A survival guide in the race against robots for companies and their employees

No AI was used in the penning of this review, but it was translated into Romanian using Claude Sonnet 5's API. That version – which had to be painstakingly edited by hand for reasons we'll come back to – can be read on standard.ro

Move too slow and be left behind, too fast and get crushed by cost: this is, right now, the conundrum for companies facing the advent of ever more powerful language models. But then, even if you somehow get the pacing right, things may still end up badly – whole industries may go bust, bubbles may pop. For employees, the question is a bit easier: will a robot take my job or not?

We're exploring possible answers to these questions with a book review, two complementary angles from high-profile bankers, one scientific study, and this reviewer's own recent experience.

The setup

"A lawyer friend told me the painful part was not that her craft might be replaced. It was that the self she had built through two decades of work might no longer be needed." This startling account comes early in the book The Work That Remains by Daniel Dines – the founder and CEO of automation giant UiPath.

Sure enough, the quote is acting as Chekhov's gun, with Mr. Dines' advice for his friend coming around right before the end of the book. No spoilers here!

The book is a two-to-three hour read, freely downloadable and admittedly written "with Claude and ChatGPT", in a complex back-and-forth that the author thoroughly describes in the book. In a nutshell, to write with AI fruitfully, expect to take longer, not shorter than doing it by hand.

The things people do better

The core claim of the work is that AI has limits that will not be overcome too soon. It's prone to degrading fast, even from seemingly high levels of accuracy, and it can't really know the world, only what is explicitly said about it. For companies, this epistemology is critical: an AI agent can't know how a company works because the company itself doesn't truly know how it works.

The important knowledge resides in the mind of its employees. And there are things that people are good at in ways that machines wouldn't even know where to start: reading the room, taking cues, following hunches. Dines brings on the example of bridge, the only well-known game where humans still beat superintelligent machines: "Chess fell to AI decades ago; Go fell; poker fell, hidden cards and all. Bridge still stands — and that should give an executive pause, because bridge is a tiny world. Four players, fifty-two cards, one rulebook, and meaning shared between two people. An enterprise runs the same game with thousands of players, a rulebook nobody has fully written, and meaning shared across departments, customers, and years. If shared meaning keeps the machines from mastering the small version, consider what it means for the large one."

It's the minute, never spoken details that make a world of difference. Here are three examples that draw from real-life:

Reading every book about skiing is not skiing. Reading every customer escalation is not twenty years in customer support.

A child learning to cross a street is the purest case: the lesson is not «look both ways». It is the parent's grip tightening at the curb, the pause that lasts longer on this street than on that one, the sharpness of a correction when the situation deserved it. The knowledge is a pattern of attention. Only a sliver of it was ever said out loud.

A general-purpose model with twenty years of one company's history loaded into memory is not the same as a system shaped by twenty years inside that company. The information is present. The transformation is not. A chef who has cooked Italian food for twenty years and a chef who has cooked Japanese food for twenty years, given the same recipe, produce two recognizably different dishes.

Then there's a layer of consequences. AI has no skin in the game, so it will most often fail at things that can't be rehearsed safely and that don't have clear, predictable outcomes: "how a chess engine became superhuman: it played, and lost, millions of games before it was good. The enterprise offers no such luxury. Every lost game here is a real payment, a real customer, a real regulator — nobody gets a million free losses."

In order for AI to work efficiently, there needs to be a thorough map of the environment it operates and a set of rails so it does it safely, the book suggests.

This clicks in place nicely with what people working in the frontlines of industrial automation had to say about their own challenges, in a Standard article about AI-driven forklifts, one of the current sweet-spots between useful capability and potential generated value: „Where safety is on the line, we don't leave anything to chance. We iterate through every possible scenario so it never makes a bad call”.

The challenge for UiPath

Mr. Dines' book puts forward a three-pronged path for the future: "AI proposes. Humans decide. Automation executes." A disclaimer that he makes becomes warranted at this point: "I run an automation company. That is my declared bias."

Right, so let's unwrap this: Generative and language models stole the show, taunting that they might never give it back. It's basically this widespread belief that "no need for software, everything will be done with AI". Daniel Dines pushes back on that, saying it's too expensive, it leads nowhere and the right architecture includes the services that his own company provides. Hence the "declared bias": being proven true is existential.

UiPath (PATH), the Romanian-born automation giant listed on the New York Stock Exchange, has been in the news lately with its share price swings up and down. The company is now down 80% from its debut peaks 5 years ago, and AI hype contributes to that as expectations beat fundamentals. Automation done right can save hours of an empoloyee's repetitive work daily, in a measurable way. But right now plenty of companies simply expect everything "to feature AI" – useful or not – just so that they don't miss out. All this had the company start a transition from its Robotic Process Automation model to Agentic Automation, a process where AI is used to orchestrate automated pipelines.

To be fair, the market context is existential for frontier labs too, who may have overinvested with little to show for it. Messrs Altman and Amodei may often keep expectations high with their hype and doom, or when they allegedly forget to keep their bots leashed, but the bubble crash that might come is all the talk now, with Bridgewater's hedge fund veteran Ray Dalio now in the headlines saying that it's imminent.

The vibe in Romania

Bankers, of course, are hedging their words. Here's from Leonardo Badea, the second-in-charge at the National Bank of Romania, who has lately started to publish his own analyses, such as The Implications of AI on Monetary Policy (3-minute read, free signup required).

"My reading is this: the impact of AI will be profound, but delayed and uneven. For central bank policy, this lag has an important impact. While the money is being spent on data centres and chips, AI is a demand shock: it lifts investment and can add to cost and price pressures. Later, as adoption lowers other costs and raises output, the same force turns disinflationary. The difficulty is that a central bank may need to act before it can tell which regime it is in, so as not to be behind the curve," the NBR's first vicegovernor says.

Mr. Badea shared some of his findings at a conference held by NewMoney magazine at the University of Economic Studies in Bucharest – ASE, an institution that has been running a dedicated programme for applied AI since 2020. An analysis regarding AI workforce readiness in CEE countries – conducted by a team of the University's researchers that includes former Standard contributors Aura-Gabriela Socol and Cristian Socol – placed Romania at or near bottom among peers accross variables, in a race led by the Baltics and Slovenia.

Here's one particularly interesting take by Florian Libocor, the former chief economist of BRD: Artificial Intelligence and Reconfiguration – Employment Between Transformation and Disappearance (5-minute read, LinkedIn signup needed). "Most debates focus on the disappearance of a large number of jobs. However, I believe that the deepest fear should not be that AI will replace human beings, but that this disruptive force will transform the economy and society much faster than people and institutions can possibly adapt," says Mr. Libocor.

He pushes back against common wisdom clichés like plumbers are safe: "Initially, standardised cognitive work will be reorganised before complex physical work. In other words, white-collar workers will not necessarily be the first to disappear." Specifically regarding Romania, "pressure may emerge in routine IT — particularly the 'middle' Low-Code/No-Code segment, which is tending to disappear — and in BPO (Business Process Outsourcing); recruitment for administrative roles may slow," the author writes.

The most striking thing, however, is the timing: the paper came out exactly one week after Mr. Libocor was laid off from BRD, with the analysis department disbanded completely, an unprecedented move that took the banking sector by storm. Granted, no AI was mentioned, but this still begs the question: would the bank now be tempted to use LLMs for economic analysis?

The takeaway

Which brings us back to Mr. Dines' advice for employers and employees.

Companies may benefit short-term from bringing in generative AI and language models to replace people, but cost will arrive long-term, down the line. "Cut headcount because agents are expected to absorb the work within twelve months, and you may hollow out the institution before the capability arrives. When the AI underdelivers, the people who knew how the place actually worked are gone."

Going the opposite route might even be the right call: "Companies that stop hiring juniors will look lean immediately. The deeper cost arrives later, when the senior layer hollows out and rebuilding the pipeline takes a decade. The labor market adds a twist. If most companies stop growing juniors, the few that maintain serious pipelines gain a supply of future senior talent the cost-cutters cannot manufacture quickly," the book says.

Workers should seek what is truly human about AI, and examples abound within the book, but this sums it up: "The most durable work is whatever the institution needs carried by name — trust, culture, mentorship, accountability — because no improvement in models manufactures a self with something to lose." Also, don't bet on ignoring the issue because it's not going away, he warns: "People who combine the durable human layer with AI-native speed will compound an advantage over those who treat AI only as a threat or a convenience."

For a conclusion, here's a piece from the introduction: "Whoever tells you [AI] is just a better version of autocomplete has not worked with an AI agent lately. Whoever tells you it means the enterprise now runs itself has never run an enterprise. Both are loud. This book is for the people who have to decide between them."

I failed to make an AI newspaper so you don't have to

The story thus far matches this reviewer's own recent experience, with some twists.

Last year, I discovered the capabilities of agentic AI (with all the fancy stuff – vector databases, feedback loops, retrieval-augumented generation). It seemed like a whole new world had opened. Over the course of three months, I staged a simulated pipeline for an AI-generated newspaper: scrape sources, ingest press releases, search for background, fact-check, apply in-house style and policy. It was all fine and dandy, with the test proving that the bot crew could spit out to the tune of 50 to 100 pieces daily mimicking The Economist. The trouble was manifold.

At the stage where I stopped the experiment, articles suffered from a lack of authentic human voices (so-called quotes from press releases are mostly useless) — people being bold, assertive, defensive or concerned. Sure, we could have the bots mail out questions, but no one will ever bother replying to that.

Alright, let's crawl LinkedIn for opinion and sentiment! But LinkedIn was just showcasing its option to create posts with AI, and a flood of AI-generated slop soon ensued, making the social network unreadable. It hit them hard, and now they've done a 180-degree pivot to add a button flagging that something "[s]eems like AI slop".

One funnny note: In Romanian, the button is translated to the effect of "[s]eems like an AI error", which is itself a self-referential yet likely unintended showcase of AI slop.

The other problem with that is a self-selection bias, where the loudest are not necessarily the most competent. Most of the times, value comes from asking the right people the right questions.

Obvious AI tell-tales — the it's not this, it's that. the quietly. the delve into. these chopped sentences. those long dashes — were programmatically weeded out. Still, the uncanny feeling reading the output was there. There is a loss in texture that I found no way to weave back, even though I asked Claude and ChatGPT themselves to search far and wide for a solution.

All they could reliably tell me was that "it is a known issue and point of pain," that they are struck with this terrible degenerative disease conspicuously called MAD – Model Autophagy Disorder. Its pathology is heart-wrenching: after a number of iterations re-ingesting its own output, sometimes as few as five, an LLM's output becomes utter nonsense, with the model effectively collapsing into itself.

With Romanian, the issue is particularly worse. The data corpus is bad: too much word-salad, copy-pasta, boilerplate and gobbledygook made it to print and web, and from there into training data. It sounds bad from the first run, no matter how "good" the model is (here's a tip, though: Chinese Kimi sometimes owns them all). Some tried to brute-force their way out of this, and only when the bill hit, did they google "tokenmaxxing".

There were some ways out, tedious and expensive, but then what? If it works, then everyone does it, triggering a race to the bottom, that will likely be won once more by Google.

Even if none of this were true, there's still one thing left: your writing is your thinking (read Juliette Ryan's brilliant "The Mind Virus" and Zawn Villines' compelling "AI 'Writing' Is Not Writing" on Substack!). I mean, sure, we ride cars and jets, but that will never replace running. As a matter of fact, more people started running marathons as transportation got faster and cheaper. They're all different means to separate purposes.

A good prompt will usually take more effort than doing the writing itself, so why not just write? Sure, one may take issue with that because it lacks scalability, but a system prompt is never a good prompt; it always needs to chain further, and it will fail sooner rather than later.

Just like LinkedIn, I dropped my AI ambitions and went the opposite way. This may actually be the right time to do good old, trustworthy, on-the-ground reporting. Generative work will find its uses, and already has plenty in journalism. But intent and agency tightly stay with humans – "not because" we're terrified of being left out, "but because" there's really no point to anything else.