There’s a bit of backlash these days because AI doesn’t quite deliver on some activities, and the promise that throwing money and GPUs at it doesn’t seem too realistic. There are good reasons for that, that we’ll discuss at some point on this very blog. But let’s focus on what works and what doesn’t, and how it would change organizations.
Any process depends on its inputs, and LLM processing is no different : it can’t do anything if it doesn’t have data to process, so :
- It won’t answer a question you have not asked
- It won’t find information it was not trained on
So if your job is made only of processing data from a database and each of your action can be described in words, there’s a pretty high chance that your job can be done by an AI. It’s not fun, but it’s OK, because there are many jobs that cannot be put into text and/or numbers, and as funny as it may seem, an awful lot of companies are run by some kind of magic.
I’ve spent last couple of years doing carbon assessment reports, which is a very boring exercise, at least the part where you collect data. To do so, you spend a lot of time talking with the admin people, and generate a lot of friction by asking them to give you information about their company, information that is very difficult to find because it’s buried in a so called information system by a colleague that happens to have left the company. That’s not some data that will be used to train a model.
It’s not much better on the operation side, except the files, as opposed to the well standardized (lol) accounting and HR forms, are mostly blobs of gigabytes that cannot even be reused for the next production. I know a VFX studio that had a contract with an online 3D model shop because it was easier to send their own CG models there so artists could retrieve them at discounted price than trying to organize an asset management where they could find their own creations…
But that’s only data and processes, how would you describe the organisation ? I’ve spent time in companies where asking for an org chart would trigger bursts of laughter. So there are many reasons why those 95% of AI project failing : on top of just serving mostly the FOMO, top down projects will be very difficult to setup just because in most case there’s not much data to train the model on, so you won’t go past what the GPTs of the world can do for you.
But we can already see a lot of bottom up implementation happening, and it’s much more like the smartphone : at some point a lot of people use it, and ask their IT dept to have the company on their GPT pretty much the same as they wanted the CRM and emails to be on the phone. GPTs work really well in most cases as sidekicks, and with enough context they can be your dumb colleague or trainee with a big memory and knowledge on a lot of things, but can’t really do anything interesting if you don’t ask and verify. Twice.
Will that change organizations? Yes, certainly, because it changes the relation to time. A lot of tasks can be vastly accelerated by LLMs, but some cannot be compressed, and it will take time to adapt. A lot of the time we spend in meetings and writing emails is to align and structure a vision : it’s now much easier to harass your virtual trainee in natural language with a couple hundreds of iterations that can be processed on a relatively small infrastructure to get something closer to your idea so you’re more comfortable to share it with colleagues. A lot of compliance related tasks can be checked automatically, pretty much as simple automations, things that once were done as a Unix script, then in a Windows bat file, and these days with Shortcut on an iPhone, will be done by your auntie by asking her phone to do something, and she will not even know AI is involved. That will be the same for professional tasks, the way emails have replaced the mail.
Most companies have computers, few have a structured digital information system that is worth more than its analog equivalent. Can AI fix that? Maybe, but probably not current generation of LLMs, because it’s not about having more info, it’s about having relevant info and structure. More on that later.