The synthetic organisation: When machines do the work and humans sign off

By Dr Vaishnavi Gautam, Assistant Professor in OBHR area, IMI Bhubaneswar

As AI drafts, summarises and scores work largely for other AI systems, people risk becoming little more than signatories. The firms that prosper will be those that keep human judgement where it matters most.

The office that is not empty
The office of the future may not be empty. It may instead be full of people approving work they did not create, metrics they do not fully understand and decisions they do not truly own.

Consider how a graduate now applies for a job. The application is drafted with AI. The employer’s screening software reads it, ranks it and rejects most of the pile. An automated tool summarises the interview. A hiring manager, pressed for time, approves a shortlist that no person assembled. Each step is efficient and everyone involved appears productive. Yet it is difficult to say where, in this chain, anyone actually thought. It is the early form of what might be called the synthetic organisation: a firm in which machines produce work for other machines, and people enter chiefly to sign off.

Bureaucracy without readers
Organisations have always generated paperwork that few people read. What has changed is its cost and its audience. Generative AI makes reports, proposals, summaries and compliance evidence almost free to produce, and their first reader is increasingly another system that summarises, scores or acts on them rather than a person. The result is ‘synthetic bureaucracy’, documentation produced largely so that other software can process it.

Two forces have accelerated this synthetic bureaucracy. First is ‘workslop’, the term for AI-generated work that appears finished but lacks substance. In their survey of American desk workers, about 40% had received such work in the previous month, and each instance took nearly two hours to repair. Second is ‘measurement’. Dashboards track outputs, response times and activity, precisely the signals AI can inflate at almost no cost. When metrics reward what machines make cheap, organisations receive more of it and mistake volume for value.

Approval is not judgement
The customary safeguard is to keep a human in the loop. A manager presented with a polished recommendation and a confident score, after the system has already framed the options, is seldom exercising independent judgement. Research on automation bias has long shown that people tend to defer to machine outputs, particularly under time pressure and when those outputs appear authoritative.

Regulators have recognised the distinction. For instance, European data protection authorities have made clear that a person who merely rubber-stamps an automated decision does not turn it into a human one. Review must be meaningful, carried out by someone with the authority and the information to change the outcome. When a synthetic process rejects a candidate, declines a loan or downgrades an employee, a signature at the bottom does not reveal who decided. It reveals only who will be held responsible.

India’s apprenticeship problem
For India, the stakes are unusually high. Its IT services firms, consulting practices and global capability centres rest on a pyramid: a broad base of early-career professionals who learn by researching, drafting, testing and documenting. These are precisely the tasks generative AI now performs best.

The danger is not only fewer entry-level roles. It is that the roles which remain turn into validation work: checking machine drafts, approving machine summaries and correcting machine code. Validation cultivates approval rather than judgement, and judgement develops only by wrestling with a problem before the answer is available. A generation that spends its formative years signing off may reach senior positions fluent in review but unpractised in decision-making. India’s knowledge industries could protect this year’s margins while eroding the expertise that clients will pay for a decade from now.

Where humans must decide
The answer is not to slow the adoption of AI. It is to be deliberate about where human judgement sits.

First, map the points of consequence. Hiring, credit, safety, performance appraisal, pricing and strategy involve decisions whose errors are costly and difficult to reverse. Boards and leadership teams should identify these explicitly rather than leave them to whichever team happened to configure the software. In most firms, this map has not yet been drawn.

Second, place judgement before the recommendation, not after it. At these points, the accountable person should record a view, or at least the criteria that would change their mind, before seeing the system’s output. Friction of this kind is not inefficiency. It is the price of a decision that someone genuinely owns.

Third, audit machine-to-machine loops. For each major workflow, leaders should ask which documents are written by AI and read only by AI. Some loops are harmless, such as meeting notes summarised for a project dashboard. Others are where errors compound unseen and accountability quietly disappear.

The competitive advantage of the coming decade will not belong to the firms that automate most. It will belong to those that retain human judgement at the points of consequence and refuse to mistake a signature for a decision.