AI helps humans ‘remember 35,000 decisions’ a day

Artificial intelligence (AI) is becoming smarter, cheaper and more capable all the time. So why are the productivity gains proving so elusive? If AI can make individual workers dramatically more efficient, why aren’t companies becoming dramatically more productive too?

During a recent London Business School Corporate 100 session, Keyvan Vakili, Associate Professor of Strategy and Entrepreneurship and Academic Director of the Data Science and AI Initiative, spelled out why the true gains are hiding in the thousands of everyday decisions that underpin knowledge work.

The data land grab

Vakili began by unpacking the technology itself. Building a large language model begins, he said, with an enormous data grab. That means scraping almost everything publicly available online for the model to feast on during training.

Companies with access to proprietary data have an edge here. Google, for example, can tap into YouTube’s vast archive of videos, converting speech into training data.

However, that also creates a vulnerability. “For this tool, ground truth is whatever people wrote on the internet,” Vakili stated. “That basically means you can do very interesting, funny or dangerous content poisoning too.”

He illustrated quite vividly the point with Google’s early AI summaries, where coordinated fake Reddit posts convinced the system to recommend adding glue to pasta sauce to thicken it up.

The productivity paradox

Recent advances, though, have made AI so much more capable than that. On individual tasks there is now little doubt that AI can lift productivity up, sometimes dramatically so.

Vakili pointed to studies showing customer support agents resolving queries 15% faster and knowledge workers completing sophisticated writing tasks 40% faster and at a finer quality.

“Nowadays we have plenty of academic evidence of this sort,” he said. Yet AI’s productivity marvel has, so far, largely failed to materialise at the macroeconomic or even organisational level, Vakili remarked, citing macroeconomic data, academic studies and reports from company chief executives.

The problem, he argued, is not the technology itself.

The irony is that this paradox is showing up even as models are becoming far cheaper and more capable. Vakili estimated today’s frontier models are many times more “intelligent” than GPT-4 three years ago while costing more than 90% less (per million tokens) to deploy.

Vakili argued that AI is still stuck on the downward slope of the productivity ‘J-curve’, as economists call it. Firms first pour time and money into changing the way they work, retraining staff and overhauling internal systems. But those intangible investments suppress measured productivity before eventually pushing it higher.

The hidden decisions

So where are those productivity gains hiding? Vakili commented they can be found in the invisible decisions embedded in everyday knowledge work.

“This is a massive bucket of missed opportunities,” he said. We make far more decisions and produce far more content than we realise. Vakili drew on research suggesting people consciously remember only around 70 decisions a day. The true figure is closer to 35,000 once the rapid, automatic choices are added to the tally.

Take the seemingly humble email. Every one that lands in our inbox triggers dozens of tiny decisions: whether to open it now or later, skim it or read it carefully, reply, forward or delegate it, who to copy, how formal to be and when to hit send. Individually, they barely register. Combined, Vakili stated they consume around 250,000 hours a year in a 5,000-person business with each employee getting 100 emails each a day.

“We are very bad at seeing this,” he said. “This is automatic. This is just background noise in our lives.”

The original iPhone solved precisely this problem. Apple’s engineers realised that typing on a touchscreen involved countless micro-decisions, and their solution was to make the letters users were most likely to press easier to hit by enlarging their invisible touch areas. “Without that algorithm, this phone would have probably never existed,” Vakili explained.

Jobs versus tasks

AI could have profound implications for the workforce. But the common mistake, Vakili noted, is thinking about AI in terms of jobs at all.

Radiologists were supposed to be among AI’s first casualties. But Vakili said reading scans is just one of many tasks they carry out. AI can help with a narrow diagnosis, but it is a long way from replacing the job itself.

And higher productivity does not automatically lead to layoffs, Vakili added. “Productivity cuts both ways.” Companies can use those gains to fatten their margins, hire more staff or leave headcount where it is. “It is a strategic choice.”

Exposure does not necessarily mean extinction. Software developers are “the most exposed to GenAI” yet also are “the occupation with the highest growth rates in the labour market”, according to Vakili. “If you do more software development, you need software engineers.”

Even if AI permeates more individual roles within an organisation, they are made up of many complicated, interlocking processes with tasks that cut across teams and departments, Vakili reasons. Making a few tasks more productive here and there doesn’t automatically ensure the entire process is more productive if the bottleneck is elsewhere.

Productivity v performance

One final point stood out: as calls to slam the brakes on AI development grow louder following a string of security failures, Vakili said that governing AI will matter just as much as building it.

“There’s no one accepted governance model. For every process you need one,” he said, explaining that organisations need to decide who is responsible for each.

Concluding, Vakili posits that while AI may result in increased individual productivity, it does not always translate into increased organisational performance.

It all comes down to how companies deploy this extra productivity. Just because we can release more apps or music with AI, it does not mean demand or consumption increases.