Managing AI Is Its Own Job

A recent Fast Company piece captured something that a lot of enterprise teams are feeling but haven't quite named: managing AI has become its own job. The article cites a Workday study of 3,200 employees that found over a third of time saved through AI is offset by rework - prompting, checking, correcting, and re-doing output that wasn't quite right. Workday calls it the “AI tax on productivity.”

That framing resonated with us because it describes a pattern we've been watching for a while now. Organizations adopt AI tools expecting efficiency gains. They get some. But they also get a new layer of work that nobody budgeted for: the labor of supervising the AI.

The Supervision Trap

Here's how it plays out in practice. An AI tool generates a draft, an extraction, a summary, or a recommendation. Someone on the team reviews it. They catch an error - a hallucinated citation, a misread number, a confident-sounding answer that's subtly wrong. They fix it. They re-run the prompt. They check again.

Each individual cycle is fast. But across an organization, across dozens of workflows, the cumulative cost is enormous. And the nature of the work is particularly draining - it's not creative, it's not strategic, it's quality control on a system that was supposed to eliminate quality control.

The Fast Company article quotes Rumman Chowdhury, former U.S. Science Envoy for AI, describing the pattern from the employee's perspective: “Yeah, it's producing stuff - and then I have to spend three hours going through every citation and making sure it's not a hallucination.”

Most organizations measure AI ROI by gross time saved - how much faster did the task get done? That metric doesn't account for the rework. When you measure net value - time saved minus time lost to supervision and correction - the picture often looks very different from what was promised.

Where the Tax Comes From

The AI tax isn't random. It has a specific source, and understanding it matters for knowing how to avoid it.

The tax is highest in systems where AI generates output probabilistically - where every run produces a slightly different result, where confidence is high but consistency is low, and where there's no way to verify the output without a human reading through it line by line.

This is the default mode for most generative AI tools. They're designed to produce plausible, natural-sounding output. They're not designed to produce the same output twice. When you use them for creative tasks - brainstorming, first drafts, exploration - that variability is a feature. When you use them for operational tasks - data extraction, document processing, compliance checks - that variability is the tax.

The ServiceNow CDIO quoted in the Fast Company piece caught her AI tool making a basic math error. She flagged it, gave it a thumbs-down, and fed it back into a correction loop. That's a responsible approach. It's also an expensive one, at scale, repeated across thousands of tasks and hundreds of employees.

Where the Tax Disappears

The AI tax is not inevitable. It's a design choice - or more precisely, it's the consequence of applying the wrong kind of AI to the wrong kind of problem.

The IBM Consulting approach from the same article is instructive. They identified over 200 potential AI use cases for a client, measured each against ROI, cut half immediately, and found that the top 10 drove 80% of the total value. The lesson: AI deployed broadly and optimistically creates overhead. AI deployed narrowly and deliberately creates value.

We'd add a second dimension: the tax disappears when the AI's output is deterministic rather than probabilistic. When the system gives you the same answer every time, traces it back to the source, and flags uncertainty explicitly - there's nothing to check. The supervision layer goes away because trust is built into the architecture.

This is the core of what we build at YellowPad. When we extract data from a document - a date, a dollar amount, a key provision - the output is a structured record, not a generated interpretation. Same question, same answer, every time. Auditable back to the exact source text. A system built this way doesn't create an AI tax because it doesn't create output that needs to be supervised.

The distinction matters because the industry conversation tends to treat “AI” as a monolithic category. It's not. A system that generates a different summary every time you ask is a fundamentally different tool from a system that records a fact once and returns it consistently. Both use AI. One creates an ongoing supervision cost. The other eliminates it.

The Question Organizations Should Be Asking

The Fast Company piece frames the AI tax as a training problem - employees need better instruction on how to use AI tools effectively. There's truth in that. But the deeper question is whether the tool should require that level of human management in the first place.

If your AI system needs a skilled operator to prompt it correctly, evaluate its output, and catch its errors - you haven't automated a job. You've created a new one. The operator might be faster than the person who did the original task manually. But you're still paying for a human in the loop, and you're paying for a particularly tedious kind of human work: the work of not quite trusting your tools.

The organizations that will get the most value from AI aren't the ones that train their employees to be better AI supervisors. They're the ones that deploy AI in ways that don't require supervision - systems where the output is reliable enough to act on without a human re-checking every result.

That's a higher bar. It's also the only bar that eliminates the AI tax rather than just reducing it.