AI Gave Professionals an Executor Before They Learned to Manage One

The Accountant, Codex and the Missing Second Instruction

In a classroom exercise, an accountant asked Codex to complete a task. The system returned something that did not match her requirement. She noticed the problem — and then stopped. She did not explain what was wrong, did not ask for a different result, and issued no further instruction.

The moment matters, according to a lecturer who teaches AI use on MBA, EMBA and DBA programmes at four business schools in the region, because it separates two distinct abilities. The accountant could use the tool and could recognise a bad result. What was missing was the conversion of “this is not right” into “do it differently.” To the machine, her silence contained no information.

The essay argues that this is a corporate AI adoption barrier that appears after the first answer, not before it. The old rule “if you want it done well, do it yourself” now breaks down: closing the chat and redoing work manually saves quality but changes nothing about the production method and yields no productivity gain. The updated rule is closer to “if you want it done well, manage it yourself.”

A July 2026 study by Emma Wiles of Boston University and colleagues at BCG adds evidence on the oversight side. In a randomised experiment with 813 HR and finance managers, participants reviewed the same flawed documents while being told the author was an AI tool, an “AI employee” called ALEX-3, or a human. In companies that had already put AI agents on organisational charts, the “AI employee” framing was associated with 18% fewer errors detected and some responsibility shifted to the system.

Why the AI Adoption Bottleneck Sits After the First Answer

Delegation Has Reached Specialists Before Management Skill Has

The core claim is that AI made delegation a mass activity without first making management skill mass. A bookkeeper now assigns a calculation to a model, a lawyer assigns document review, a programmer assigns a code change. Previously, regular delegation arrived with a managerial position and human subordinates; now the specialist becomes the client and organiser of part of the work. That is why the article describes the required capability as “profession + management + AI,” not three separate training topics: professional judgement evaluates the result, management organises how it is produced and accepted, and the AI component accounts for the machine’s lack of implicit context and its risk of plausible error.

The Risk of Turning a Tool Into a Colleague

The Wiles/BCG experiment suggests a subtle oversight risk. The finding does not show that AI output was better; it shows that, in AI-mature organisations, reviewers found fewer errors and shifted responsibility when output was labelled as coming from an “AI employee.” That matters for finance and HR documents, where undetected mistakes carry real cost. The essay extends the warning to ready-made agents: an “analyst” or “sales” agent already contains someone else’s assumptions about goals, quality, sources and risk. Unless those criteria are checked, a company may import a foreign model of its own work.

Correcting the Process, Not Just the File

For a one-off report, manually correcting the output is enough. For a monthly report, it is not. The essay compares repeated manual correction to filing each part instead of resetting the machine. The higher-value move is to change the instruction, source, template or control procedure, then reproduce the corrected order. A specialist may not have authority to approve a new instruction, but should at least identify the systemic cause and pass it to whoever controls the process. Otherwise each corrected file saves one report, while a corrected instruction saves all future ones.

Closing the Gap: From ‘This Is Wrong’ to a Working Instruction

For professionals and managers, the piece points to several concrete practices:

• After a wrong output, give the next instruction in actionable form: state the observed gap, the required result, what should change, and ask for a revised version.
• Define the target result, data and constraints before delegating, because “do it as usual” carries little information for a system with no shared history.
• If the same defect recurs, change the instruction, source, template or control procedure rather than only the final document; the corrected process applies to all future outputs.
• Where AI agents are treated as employees, verify who checks outputs and who owns responsibility for errors, especially in HR and finance document review.
• Make the training goal the connected capability of judgement, management and AI-specific constraints, not just prompt techniques for individual tools.