The Redeployment Problem

Fifteen months ago the loudest voices in AI were forecasting mass job elimination. That forecast has quietly been withdrawn.

The Wall Street Journal reported that the share of CEOs expecting significant AI-driven headcount reductions fell from 46% in January 2025 to 20% in May 2026, on EY-Parthenon polling of 1,200 executives across 21 countries. MIT economist David Autor offered two readings in the same piece: leaders “have noticed the labor market is not imploding as fast as they claimed,” or they “may have realized it was bad business to say that your great new product will destroy the economy.”

Both readings can be true and still miss what operators are seeing on the ground. AI is replacing tasks rather than whole jobs, and the companies furthest along are redeploying people instead of eliminating them. Hilary Gosher of Insight Partners, whose team works inside a large portfolio of software companies, has described the same pattern.

Redeployment is the right goal. It is also gated by infrastructure most enterprises have not budgeted for.

Task replacement is not redeployment

In the first wave of enterprise AI deployment, a lot of teams put ChatGPT or Copilot in front of knowledge workers and measured productivity gains. The numbers came in strong. In a controlled study published in Organization Science, researchers working with Boston Consulting Group tracked 758 consultants using GPT-4 on realistic tasks. For tasks within the AI’s capability envelope, GPT-4-augmented consultants completed 12% more tasks, worked over 25% faster, and produced work rated about 32% higher in quality than their unaugmented peers.

That is task replacement: AI does a portion of the work that used to sit on a human’s plate, and the same person does the same job with better throughput.

Redeployment is different. The task comes off the human’s plate entirely, the person moves to work that was not getting done before, and the AI runs the original task unsupervised.

Task replacement produces marginal productivity gains. Redeployment produces durable ROI. Most enterprises are stuck at task replacement, and the reason is trust rather than retraining.

The trust threshold

Every AI deployment eventually reaches the same decision: do we trust this enough to remove the human check?

Below the threshold, humans supervise every output. AI accelerates the work without freeing the person doing it, so the productivity gain is real but bounded. Above the threshold, AI runs unsupervised and the gain compounds, because the human moves to different work entirely. The failure mode also gets more expensive. Any mistake now happens without supervision, and if you cannot reconstruct what the AI did or why, the mistake becomes unrecoverable.

Most enterprises stay below the threshold because the cost of a single unrecoverable AI mistake is high enough that, risk-adjusted, staying put looks safer than crossing. It has little to do with whether their AI is any good.

In February 2024, the British Columbia Civil Resolution Tribunal held Air Canada liable for information its customer-service chatbot gave a bereaved customer. The chatbot invented a bereavement fare policy that did not exist. The airline argued the chatbot was “a separate entity” responsible for its own outputs. The tribunal rejected that and ordered Air Canada to pay damages of a few hundred dollars.

The precedent outlived the damages. Any organization deploying AI in a customer-facing context now knows that “the AI did it” is not a defense, and the same logic applies internally the moment the CFO or the audit committee asks what happened.

Audit is a productivity problem

Audit infrastructure typically gets sold as compliance spend, something you buy because a regulator or an auditor is asking for it, owned by the security or compliance team.

Read against the redeployment problem, it is a labor-productivity investment.

The human stays in the loop mainly to reconstruct what happened when something goes wrong, rather than to catch errors in the moment. A supervised AI is a recoverable AI. Without audit infrastructure an unsupervised AI is not recoverable, and if you cannot answer “what did the AI do and why,” you cannot defend the outcome to the customer, the regulator, or your own leadership.

Audit infrastructure inverts that. A tamper-evident record of what the AI did lets an organization take the human out of the loop and still explain the outcome afterward, which is what makes redeployment safe to attempt.

The NIST AI Risk Management Framework, published in January 2023, splits the core functions into GOVERN, MAP, MEASURE, and MANAGE. MEASURE and MANAGE are where operational risk gets tracked and treated. Both depend on a record of what the AI system did. Without that record, MEASURE and MANAGE are aspirational.

Cost per outcome requires attribution

Two metrics separate durable AI ROI from the appearance of it: cost per outcome, and output per employee. Both require attribution.

Cost per output is easy. You count tokens or API calls and divide by units produced. Cost per outcome is harder. It requires knowing which AI action produced which business result, at what cost, with what side effects. Was the model call that generated the customer response the same one that led to the resolved ticket, or did a human fix the AI’s mistake before the customer noticed? Was the revenue attributable to the campaign the AI wrote, or to the marketer who rewrote 40% of it?

Attribution requires an audit trail with granularity: which model call, which prompt, which output, which downstream action. Without that, cost per outcome is a marketing phrase.

Output per employee has the same dependency. In an AI-augmented workflow, output is a compound of human and AI work, and separating the two requires attribution at the task level. An organization that cannot say what its AI contributed cannot claim the productivity gain.

What we built Gate AI for

We built Gate AI to be the infrastructure layer that makes redeployment possible.

The core mechanic is a drop-in gateway that every AI request routes through. Every request gets prompt-injection defense, output-side scanning for secrets and PII, and a tamper-evident audit trail anchored to Constellation’s Digital Evidence layer. The audit trail is a cryptographic record that a regulator, a customer, or the organization itself can verify without trusting the operator, rather than a log file on someone’s disk.

Compression cuts token costs on every request, typically 20% or more. That matters for cost-per-outcome measurement, because cost is captured at the request level, per model and per team. Attribution comes with the audit trail rather than on top of it.

The security layer is the entry ticket. The audit layer is the productivity story.

Redeployment is an infrastructure problem

The companies furthest ahead on these metrics shipped the trust infrastructure before they needed it. Because the AI’s work is verifiable, they can move people off tasks. Attribution is built in, so cost per outcome is a number they actually have. And when a result gets challenged, the record is already there.

The companies stuck at task replacement are stuck for reasons that have nothing to do with model quality. They cannot cross the trust threshold, so their humans stay in the loop and their productivity gains stay marginal. Their pilots produce nice charts and no organizational change.

The 46-to-20 shift in CEO messaging reflects more than tempered expectations. It is executives discovering that the second half of the AI productivity story runs on infrastructure most of them have not built, which is the infrastructure Gate AI ships.