a full-fledged strategy and doesn't provide any competitive differentiation. Any company that truly treats data as an asset with a direct link to one or many profit centers should get this point..
Much of this is so called "Shadow AI", where organizations aren't even knowledgeable of how their own employees already using AI tech
Sammy B.
3mo
It's interesting to see how the GenAI landscape is evolving. The reliance on vendors and engineering teams for deployment highlights a shift in strategy.
Data and Technology | Financial Crime, Sanctions, AML & KYC SME | Harvard Business Review Advisory Council member
3moWhat I would be more concerned about with any implementation of GenAI as a data team: * How can we safely create value with this technology? - Speed, Accuracy, Methodology, Ethics etc * How do we protect ourselves when using this technology? - Compliance, Legality, Liability, Data Leakage, Data Sovereignty etc There are pros/cons to building built in-house or a vendor-supplier product. It takes lots of resources to support in-house products. Many good 4GL GenAI examples of Reporting tools have arrived on the market accelerating common simpler tasks (inc basic data science work). Given the mess of copyright, and materials being apportioned for GenAI technologies someone must help navigate the data complexities to ensure it protects and serves the business correctly. Depending on the data team's role, it may not be to build the products - either way, any products are evaluated against the capability to deliver on the business objectives. The agenda for the data team is usually set much higher up the food chain in the C-suite/Board/Shareholders. A vendor's job is to market the benefits and narrative in a highly visible way. The real hidden work goes on behind the scenes usually with someone who knows things about data and AI.