Every major study of AI and work asks some version of one question: will AI automate this task, or augment the person doing it?

The World Economic Forum’s Future of Jobs Report 2025 asks it. Anthropic’s Economic Index, which measures how AI is actually being used, asks it. On stage in Davos this January, technology CEOs answered it the same way: AI augments people, it doesn’t replace them.

None of them asks the question that comes first: should this work exist at all?

On Thursday, I argued that elimination is the missing objective in how we train and deploy AI. This piece looks at the same gap from the other side: how we measure AI’s impact on work.

What the declining-jobs list really shows

WEF’s 2025 list of the fastest-declining roles reads like this: postal service clerks, bank tellers, data entry clerks, cashiers, accounting and payroll clerks, stock-keeping clerks. The report credits automation, digital payments and self-service.

Look at that list as a process architect instead of a labor economist, and it isn’t a list of jobs being automated. It’s a list of workflows that exist because of how information used to move.

  • Data entry exists because data was captured in one place and needed in another.
  • Reconciliation exists because two systems disagree.
  • Teller transactions exist because the customer had to come to the money.

Automate those workflows and you get a faster version of an unnecessary process. Augment them and you get a smarter person running an unnecessary process. Neither changes what the business delivers to its customer.

The bank that never built the workflow

Digital banks didn’t automate tellers. They designed a bank without branches, so the teller workflow never existed. Onboarding, identity checks, credit decisions and servicing were built as one digital flow instead of a chain of handoffs.

That shows up where it matters: for the customer. As of October 1, 2026, Bankrate puts the national average savings yield at 0.65% APY, while the best high-yield accounts, mostly at online banks, pay around 4%. Lower overhead is the reason usually given.

To be fair, not all of that gap is efficiency. Some of it is customer-acquisition spending, and some of it is that incumbents can pay less because their deposits are sticky. But efficiency is what makes the higher rate sustainable rather than promotional.

That is elimination expressed as a customer outcome: a structural advantage passed straight to the customer. That is what peak business performance should look like.

Why the research can’t see elimination

Task-level frameworks inherit today’s org chart. They start from existing occupations and their task lists, then score each task. A task that should disappear doesn’t get a score of its own. It just gets labeled “automatable.”

Usage data has the same blind spot. Anthropic’s January 2026 index found augmentation narrowly ahead of automation on Claude.ai, roughly 52% to 45%. On the business API side, the picture flips: about three-quarters of interactions were automation, which Anthropic reads as businesses automating routine back-office workflows like email, document processing and scheduling. Nobody is measuring whether those workflows should exist. Both categories describe AI doing existing work or helping someone do it. There is no category for “redesigned so this step is gone.”

That gap isn’t academic. If the measurement has no column for elimination, the transformation roadmap won’t have one either.

How incumbents get there: historical plus market intelligence

A digital bank eliminated work by design, on a greenfield. An incumbent has to do it inside a running operation. That takes two kinds of intelligence.

Historical intelligence tells you what can go. Years of tickets, cases, transactions and logs show which workflows are failure demand, rework, or workarounds for gaps between systems. Each one gets a disposition: eliminate it, hand it to agentic AI, or keep a human in the loop with assistance. The order matters. Eliminate first, automate what can’t be eliminated, and assist humans on the rest. On the AI-native shared services platform I led R&D for, that order is what produced 60–80% work elimination.

Market intelligence tells you what has to go. When the business asks for “faster and cheaper,” it usually means a competitor is already there. Benchmarks like onboarding time, claims speed and order-to-delivery tell you which gap is real and how large it is. A savings rate six times the national average is a market-intelligence signal. It tells an incumbent bank exactly how much structural cost a competitor has already removed.

IT shouldn’t wait for the business to raise that gap. With both kinds of intelligence in hand, it can surface the gap first.

The people question

WEF’s real concern is the workforce, and it’s a fair one. Elimination is not a headcount plan.

Eliminate the work, redeploy the people. Move them to the work the eliminated process was crowding out: customer relationships, exception judgment, new products.

That is why elimination has to be led by change management, not by tooling. It means changing the operating model, then transforming and transitioning. It’s also why enterprises need an elimination office: someone has to own the question before anyone buys the automation.

Three questions before “automate or augment”

  • 1. Why does this workflow exist, and what failure or gap created it?
  • 2. Who already runs without it: a competitor, a digital native, another business unit?
  • 3. If we removed it, what would the customer get from the difference?

Answer those first. Then decide what’s left to automate.

Sources

This article is part of my ongoing series on elimination-led enterprise intelligence: building intelligence that drives peak business performance by eliminating work first. Earlier: the missing objective and why peak business performance starts with what you eliminate.