Fully Automated Luxury Capitalism or Just Sparkling Process Management?
A roundup of enterprise AI developments and links to some interesting lines of enquiry using agents as knowledge synthesisers in human collaboration networks
It was another week of announcements and excitement about progress in LLMs with the announcement of Claude 3.5 Sonnet, which is already showing some very impressive coding skills. But progress on applying all this raw intelligence to the transformation of work continues to lag behind.
Unbundling of Tasks and Management
The irrepressible Box CEO Aaron Levie talked to Diginomica recently about enterprise AI, and was bullish about the productivity gains we can expect and what these will mean for people. Describing some personal research he was able to complete much faster by working with ChatGPT as a sidekick, he said:
This is, I think, the part that will be hard to quantify about AI, but literally will be the impact of AI, which is I didn't call somebody to do it, I could do it myself. And so I moved something forward faster than I would have, if I had sent the task out, waited three days for the task to get completed, and then we moved things forward. You multiply that by everybody in the economy, and I think what you end up seeing is just literally an acceleration of progress. And that will be the ultimate measure of probably what AI does.
He also mentioned that the only part of his work that depletes his energy is precisely the kind of repetitive tasks that will be automated, which will improve the balance between boring work and fulfilling work for many people:
It's the trifecta. You're going to do more, you're not going to be drained, and you're going to have a lot more fun in the work that you're doing.
As well as jobs as bundles of tasks that will probably become un-bundled, this also promises to re-configure many aspects of how we manage people inside our organisations. With more automation of management grunt work, perhaps managers can spend more time on the personal coaching and line management roles that often suffer from lack of time. Or, as Ethan Mollick controversially argued in the FT recently, perhaps AI can help with that as well:
AI may also help managers directly. Its capabilities in empathy, summarisation, and customisation make it a powerful tool for coaching and mentoring. AI can provide personalised feedback, help employees navigate complex situations, and offer guidance tailored to individual needs and learning styles. It can also watch everything an employee does and offer comments.
By leveraging AI in this way, as a coach and mentor, organisations can scale employee development and support it to a degree that was previously impossible — creating freedom from boring tasks along the way. Done wrong, however, it risks creating a panopticon, where employees feel constantly monitored and judged by an all-seeing AI.
More to do on Organisational Readiness
But let’s not get ahead of ourselves. The work still to do on organisational readiness is not happening fast enough, probably because it is not exciting or headline grabbing, and also doesn’t fit neatly into existing silos and domains of control.
Larry Dignan’s excellent newsletter for Constellation Research is a good read on the current position and views of typical CIOs regarding AI, and this week he helpfully listed some of the emerging good practice and insights he is picking up from his enterprise contacts, including:
Build a library of use cases
GenAI is a tool to solve problems, but isn't more than that
Generative AI projects require a lot of human labor that's often overlooked
AI is just part of the management team now
Compute is moving beyond just the cloud - optionality is the word
LLMOps is emerging and will converge with MLOps in the future
Focus on long-lasting use cases and business value instead of infrastructure
And More to do on Infrastructure
There is also a lot more to do on data and technical infrastructure, as this TechCrunch piece makes clear, which is generally not yet in a place to make the most of custom LLM training and retrieval.
Perhaps this helps explain some of the other news this week, such as OpenAI agreeing to acquire Rockset to help with connected data infrastructure:
Rockset, which was co-founded in 2016 by ex-Facebook engineers Venkat Venkataramani and Tudor Bosman and database architect Dhruba Borthakur, created tools that allowed companies to automatically ingest data from databases and public cloud storage services and then index that data for search and analytics applications.
Rockset’s database platform underpinned things like recommendation engines, logistics-tracking dashboards and — especially relevant to OpenAI — chatbots in domains such as fintech and e-commerce.
But there is also more evidence emerging of firms building their own LLMs using widely available smaller models, since the cost of training specialised internal LLMs is orders of magnitude smaller than the foundational models we have become used to. This will help many organisations get a foothold in relevant, custom use cases for enterprise AI.
Agents and Synthesisers in Human Networks
As I mentioned last week, if we can overcome these impediments to early progress, much of the really exciting near-term potential lies in multi-agent deployments where small, specialised autonomous agents can hand-off tasks to each other to help people complete complex processes and workflows.
But the UI for this is unlikely to be just chatbots.
Zeya Yang at a16z Enterprise wrote an interesting piece about how these “SythAI” apps might own the workflow in this kind of scenario, and even proactively adapt the user experience to the specific context in which the workflow operates, extrapolating from examples like Figma, Macro and Claygent where AI has been subsumed within specific application experiences.
Taking this even a step further, having reliable, proactive AI capabilities should change how we interact with products at a more fundamental level. One of our favorite thought experiments on the team is to imagine how an AI-powered CRM would manifest. In the most extreme case, an AI CRM would not at all resemble a CRM as we know it today.
There are echoes here of Apple’s presumed strategy in the consumer market for Apple Intelligence - LLM agnostic and focused on the user experience in specific narrow domains, rather than creating one over-arching general AI tool. If they can show what kinds of experiences and utility are possible, then perhaps we will see more focus on narrow agents being combined on shared platforms in the enterprise as well.
But agents are not just about helping people navigate and fulfil the byzantine process requirements of complex enterprises. An area that interests me a lot is pro-active, semi-autonomous knowledge assistants and concierge bots. And similarly, I don’t think text or voice chat will be the predominant UI model for these in all cases.
A typical example of how people are thinking about knowledge agents is this recent piece by Boris Kontsevoi in Forbes, but honestly this is just table stakes when it comes to what is possible in this domain.
Beyond search and knowledge assistants, we are already seeing the potential for autonomous agents to play a role in knowledge synthesis and semantic discovery within human (or organisational) networks. This paper by Atsushi Ueshima, Matthew I. Jones & Nicholas A. Christakis in Nature starts with a simple example, but it points to a fascinating area of discovery and some potentially game-changing use cases.