Is there a Pilot on the Plane?

In 2020, Airbus successfully showcased the first fully autonomous flight via the ATTOL project. The system was trained on 450 human flights. So why 6 years later do we still have an entire crew flying in?

All A320, A330 and A350 are already equipped with automatic landing systems. And with an Autoflight system for high altitude flying. We can already say a lot of the flight is already autonomous.

And if you fly private: Garmin Autoland integrates as a simple button for anyone to push, that will automatically direct the plane to the nearest airport, handle communication with the tower, and land.

The role of the pilot is not to operate the plane anymore. It's to take the necessary actions if the system breaks. And to be held responsible for the situation.

The Autonomous Revolution

Autonomous systems provide output based on the tools at their disposal. Since LLMs process unstructured data with a General Intelligence, the costs of building autonomous systems have dropped substantially.

Lock up the most intelligent man for 20 years, he'll not seem smart right out. But with the right tools, he'll get back up with great insights in a matter of time.

Combining a great toolbox with an appropriate learning loop, an LLM can make informed decisions and is always up for work.

Practical Example

At ARTE One, we built such an autonomous system for compliance work.

The majority of tasks our human team would handle comes from email outreach. So at each incoming email, our autonomous system classifies the conversation by importance and creates a three liner blurb.

Then, it attaches the conversation to the right project (eg: Cayman fund incorporation). Reading on all the project context, it creates a to-do item. One item could be to draft a response back for example.

If the agent needs more context about the response, it can search on its drive for all information regarding the company. If the email asks for a specific document (eg. KYC Pack) it can serve it away.

And if the document is non-existing, it can create a perfect PDF with information from its context, past document, or even ping the appropriate person at the company for missing information.

Every day at midnight, an outside LLM looks at the daily interactions happened, and modifies the tooling or prompt accordingly to better the performance of the system and append its memory. Like humans would.

Focus on the Learning Loop

A little note about the learning loop, as most of the value from the autonomous system comes from its ability to adapt strictly to the environment it is put out.

We have a dataset of events (eg. response to an incoming email) along with great behavior and wrong behaviors. The system proposes an answer, and it is checked against the expected behavior in chronological order, the way a backtest replays market days.

Then the system can memorize what it did wrong, and change its behavior. Would a new tooling help the case for example? Would adapting the wording of the mission streamline the operations?

The agent should, at regular intervals, adapt the learning loop to new situations the agent faced. This way, the agent can adapt to its environment, the same way an employee would.

The Role of the Pilot

In the past years, we've seen countless theories on the future of human labor. One sentence well heard is "You will not be replaced by AI. You will be replaced by someone using AI."

This perfectly describes the impact of assisted usage. By assisted usage, we mean using AI in accordance with your tasks. And it is true in a sense: all non assisted work will tend to under-perform assisted work. It's already the case.

With general autonomous usage being implemented sooner than later, human labor relying on some kind of intelligence will be replaced by companies adopting autonomous usage.

So why do we need humans if not for intelligence? We showcased a system able to adapt to any situation and grasp deep knowledge both on the macro situation and specific to the company.

But if the system fails, especially in compliance, damages could be catastrophic. This is why we keep the pilot in the loop: his role shifts from an operator, to a manager.

Before an email is sent, or a document is created, the manager has to approve it. But reviewing operations made by an autonomous system is not a full time job.

Reviewing several is.

Moving up the hierarchy

Pre-AI, human lands at a job as an operator. He then manages a team of operators, and finally an entire department. And the best CFOs are not generally the ones which calculate VAT the best.

The more we move up the hierarchy, the more abstraction we integrate relative to work being done. And the more the role looks like politics: filling up the space in orderly systems, to grease it and improve human collaborations.

Naturally as the role of operator will be filled up by autonomous systems, the current workers with domain knowledge will move up the hierarchy to become managers.

They will monitor the work of their subordinates. Much as is already the case today, a great part of a manager job is to handle situations where the system created by the company failed or could have done better.

The Role of Taste

Because one thing for sure, a new recruit - or even an experienced one - does not have the higher context of the company. The things we say between lines or informally (this is where politics comes in handy).

The role of the manager is also to add taste onto the work of operators. Taste can be described as doing the right thing. And it's best to know the right thing when you have the overall context.

Taste can be shown on a daily basis, but even on the data used to train the autonomous system. As most of the value derived from the autonomous system comes from its training, the role of the manager will be to select the best cases and the worst ones to indicate best practices.

And I expect we'll see more and more acquisitions of companies, mainly for their internal data to train on: email communications, Slack messages, internal documents...

The More the Better

What happens when greatness becomes a commodity?

Up until today, qualitative work beats quantitative. But when quality becomes the baseline, the only differentiator is to produce the best quality, at scale.

The companies which will differentiate themselves and win markets are the ones which will adapt to this new system, where actual output will be the measure of greatness.

As with the Asian industrial revolution, where labor was cheaper with same quality, the companies which did more succeeded by delocalizing their factories. The ones which did not, died.

The training data will be bought out or created for systems to learn from. Then the role will shift from operator, to reviewer. And finally, to maintenance.

And then?

One could say, autonomous systems could easily orchestrate and replace the newly managerial role of the human. I suspect they could, but not without impacting the quality of the output.

As we tend to attach more value to work produced by a human. For accountability reasons, as we have someone to talk with who understands us, and to make society whole.

This managerial role is very likely to persist a great amount of time, not out of technical necessity, rather because accountability and value needs a human name attached.

It will shrink as institutions, not brains, adapt.