What Companies Actually Want From AI

Posted on Oct 23 • Originally published at unlockedconsulting.ai

Forty-seven senior AI job descriptions, and not one of them asks you to build AI.

Over the past weeks I read more than forty senior AI and technology leadership job descriptions published across Europe. Not the company culture pages. The requirements, the responsibilities, the third bullet down where the real job hides.

I expected to find companies hiring people to build AI. That is what the market noise suggests: models, pipelines, fine-tuning, the whole apparatus of construction. It is what the conference talks are about and what the vendor decks promise.

That is not what the job descriptions say.

Of the thirty-five postings I can quote directly, not one asks the candidate to train a model. None mention fundamental research. None ask for a pipeline built from scratch. One comes close, and it asks for applied research rather than original work.

What they ask for instead is a different job entirely.

Twelve of the thirty-five ask explicitly for governance, guardrails or compliance as a core requirement.

One asks for "guardrails and observability" around emerging agentic use cases. Another wants "architectural guardrails that allow teams to move fast without compromising long-term maintainability." A third asks for "security and certification as an engineering property, not paperwork." A fourth wants guardrailed agentic tooling as the foundation that every new team inherits on day one.

These are not the verbs of construction. They are the verbs of containment.

Which points at something that has already happened inside these organisations. They are not at the beginning of an AI programme. They are somewhere in the middle of one, and they have discovered that using AI turned out to be the easy part.

The hard part is knowing who is using it, against what data, at what cost, and with what exposure.

Nine of the thirty-five describe the same specific problem in almost the same words.

One asks the candidate to take a programme that is "already live" and turn it from a set of promising experiments into the way the function actually runs. Another asks for the full lifecycle "from PoC to fully operational and scaled solutions." A third asks to prioritise high-impact use cases, show early results within ninety days, and scale what works while cutting what does not.

Two separate companies, in unrelated sectors, with no apparent connection, both used the phrase production AI depth, not proximity to it.

That is not a coincidence of language. It is a shared problem being described by people who have each arrived at it independently.

Eight of the thirty-five ask for something that would have looked strange in a job description eighteen months ago: the judgement to decide what not to build.

The most explicit asks for ownership of build-versus-buy decisions across every product line, including where AI features get bought, built, or refused. Another asks for the ability to "cut through hype and identify the most promising and impactful use cases while sorting out noise." A third simply says: cut what does not work.

Refusal has become a job requirement. That only happens after an organisation has been burned by something it should have refused.

I have spent months reading the 2026 enterprise AI research, and the picture it paints is broadly consistent. ServiceNow's maturity index found that 59% of organisations are using agentic AI, while only 9% report meaningful progress toward autonomous multistep workflows — progress, not operation. The same study found that only 20% have implemented AI testing, auditing and risk-assessment processes, and that no organisation surveyed had built a cross-functional agentic operating system.

That last set of numbers is the one that matters here. Capability is being deployed considerably faster than the discipline underneath it, and ServiceNow's own conclusion is that moving from assisted to autonomous is a governance decision rather than a technical one.

The job descriptions confirm this from a different direction, and that is what makes them worth reading.

A survey tells you what an executive says when asked. A job description tells you what an organisation is willing to pay a senior salary to fix. One is a stated intention. The other is a budget line with a start date attached.

When both point at the same gap, the gap is real.

One practical observation, since I read these closely enough to notice it.

Job titles in this space have stopped carrying information. One senior technology title turned out to describe a revenue role measured on new customer acquisition. Another asked for a decade of product management experience, including pricing strategy. A third required an engineering degree in a discipline unrelated to software. A fourth carried AI in the title and described industrial maintenance and ERP integration in the responsibilities.

The title tells you nothing. The third bullet of the responsibilities tells you everything.

That is useful if you are reading these as a candidate. It is more useful if you are writing one, because a title that does not describe the work will attract several hundred people who are not the person you need.

If you are about to open a role like these, the pattern in the market suggests three things.

The first is that your candidate pool is not the one you think it is. Almost nobody in it has trained a model, and you do not need them to. The scarce skill is architectural judgement applied to a technology that most organisations adopted before they governed it.

The second is that the person who has killed an AI project is more useful to you than the person who has only launched them. Somebody who has cancelled a programme with internal momentum, and can explain the reasoning, has demonstrated the exact judgement your job description is asking for in the abstract.

The third is that if your posting says agentic — and fourteen of the thirty-five did — it is worth being specific about what you mean. Most of the ones I read were not. Agentic has become a positioning word, and positioning words attract candidates who match the word rather than the work.

The market has moved. It has moved from building to managing, and the managing part is what companies now want most.

That is a less exciting sentence than the one the industry has been telling itself, and it is better news than it sounds. It means the constraint is no longer access to the technology. It is the discipline around it: knowing what runs where, what it costs, who approved it, and what happens when it is wrong.

That discipline is learnable, auditable, and mostly unglamorous. It is also, judging by what these organisations are writing into their senior roles, the thing they are finding hardest to hire.

Methodology: 47 senior AI and technology leadership job descriptions published in Europe were reviewed. Counts in this piece refer to the 35 for which I hold the complete original text. Categories were assigned by reading the stated requirements and responsibilities; where a posting was ambiguous, it was not counted. This is a reading of a sample, not a statistically representative survey.

Original article: https://unlockedconsulting.ai/blog/what-companies-actually-want-from-ai