In deploying AI and automation to upgrade our old organisational systems of work coordination, we should embrace the fact that we are still in the Centaur chess stage, that it might last a while, and that this is probably a good thing. Enterprise AI is built on the same technology as open-world generative AI, but it is not the same thing.

After chess grandmaster Gary Kasparov was beaten by a computer in 1997, there was a period when the best chess players were neither human nor machine, but humans + machines. As Nicky Case put it in an essay about Centaurs in the Journal of Design & Science:

When you create a Human+AI team, the hard part isn’t the “AI”. It isn’t even the “Human”. It’s the “+”.

So, how do you find the best “+” for humans and AI? How do you combine humans’ and AI’s individual strengths, to overcome their individual weaknesses? Well, to do that, we first need to know exactly what humans’ and AI’s strengths and weaknesses are.

Even if AI development paused today, the technology would be sufficiently advanced to transform the structures and management systems of most organisations, saving trillions of dollars whilst improving productivity – but it would still take years to overcome the change resistance and vested interests of the bureaucrats.

The tech is not the issue. It is how we adopt it in pursuit of new and better ways of working.

In combinatorial innovation, added value comes from how we compose and combine components to deliver radically improved products or experiences. Enterprise AI is a combinatorial challenge, rather than an invention challenge, so we should focus on exploiting what we already have rather than just standing with our faces pressed to OpenAI’s workshop window, waiting for an omniscient AGI to emerge like some long-predicted deity.

“Battle of Centaurs and Wild Beasts” estimated to have been made 120-130A.D. for the dining-room of Hadrian’s Villa (via Wikipedia)


Is that AGI or just several small AIs on top of each other in a trench coat?

Last week, Meta’s chief AI scientist and Turing award winner, Yann LeCun, told an event in London that Large Language Models (LLMs) will not become Artificial General Intelligence (AGI):

He pointed to a quartet of cognitive challenges: reasoning, planning, persistent memory, and understanding the physical world.

“Those are four essential characteristics of human intelligence — also animal intelligence, for that matter — that current AI systems can’t do,” he said.

Benedict Evans added some historical context in the FT to this re-evaluation of whether AGI really is just over the next hill (like Tesla’s promise of Full Self Driving) or whether much of it is really just sparkling automation and not AI champagne:

The problem is that we don’t have a coherent theoretical model of what general intelligence really is, nor why people are better at it than dogs. Equally, we don’t know why LLMs seem to work so well, and we don’t know how much they can improve. We have many theories for parts of these, but we don’t know the whole system. We can’t plot people and ChatGPT on a chart and say when one will reach the other.

Instead of chasing the dream/nightmare of super-human intelligence, I think we will see more discussion of how we develop a wide variety of narrowly scoped machine intelligences (or AI ‘skills’) and then use both software agents and people to decide what combination to use in a particular situation. We already have the concept of ‘Mixture of Experts’ inside LLMs, where the AI decides which experts to use on a specific type of challenge or question, but breaking this agent logic out to make the process more transparent feels closer to an evolutionary architecture, where each component seeks to improve itself within a connected ecosystem.

Smart people and teams using narrow AIs will produce better outcomes than either people or AIs alone; and if we prioritise human outcomes, creativity and imagination over ever more powerful raw computing for its own sake, that sounds like a pretty good approach. Even if Centaur chess is not the pinnacle of theoretical chess strategy, what’s the point of the game without the players?

Towards centaur service teams

We should be trying to build more Centaur services moulded around high-performance teams who own a clear outcome or product (whether internal or externally-facing), with a combination of narrow AIs, automations and agent logic providing a platform on which the team can work its magic and scale their impact.

Inside large enterprises, senior leaders have their work cut out getting the organisation AI-ready in terms of connected systems, data, knowledge and security, but it is not too early to start thinking about how to design Centaur services that can apply AI to practical, productive use cases within their existing ways of working.

One area where we will see this sooner rather than later is probably in the way we imagine and code internal apps. Andrew Ng recently shared an interesting story about his agent setup for running code:

Many people had a “ChatGPT moment” shortly after ChatGPT was released, when they played with it and were surprised that it significantly exceeded their expectation of what AI can do. If you have not yet had a similar “AI Agentic moment,” I hope you will soon. I had one several months ago, when I presented a live demo of a research agent I had implemented that had access to various online search tools.

I had tested this agent multiple times privately, during which it consistently used a web search tool to gather information and wrote up a summary. During the live demo, though, the web search API unexpectedly returned with a rate limiting error. I thought my demo was about to fail publicly, and I dreaded what was to come next. To my surprise, the agent pivoted deftly to a Wikipedia search tool — which I had forgotten I’d given it — and completed the task using Wikipedia instead of web search.

This was an AI Agentic moment of surprise for me. I think many people who haven’t experienced such a moment yet will do so in the coming months. It’s a beautiful thing when you see an agent autonomously decide to do things in ways that you had not anticipated, and succeed as a result!

But given the fuzzy nature of some probabilistic LLM coding, as Mike Loukides writes for O’Reilly, if we automate coding then we also need better QA teams. There is also a deeper question lurking here about how this changes the way we see coding that points to potential for a real democratisation of programming in the emerging AI-augmented low-/no-code/prompt-led area:

The important part of software development is understanding the problem you’re trying to solve. Grinding out test suites in a QA group doesn’t help much if the software you’re testing doesn’t solve the right problem.

Software developers will need to devote more time to testing and QA. That’s a given. But if all we get out of AI is the ability to do what we can already do, we’re playing a losing game. The only way to win is to do a better job of understanding the problems we need to solve.

It is truly exciting to think about what well-supported, autonomous teams could achieve with this kind of stack inside organisations today, let alone tomorrow.

And if we look beyond prompts as the new command lines interfaces and towards conversational interaction with agents and LLMs, there is a lot of interesting work and ideas floating around about how narrow AIs can shape conversations to guide people towards questions they are able to answer, rather than just presenting a blank text prompt and running the risk of disappointing the user. This echoes the way video game AI characters often initiate guided conversations to achieve scripted goals even within open-world environments – it feels limitless, but in fact clever conversational prompts are used to guide players through a constrained series of interactions and options.

Reinventing organisations one team at a time

Ethan Mollick is bullish about AI as a catalyst for organisational change, partly as a result of his own practical experience augmenting teams with AI at Wharton Interactive:

The changes are profound. Theoretical discussions become practical. Drudge work is removed. And, even more importantly, hours of meetings are eliminated, and the remaining meetings are more impactful and useful. A process that used to take us a week can be reduced to a day or two. And these are just off-the-shelf projects built with GPT-4, not the most imaginative version of this sort of future. We already can envision a world where autonomous AI agents start with a concept and go all the way to code and deployment with minimal human intervention. This is, in fact, a stated goal of OpenAI’s next phase of product development. It is likely that, if GenAI’s hallucination rates decrease in future versions, entire tasks can be outsourced largely to these agents, with humans acting as supervisors.

This is the place to start in most cases: at the team level.

I had a catch up last week with Boundaryless, who are doing some great research into alternative organisational models and architectures. Their recent post on how to structure units and teams in a platform organisation makes a link between team types in the Team Topologies approach and the different roles needed in developing an organisational service platform – which is the best basic architecture for orchestrating a variety of AIs, automations and other digital services.

Each of the four main Team Topologies team types presents opportunities for Centaur services to support their work:

  • Stream-aligned teams – aligned to a flow of work in a business domain – can use AIs to automate basic process work, improve information flow and create their own Centaur services.
  • Enabling teams – help Stream-aligned teams overcome obstacles – can identify missing capabilities and use cases where AI can help fill the gaps.
  • Platform teams – provide internal services and products to be used by Stream-aligned teams – can scope and design narrow AIs and automation services that delivery teams can build into their own Centaur services.
  • Complicated Subsystem teams – run key systems requiring specialist knowledge – can apply AI to develop core services and capabilities that other teams can build on.

For Stream-aligned teams wanting to think of ways to incorporate simple AIs and automation recipes to create their own Centaur services, they might begin with process mapping and identifying scope for repeatable or standardised process steps that are ripe for automation.

For enabling teams, capability mapping can help define the missing pieces of the puzzle, and they might work with platform teams to develop the AIs they need to fill the gaps for delivery teams.

Platform teams can look across the wider organisation and identify common service components that different domains can use in their own Centaur service teams, and make sure they are available on the service platform and interoperable through common APIs and data schemas/ standards.

And for teams charged with running complex sub-systems, their build/buy/rent decisions will be key to providing the organisation with the right combination of utility, cost and flexibility for building out the AI and automation landscape that the other teams will build on.

In future technical deep-dive editions, we will look at some capability maps (see last week’s map for AI-driven talent discovery) and playbooks that could help teams design their own human-AI hybrid services and get them up and running.

But a key differentiator for any organisation looking to do this will be the organisational architecture and culture they start with. Highly siloed, centralised firms will really struggle to create connected services and capabilities at lower levels, but those organisations that already have a degree of decentralisation and encourage autonomous, agile teams will have an advantage.

Next week, we will share a longitudinal case study (i.e. not just the initial breathless press release!) of such an organisation to see what executives can learn from the ways that some incumbent firms have tried to create a healthier, more connected workplace and organisational operating system.

We hope you will continue to join us in this journey and share the learning as widely as possible. If you have any questions or suggestions about how we can better pursue our learning mission, please let us know - we would love to talk to you.