The mistake of demanding that machines think like us
Susskind's argument, which he set out at Exponential Day in July 2026, rests on examples that are now history: Deep Blue beating Kasparov, Watson winning Jeopardy!. Neither system played chess or answered questions the way a person does. They did not understand: they calculated and retrieved. And they still beat the best.
The lesson is not about chess. It is about how we assess whether a task is automatable. The question usually asked in a committee is "could a machine do what our best analyst does?", and the answer is usually no, rightly. The correct question is different: "is there some way — possibly very unlike ours — of producing an equivalent result?". There, the answer frequently changes.
This shift explains why so many companies are caught out. It is not that AI has learned to reason like a senior professional. It is that much of the output a senior professional produces can be reached by another route, and that route does not require the capabilities we assumed were essential.
The unit of analysis is the task, not the profession
From this comes the most useful part of the thesis. Susskind argues professions do not disappear at a stroke: they transform task by task, and the challenge is not mass unemployment but the mass transformation of professional profiles.
It is a distinction that sounds academic and is intensely practical. It completely changes what you do on Monday morning.
If the unit is the profession, decisions are binary and politically toxic: do we keep this department or not? The conversation turns defensive, people protect themselves, and nothing moves.
If the unit is the task, the conversation is different: of the seventeen things this job does, which change, which disappear and — the one almost nobody asks — which appear? That conversation can be had without threatening anyone, it produces a concrete redesign, and it is the only one we have seen lead to real adoption.
Where our reading adds a caveat
There is a point where field evidence qualifies the optimism implicit in "transformation rather than substitution".
Profile transformation is neither automatic nor free. A profession surviving in redistributed form does not mean the specific people practising it today will move frictionlessly to the new version. The task shift concentrates the load on a kind of work — judgement, verification, exception handling, responsibility for what the machine signs — that not everyone wants or can take on at the same pace.
In operational terms: the thesis is right in aggregate and harsh at the individual level. A company that uses it as a sedative — "no need to worry, there will only be transformation" — and does not invest in the transition will meet the harsh part having prepared nothing.
That is why, in the people layer of our maturity series, we insist on something that sounds unfriendly: if the project will change someone's work, say so. The credibility lost by promising nothing will change does not come back.
His answer, and its limit for a company
His main recommendation is educational: rethinking training so it is not concentrated in the early years of life but accompanies an entire career. It is hard to disagree at the level of public policy.
At the level of a mid-sized company, however, "continuous learning" is exactly the kind of answer that produces activity without result. We already see what happens when it is translated literally: literacy sessions for the whole workforce, high attendance, zero operational change three months later.
What does work, and is the operational translation we propose: training anchored to a specific process being redesigned, for the small group that runs it, at the moment they run it. It is no less ambitious than lifelong learning; it is lifelong learning with somewhere to land.
On AGI, briefly
Susskind holds that artificial general intelligence deserves to be taken seriously — not as a certainty, but as a credible scenario given the level of investment, the concentration of talent and the rate of acceleration.
Our operational position is that this discussion should not condition any business decision in the next twelve months. If it arrives, it will change so much that planning for it today is sterile; and if it does not, the company that deferred concrete decisions waiting for it will have lost two years. What decides the outcome within the horizon a leadership team can govern is far more prosaic: which process you measure, what context you have written down, and who signs the outputs.
What to do with all this on Monday
Three actions that follow directly from the thesis and fit in a quarter:
- Break down a job, not a department. Pick one concrete role in the process you are already looking at and list its tasks. Mark which change, which disappear and which appear. The exercise takes an afternoon and usually dismantles half the committee's assumptions.
- Look for the equivalent result, not the imitation. When assessing whether something is automatable, do not ask whether the machine can do it the way your best person does. Ask whether there is another route to the same result, with human verification where it matters.
- Name the transition. Who supports the people whose work changes, with what time and on what timeline. If nobody holds that brief, profile transformation stays a good conference argument.
Conclusion
Susskind's thesis disarms the two comfortable positions that coexist in many boardrooms: that of the person who believes their sector is safe because their work requires judgement, and that of the person who believes they must prepare for mass redundancies. Neither describes what is happening. What is happening is a redistribution of tasks within jobs, at a speed that depends, largely, on decisions each company makes.
That is reasonably good news, on one condition: that somebody does the work of looking at the job task by task. Technology does not do that work.
Frequently asked questions
Will AI destroy jobs in mid-sized companies?
The best-supported thesis today — and Susskind's — is that the dominant effect is not the disappearance of professions but the redistribution of tasks within them. That does not make it painless: the load shifts towards judgement, verification and responsibility, and that transition has to be supported explicitly.
Which jobs are affected by generative AI?
Unlike earlier waves of automation, it reaches non-routine tasks belonging to skilled professions. The useful boundary is no longer manual versus intellectual, but tasks with a verifiable output versus tasks whose value lies in relationship, accountability or judgement in ambiguous contexts.
Does AI have to think like a person to replace a task?
No, and assuming otherwise is the most common assessment error. Deep Blue and Watson beat the best humans by routes entirely unlike human ones. The right question when analysing a process is not whether the machine will reason like your best professional, but whether another route to an equivalent result exists.
How does a company prepare for profile transformation?
By breaking specific jobs into tasks before deploying tools, redesigning the process around the group that performs it, and naming someone responsible for supporting the transition. General AI literacy has cultural value, but without a process to land in it produces no operational change.
Jordi García is Tech Lead at onext. He works on bringing AI into governed production across development and product teams —with Spec-Driven Development, context engineering and human verification at every step— and authors onext's technical insights on the method, quality and cost of applied AI.
LinkedIn →