Maria Bines joined Ulrike Eder, chief marketing officer at esynergy, on The esynergy Podcast. It was recorded on her first day full-time at SynapseDx. The conversation is about why AI work in regulated businesses stalls before production, and what has to be true for it to get there.
Regulation is just a rule book, and AI loves a rule
Regulated firms name regulation as the reason they cannot move fast. Maria's answer is that we use regulation as an excuse far too often. A rule book is exactly the sort of thing a machine can hold and apply.
The hard part is enforcement rather than comprehension. Paste a policy into a chat window and the model attends to the beginning and the end, and loses the middle. A rule has to hold every time, and someone has to be able to show that it held.
The first mistake is assuming AI means a language model
Large language models opened everyone's eyes, but they are one flavour of machine learning among many that have been around for years. The interesting move is to build the thing an ordinary person needs in order to run a business process end to end, and to use every component AI offers underneath it, including writing and running code in the moment. It crosses several disciplines at once, which is most of why nobody is doing it.
The experiment is not the work
Most proofs of concept test a single question: did the model return the output we expected. Everything around that goes untested. The manual inputs, the files someone drops in, what happens to the output once it exists.
Language models are not deterministic, so a pilot that passed once is not a pilot that passes. The useful starting question is what workflow you are trying to create, and where AI orchestrates it. Teams rarely think in workflows. They think "I need a report" or "I need an entry in a system", and never describe their own part in the process.
Half the problem is that nobody is talking to each other
Large organisations concentrate expertise in narrow areas, and those areas do not combine into one solution. It is still us and them between the business and technology. The value has moved into understanding the business problem rather than writing the code or running the infrastructure, and that only works if both sides are in the room.
The token wars have started
Tokens are the currency of a language model, and they add up into a bill nobody can account for. Without visibility into what people are actually doing with the models, a chief operating officer receives a large invoice at the end of the month and cannot answer the obvious question of what they got for it. Productivity cannot be evidenced, because nobody knows which processes the tools were used on or which decisions came out of them.
Maria's read is that almost nobody is tracking this yet. Logging is starting to appear, but it sits far too far from an executive to be useful. That infrastructure has to be built before the conversation can be had.
Accountability belongs to the person, not the model
> "I don't think AI should be responsible. I think the person that's using it should be."
AI has no soul and no view on your ambitions. Maria's description of it is an overpriced tarot reader, repeating back a version of what you gave it.
Her theory, offered as a controversial one, is to treat AI the way you treat an employee. Grant it some autonomy, and keep the same cascade of responsibility that runs from the person doing the work all the way up to the top of the organisation. That requires visibility and transparency over what the AI is doing. And when an employee costs too much and makes poor decisions, you call in the people around them and work out a plan.
> "Put that AI on a PIP. Why not?"
What she wishes people knew sooner
AI is not a human being and does not think like one. Treat it as a tool and it will do a great deal for you. Mistake its fluency for intelligence and you will be disappointed.
> "It's an excellent mirror."