Sideways is the New Up: A Third Tier of Talent for the AI Era

In 2025 and 2026, the corporate world ran a vast, unplanned experiment in AI-driven workforce reduction. The results are now in, and they converge from every direction.

Roughly two-thirds of companies that laid off workers after implementing AI are rehiring some of them, according to a Careerminds survey of 600 HR professionals. Forrester finds that 55% of executives regret AI-driven cuts and predicts half of AI-attributed layoffs will be reversed by the end of 2026. Gartner projects that half of the companies citing AI for headcount reduction will restaff similar functions by 2027, often under new titles. Robert Half found 29% of companies rehiring for the exact roles they had cut. About a third of firms spent more on restaffing than the layoffs saved, and the returning "AI-native" roles command salary premiums of 20–35% over the positions they replaced. No single survey carries the claim. Their convergence does. The industry has a name for it now: the AI boomerang.

Klarna remains the emblematic case. Its CEO publicly celebrated an AI assistant doing the work of 700 service agents, then publicly reversed, admitting that cost had become too dominant a factor in the company's thinking and that the result was lower-quality service. Klarna began rehiring the humans it had dismissed. But Klarna is no longer an outlier anecdote. It is the median experience.

Read correctly, the boomerang is not a story about AI failing. The technology did automate the tasks. It is a story about organisations discovering, the expensive way, that the people performing those tasks were carrying something the AI was not: context, judgment, institutional memory, and the practical knowledge of how the work really flows. Rather than a failure story, this is the most honest data point in the AI labour debate, and the starting point for a more useful conversation.

The Bench Problem

Nvidia CEO Jensen Huang has argued that while some jobs will be eliminated and many created, every single job will be impacted: "You won't lose your job to AI, you'll lose it to someone who uses AI."

The debate between AI accelerationists and decelerationists is, at its core, an argument about time. We already know that jobs will be disrupted and that workers will be benched. The extreme optimists believe the interregnum will pass faster than you can say "Jevons paradox." The doomers fear permanent dystopia. Both are, in essence, debating duration. Neither is asking the question confronted daily by those at the coal face: how do you manage the bench? How long do capable people sit idle, at what cost, and to whose ultimate benefit, yours or a competitor's?

The Other Side of the Ledger

While organisations were shedding practitioners, an acute shortage emerged of a very particular profile: the person who can sit inside a real organisation, with its legacy systems, data quirks, and human resistance, and make AI work in production. The market calls this profile the Forward Deployed Engineer. Job postings for it grew by more than 800% in nine months against a candidate pool that grew by roughly half. Senior practitioners clear $300,000 to $600,000 in total compensation. OpenAI has assembled a $4 billion deployment company backed by TPG and McKinsey; Anthropic has a $1.5 billion enterprise services venture with Blackstone. The driver is the deployment gap itself: an MIT study found that some 95% of enterprise AI pilots fail to reach production. That is a failure of implementation and adoption, not of model capability.

Now hold the two facts side by side. The market will pay half a million dollars for an outsider with implementation skill and no institutional context, while organisations lay off insiders with deep institutional context who have demonstrated implementation skill by automating their own work, then rehire strangers at a premium when the gap becomes undeniable. This is not a labour market functioning. It is a labour market failing to see an asset class.

The Binary That's Breaking

Enterprise design has historically relied on two talent archetypes during technology transformations. The internal line worker brings deep institutional knowledge, anchored to a single function, but is highly vulnerable to automation. The external strategic adviser offers technical proficiency with little local context, at significant cost, and burns four to six weeks of billable time simply learning how an organisation actually operates.

Between these two archetypes sits a critical gap, accentuated by the scale of AI disruption. What the era demands is a third category: high technical proficiency, native institutional context, and permanent horizontal leverage.

The Third Way Most Companies Already Tried

Many organisations have already moved toward this gap with champion networks: peer advocates who promote AI tools alongside their day jobs. The instinct is exactly right, and the results prove the mechanism. One global bank built a network of 4,000 peer accelerators and achieved over 70% adoption of approved tools across 182,000 employees.

But champion programmes hit a structural ceiling. Champions are advocates, not implementers; they can demonstrate a prompt but cannot rebuild a workflow. Their duties are capped at a few hours a week, and the day job always wins. They hold no mandate beyond their own team, no budget line, no authority to carry a solved problem across a boundary. And they are compensated in badges and visibility, in a market paying a measured 56% external wage premium for AI skills, which means champion programmes systematically train their best people for the competition.

None of this makes champion networks a failure. It makes them Phase 0. Organisations that have invested in champions have proven the peer mechanism works and built the natural feeder pool for what comes next. Champion networks are the demo. What follows is the production system.

Enter the Sideways Deployed Specialist

The Sideways Deployed Specialist (SDS) is what you get if a Forward Deployed Engineer and a traditional super user had a child. The FDE brings implementation orientation and AI fluency. The super user brings deep institutional context, domain knowledge, and peer-level trust. The SDS inherits both, and corrects each parent's defect: unlike the FDE, they are not transient, and unlike the super user, they have freed capacity, real training, an explicit cross-boundary mandate, and money attached.

These are domain practitioners in finance, marketing, HR, procurement, operations, who have already automated a substantial portion of their own workflows using AI agents. In doing so, they have acquired something rare: firsthand, practitioner-level knowledge of what it actually takes to make AI work inside a real organisation. Once upskilled in advisory and change-management competencies, they deploy horizontally into sister functions and sister entities as peer-level implementation resources, governed by a Centre of Excellence. The conversion pipeline is supported by AI agents purpose-built for the role: coaching deployments in real time, translating domain knowledge into structured playbooks, and running practice simulations before an SDS faces their first live engagement.

Deployment comes in four forms along a maturity dial, from embedded support (office hours and co-working from the home role) through dual-track and fixed-duration sprints to fully untethered redeployment. At scale, the logic points to an internal marketplace where freed practitioners list capacity and receiving units bid for it, and ultimately to an external exchange where verified SDSs become accessible beyond the immediate portfolio, converting surplus institutional capacity into a tradeable asset.

One design detail matters more than it first appears: practitioners must be safe to be found. A rational employee who has automated 70% of their role has every incentive to hide it, because disclosure looks like evidence for redundancy. The model therefore requires a disclosure safe-harbour as standing policy: automations disclosed through the programme cannot be used against you, and disclosure is the gateway to the SDS track and its financial upside. Without that guarantee, the candidates exist but stay invisible.

Why Peer Delivery Changes Everything

The bottleneck in AI adoption is not technology or capital. It is adoption itself.

Enterprise experience is consistent in direction: traditional classroom training and technical documentation produce durable adoption rates below 20%. Side-by-side building, where an AI-proficient practitioner sits beside a colleague and rebuilds a real workflow together, lifts durable adoption dramatically. Peers carry credibility that consultants cannot rent and trainers cannot teach. They speak the language of the function, know its constraints, and have already lived the doubts the learner is about to voice.

An Honest Caveat

The SDS model is not a universal solvent. It works where peer credibility, institutional context, and freed practitioner capacity actually exist. In organisations with thin AI fluency, no champion layer, or hostile labour relations, the prerequisites have to be built first. The honest sequence is champions, then SDS, then marketplace — not a leap straight to the end-state.

Where It Runs First

The model lights up earliest in large, multi-entity organisations: banks, insurers, telcos, holding groups, and the public sector. These environments have the scale to spawn a meaningful pool of AI-proficient practitioners, the internal mobility infrastructure to deploy them, and the cross-entity surface area where horizontal leverage compounds. They also have the most to lose from the AI boomerang and the most to gain from converting freed capacity into deployable expertise.

Policy Buys Time

Policy will not stop AI-driven workforce dislocation, but it can change its tempo. Disclosure safe-harbours, retraining levies, and internal redeployment mandates buy organisations and workers the runway needed for sideways moves to outrun downward ones. The organisations that act in that window — that treat the AI boomerang as a signal rather than an embarrassment, and that build the SDS layer before the next round of cuts — will own the adoption advantage of the decade. The rest will keep paying premiums for strangers to rediscover what they already knew.