The problem with AI-generated work
One recurring criticism of artificial intelligence in the workplace is that it creates more work for managers. An employee produces a report, proposal or analysis with AI; the result looks polished, but essential research is missing, conclusions are questionable or the work does not meet the required standard. Someone else must repair it.
Perhaps we are looking at the problem from the wrong direction. Rather than concentrating on correcting inadequate output afterwards, why not improve how the work is specified, undertaken and verified?
Start with a specification
Before beginning a project, establish what it must involve. Depending on the task, this could include:
- Purpose and intended outcome
- Necessary research and supporting evidence
- Brainstorming and exploration of alternative ideas
- Examination of similar products, projects or solutions
- Competitor research, where relevant
- Agreed working methods and deliverables
- Verification methods and acceptance criteria
Not every project requires every activity. The essential step is deciding what is mandatory, what is desirable and what can be omitted.
Generic specifications and desirable activities
Organisations can maintain a library of reusable specifications for articles, research, investigations, product development, business proposals and other recurring tasks. The originator selects the relevant template and adjusts it for the assignment.
A second list offers desirable activities: consultation, competitor analysis, alternative approaches, case studies, additional fact-checking or cost-benefit analysis. These are selected when they add value, rather than automatically imposed on every task.
Return to the specification at the end
The critical step is to use the same approved specification twice: first to plan the project, then to assess its result.
For every significant requirement, record (1) what must be done, (2) what evidence demonstrates completion, and (3) what standard will be accepted. For example, “research competitors” could mean “compare three relevant competitors against agreed criteria, identify sources and explain the implications”.
If requirements change during the work, record what changed and why. The final check should use the latest approved version, not an obsolete brief.
Independent verification and human acceptance
An independent checking AI may be given the specification, finished work and supporting evidence. It can report apparent omissions, inconsistent reasoning and requirements that lack evidence. The originator corrects what is justified and investigates contested findings.
Separate AI checking is helpful but does not guarantee correctness. External evidence and accountable human acceptance are still needed, particularly where decisions affect people, money, safety or legal obligations.
A practical method
Specify → Select → Execute → Verify → Review.
Specify the mandatory requirements; select useful desirable activities; execute the work; verify its accuracy and completeness; then review the result against the approved criteria. The originator should do this before the work reaches the manager.
Reducing rework rather than concealing it
The method is testable. An organisation could compare how much time managers spend correcting work, the number of returned submissions, the quality of accepted outputs and the time spent by originators checking them. A genuine productivity improvement must include the cost of verification.
The objective is not merely to make AI produce work faster. It is to increase the likelihood that satisfactory work is produced correctly the first time.
Related resource
For a reusable specification and final-review checklist, see the companion resource “AI-Assisted Project Specification and Assurance Method”.
For further information: ai-humancapability.com