What Forrester Actually Said
The headline figure — 50% of AI-related layoffs reversed by 2027 — is striking enough. What makes it sharper is the accompanying data point: 55% of employers who cut staff in favour of AI already regret it. That's not a forecast. That's current sentiment among the companies that made the bet.
What went wrong? They fired their institutional knowledge and kept the tools. The AI could produce outputs, but nobody left on staff had the domain expertise to evaluate whether those outputs were right. They eliminated the people who knew what "correct" looked like in the context of their industry, their clients, their compliance requirements.
Goldman Sachs, in its April 2026 update, provided the other side of the picture: workers displaced by AI face long job searches, and when they land, they typically find lower pay and worse conditions. That's the average. It is not a description of every outcome. The distribution is bifurcating. The people returning with skills the company now urgently needs are not negotiating from weakness.
Why Companies Are Struggling
The logic that drove the cuts seemed sound at the time. AI tools were producing outputs that looked like the work humans had been doing. The cost saving was immediate and visible. The downside — loss of judgment capacity, institutional memory, quality control — was deferred and invisible until it wasn't.
As Forrester's analysts put it, the dynamic resembles an airline cutting the pilot to reduce weight, then discovering someone needs to fly the plane. The AI can generate the proposal, draft the contract, write the marketing copy. But it will also confidently include the wrong regulatory reference, subtly misread the client brief, or produce brand messaging that sounds generically professional but is precisely wrong for the audience. And if you've eliminated the people who would have caught those errors, the errors ship.
The companies making the most effective use of AI in 2027 will be the ones that either never lost their experienced people or moved fastest to get them back. That creates a specific type of demand — for people who combine domain expertise with AI fluency. It's a combination that took time to develop and can't be manufactured quickly.
The Three Skill Categories That Matter
Category 1: AI Workflow Design
The ability to look at a business process and redesign it around AI capabilities. This is not software engineering. It's applied judgment about where AI excels — fast, consistent, tireless, good at pattern recognition — and where humans are essential — novel situations, relationship management, ethical judgment, complex tradeoffs. People with operations, project management, or consulting backgrounds can learn this faster than technologists who have never worked in business processes. The skill is designing the handoffs, not building the technology.
Category 2: Prompt Engineering and Output Evaluation
Most people still do this badly. Advanced prompt engineering means constructing inputs that produce consistent, structured, reliable outputs — and knowing how to evaluate whether those outputs are actually good. This includes recognising hallucination (confident-sounding errors), sycophancy (the model agreeing with whatever you imply rather than telling you what's true), and overconfidence on topics where uncertainty is appropriate. These failure modes are not obvious to untrained users, which is precisely why evaluating AI output is a skill worth building. Organisations are actively hiring for it.
Category 3: AI Auditing and Quality Control
As companies scale AI-generated content and automate AI-assisted decisions, they need people who can check the work systematically. Is this output factually accurate? Is it legally safe for the context it will appear in? Does it represent the brand correctly? Does it reflect the current regulatory environment? The MIT study that found 95% of AI pilots failing identified output quality as a primary failure point — not because the AI was bad, but because nobody was checking it properly. AI quality assurance is an emerging role that will attract meaningful compensation precisely because the supply of people who can do it rigorously is currently small.
The Human-AI Collaboration Premium
There's a pattern in how compensation is evolving for AI-adjacent roles that's worth naming directly. The people commanding the highest premiums in the current market are not those who know the most about AI in isolation, and they're not those who have resisted it. They're the people who can work alongside AI in a way that produces outcomes neither could achieve alone.
That means knowing when to trust the model's output and when to override it. It means being able to use AI to move faster while applying human judgment at the moments that matter. It means having enough domain expertise to recognise when something is subtly wrong and enough AI fluency to ask a better question.
This isn't a description of a rare talent. It's a description of what happens when an experienced professional takes AI tools seriously as instruments of their craft rather than either a threat or a magic solution. The barrier is mostly time and intentionality, not innate ability.
How to Start This Week
The gap between people who are building these skills and those who aren't is currently small enough to close quickly — but that window won't stay open indefinitely. The demand spike that Goldman's 2027 timeline implies will make these skills competitive. The time to acquire them is before they're in demand, not during.
- Spend 30 minutes a day for two weeks working on your most complex professional task with AI assistance. Document every point where it helped and every point where it failed. That record is the beginning of your expertise.
- Take one structured course on prompt engineering. OpenAI's promptingguide.ai is a solid free starting point. The goal is to move from intuitive use to systematic use — understanding why some approaches reliably produce better outputs than others.
- Identify the AI tool most relevant to your industry and become the most informed person on your team about it. Not the most enthusiastic — the most informed. There's a difference, and the latter is what organisations actually need.
- Document what you learn. Write it up as internal documentation, post it on LinkedIn, share it in professional communities. Public learning creates its own opportunities in ways that private learning does not.
The goal isn't to become an AI expert. It's to be the person in the room who knows how to use the tool correctly when everyone else is guessing.
That framing matters. The companies rehiring in 2027 aren't looking for AI researchers. They're looking for their previous domain experts, now equipped with the tools that were supposed to replace them. Being that person — someone with real expertise who also understands how to deploy AI within it — is achievable, it's valuable, and the preparation window is open right now.