AI World Building & The Value of Boring BizOps
3 new AI developments, musings on physical AI and world building, plus how to harvest value today in what might seem like boring business support functions
The roaring AI train didn’t slow down at the end of the year as many of us humans needed to; and as 2025 gets underway, major players are still throwing coal into the furnace:
OpenAI has been heavily hinting that they think they are moving beyond AGI (but what does AGI really mean?)
The Chinese DeepSeek model was released towards the end of the year, but we now have enough testing and analysis completed to indicate it is an important advance in compute efficiency for training (it needs 11x less compute than comparable models).
NVIDIA has open sourced a new AI video model that they are using to build world models to accelerate physical AI such as robot development and autonomy.
(Incidentally, if you were offline over the holidays, check out our 2024 wrap-up and deeper thoughts about the direction of enterprise AI in 2025)
World Building for Physical AI
The latter development is interesting because agentic AI, although developing rapidly, still lacks a reliable model of the physical world.
World building and the use of digital twins to model predicted outcomes is a key foundation in enabling AI to understand reality, rather than just re-combining knowledge and data in new forms. Tesla famously opted for a vision-only approach to understanding the world around their cars - to reduce cost by avoiding the use of LIDAR that enables 3D vision. The boss of China’s leading electric car company BYD recently said that if Elon Musk had ever driven on a Chinese highway at night, he would immediately see the need for LIDAR, which BYD are introducing in new vehicles that are about half the price of the cheapest Tesla. I would also add that trusting Tesla’s FSD with its mostly USA-based training data is not something I would advise in Portugal, Italy or Egypt where people drive a little less like Americans.
The cameras vs LIDAR question highlights the danger of taking a highly reductive view of human capabilities. Our optical systems are not just simple low-res cameras hooked up to a computer, just as our brains and consciousness are not just computers and our neural pathways are not just cables.
World building is such an interesting area of creativity. Some of the best movies and novels are so immersive precisely because they are so good at world building. And the best games take this to a new level, even without AR/VR adoption. I cried at the end of Metro Exodus, an old game that I finally played through over Christmas. I love the Far Cry and Atomic Heart worlds, with their own weird little rules. And I think Cyberpunk 2077 Phantom Liberty is one of the best C21st works of art to date, mostly for its incredible world building.
Living creatures seem to spend a lot of time creating imaginary worlds to develop their skills and understanding, whether it is a cat playing at predator with a toy or a child having a tea party for its dolls.
But we can also delude ourselves, so the worlds we imagine don’t always match objective reality. Will we (or crazy Uncle Elon) do the same to our world models to replicate our own biases? Or will we create our own separate world bubbles with their own models of reality and ethics, as the progressive business group Mondragon has done, and as crypto bros like Peter Thiel dream of achieving for altogether different reasons. In the new post-truth world, could we end up with different worlds with their own training data, creating feedback loops that divide us even further?
Human-AI collaboration in both the virtual and physical world is something we will need to explore and understand in all areas, including the enterprise. Next week we will cover this in greater depth as a management technique that is important for organisations implementing AI.
BizOps: boring work is where the value is for now
Late last year, consultancy BCG published a report entitled Where is the Value in AI? The most striking conclusion of the report is the proportion of expected value to be found in ostensibly boring business support functions:
Core functions and business operations could represent 62% of the value to be harvested through cost savings and revenue gains. But I don’t think this captures the whole picture, because the term ‘business support function’ disguises the fact that within most organisations these functions are perceived as blockers of value creation. So by reforming them and making them more automated, we might also unlock efficiencies elsewhere in the value chain.
Diginomica recently highlighted the potential for AI within finance and accounting functions:
Aurorium, a producer of specialty ingredients, uses AI on its financial data to predict order volumes and make better production planning decisions. “[AI] shows us which customers are due to place orders soon,” says Robert Franz, Senior Reporting & Systems Analyst at Aurorium. “That lets us run a bigger batch, which is more cost-effective.”
The piece also pointed to the potential for recycling time saved into more strategic activities:
It's common for a CFO to tell me their financial performance management solution saves them hours per day, days per month, and weeks per quarter. But, I've never heard one say they used that time savings to work less. Sure, they don't work late as often, but none are working four-day weeks. Instead, they’re focusing less on tactical work and more on strategic, needle-moving decisions that will guide the organization’s long-term success.
PYMNTS also recently covered the use of AI in back-office efficiency:
Traditional accounts payable (AP) and accounts receivable (AR) processes are laden with inefficiencies, from invoice approvals to payment collections. AI-powered solutions can automate invoice processing, flag discrepancies, and predict payment behaviors. Tools like machine learning algorithms can also help analyze payment patterns to improve cash flow forecasting, giving CFOs and treasurers a clear view of working capital.
And Business Insider suggested that these developments demonstrate that enterprise AI is already starting to live up to the hype:
The AI boom has added trillions of dollars to tech company valuations. Is it living up to the hype?
In some real ways, the answer is yes. This is especially true when it comes to the technical plumbing of modern companies. These are tasks that often go on behind the scenes and are either unknown or taken for granted by most non-technical people.
But this is not just about finance functions. Similar use cases and opportunities can be found in HR, legal, compliance, internal audit and even facilities management.
Our experience over the years in helping legal and internal audit functions transform by making better use of technology and connected data suggests that there are plenty of smart people who play with technology ‘under their desks’ to get things done more efficiently than their formal systems allow, and I would imagine these people are already becoming secret cyborgs, using AI to improve their work.
Rather than leave this to individual initiative, which also carries risks, it makes sense for BizOps leaders and function heads to assess each grouping from the ground up, looking at its process landscape, manual work, systems and data, to identify opportunities for smart automation and delivering a better customer experience for their colleagues inside the organisation.
The Business Insider piece above cites Vercel’s CFO using AI to write his own code to solve problems, and we are seeing many examples of people in business operations functions who are doing the same, scratching an itch or filling a gap by coding with the help of AI.
In more technological firms, there is another category of people who are also doing something similar: people who understand the basics of coding (former developers or just nerdy tinkerers), but who are now in management positions, as Zohaib Rauf - who transitioned from developer to engineering manager five years ago - describes:
Because my time has always been limited, progress on side projects had been slow in the past, and many remained unfinished as life's events caused a loss of momentum, making them harder to resume. However, in the last year (2024), I have been very productive with my side projects, quickly building the tools or projects I need and deploying them for others to use—in other words, finishing the v1 of each project.
LLMs in general have been immense booster for my productivity when it comes to side projects and more specifically the Cursor IDE has been a great editor to use these LLMs for coding.
He goes on to share some great advice for how to make the most of LLMs and get these small projects over the finishing line to be useful.
Maybe there will come a day when these side projects will create personal robots that understand our world model, but for now I would be grateful for software bots that do the boring stuff so that I can spend my time on more creative, human tasks.