how should work be redesigned around AI?

Randstad Advisory Signals from the Edge: this is a dispatch from your colleagues in 2030

Is your AI strategy stuck in pilot purgatory? In Signals from the Edge, Randstad Advisory offers six core principles to drive durable results. Find out why the best global companies ask "Should AI do this?" and how they protect indispensable human elements — like judgment, relationships and skill development — even when AI makes them optional. Get practical insights on shifting from tactical automation to strategic, positive-sum augmentation that maximizes both productivity and human value. 

Most organizations are still asking: What can AI do? Which tasks can we hand over? Where do we get the efficiency gains? These are not bad questions, but they are the wrong starting point.

They begin with the technology and work backward to the work. 

The result? AI layered on top of outdated processes. Task-level productivity gains with no measurable impact. And a workforce that is simultaneously busier using AI, but less clear about how it improves outcomes. 

The organizations pulling ahead are starting from the other end.

They begin with the work itself. Designing the human-AI partnership that combines human judgment with AI capabilities to produce outcomes neither could achieve alone. This requires completely rethinking how work gets done.

They are treating AI not as something to be deployed into work but as a catalyst for redesigning work entirely. They have access to the same tools. The difference is architectural. This is where the path divides.

And 2030 will look very different depending on which path you took.

from implications to architecture: the new design discipline

Last month, we outlined the implications of AI that nobody fully anticipated, and what you can do next. Those implications are real: the second job nobody planned for; the cognitive atrophy nobody measured, the personality fragmentation nobody designed around, the collective narrowing of thinking that individual performance metrics would never reveal. 

The question is no longer how to introduce AI, but how to redesign workflows, roles and decision-making around it. 

There is no single answer, because that requires a structured conversation about your organization specifically. But you can read on to understand what the organizations’ getting this right are actually doing, and the questions that separate deliberate choices from default ones.

could vs. should

Most organizations approach AI deployment with a capability question: What can AI do? 

The organizations getting the best results asked something different.: AI could do this, but should it? This leads to more precise questions, and answers. 

What does this task produce beyond its output? Does it build expertise, maintain relationships, develop judgment? If we hand it to AI, what happens to those things? What does the human in this role need to remain capable of, and does this decision protect that capability or erode it?

There are also two dimensions most organizations are not yet treating as diagnostic criteria at all.

The first is relational. How many people does this task involve, and does its success depend on them feeling safe, heard or understood? And beyond that, what is the nature of the interaction between them? In many tasks, the exchange between people is not incidental. It is how decisions get shaped, challenged and refined. Remove that interaction, and you do not just lose human warmth. You change the quality of decisions and outcomes.

The second is a consequence. If something goes wrong, how severe is the damage, and how easily can it be undone? 

The answers to these questions should shape how much human oversight is built into the work from the start.

Harvard Business Review research published in April 2026 puts it starkly: Automation produces early bottom-line gains, but augmentation produces superior top-line growth and better outcomes over time. 

The ones asking “Should AI do this?” are building something more durable. But should is only the first question. Once you have answered it, more questions immediately follow. How does the work get done now, and what does that change about where people spend their time and how value gets measured? And they are where most organizations stop asking.

what separates the companies that are getting it right?

The MIT Industrial Performance Center spent three years inside more than 50 organizations across healthcare, finance, retail and manufacturing to see what is actually happening when AI meets real work in real organizations.

Their finding is not that AI is transforming work; it is that most organizations are still trying to figure out what that transformation should look like. The majority are still proving the concept, trapped in what is being called pilot purgatory — the gap between a tool that works in a controlled test environment and how an organization has actually changed around it. 

Deploying a tool and redesigning a workflow are not the same thing. Most organizations are discovering that the distance between them is where transformation stalls, and where the costs of inaction quietly accumulate. 

IBM's 2025 CEO study of 2,000 executives reveals the consequence: Only 25% of AI initiatives have delivered expected ROI, and only 16% have scaled enterprise-wide. The approaches that organizations are taking to get out of pilot purgatory vary widely. 

Deloitte's State of AI in the Enterprise 2026 survey of 3,235 leaders finds that roughly one-third are deeply transforming,reinventing core processes or creating new products around AI. Another third are redesigning key processes. The remaining third are using AI at surface level, with little or no change in how work is actually done. What separates those thirds is rarely the technology. That difference shows up most clearly in how organizations structure the change itself.

Neither a purely top-down nor a bottom-up approach has proven sufficient on its own. PwC finds that ground-up crowdsourcing delivers impressive adoption rates but rarely produces meaningful business outcomes. PA Consulting finds that purely top-down approaches fail at execution; leaders define outcomes, but are often removed from the workflows AI is meant to improve. 

As MIT Sloan's Paul McDonagh-Smith observes, GenAI is simultaneously a top-down initiative and a bottom-up movement. The evidence points toward a hybrid: leadership sets strategic priorities and governance, and teams hold the mandate to redesign the workflows. 

And regardless of which approach organizations take, new INSEAD and Harvard field experiment research makes the underlying challenge precise. 

Profiting from AI is not mainly a prompting or access problem; it is a workflow design problem. Firms must discover where AI can reorganize the production system. That is the work most organizations have not yet done.

The organizations that succeeded in redesigning work share the following: 

• • Commitment to solving a well-defined problem
  • Technologies to address the specific problem
  • Domain experts that are close to the process at every stage
  • Leadership that stays actively involved as sponsors and owners of the outcome

3 problems, very different risks

MIT finds that organizations are deploying AI to address one of three recurring challenges. Understanding which one you are solving changes everything about how you design the work around it.

1. The bottleneck problem
Workers are buried in near-routine tasks. This is where AI is most defensible. At the Permanente Medical Group, ambient scribes across 2.58 million patient encounters saved approximately 15,800 hours of documentation time. Mass General Brigham saw a 41% reduction in after-hours EHR time and a 66% drop in incomplete notes. Physician burnout decreased. Patient eye contact increased. The AI handled the draft. The human handled the validation and the final judgment. That distinction is the architecture that scaled.

2. The cafeteria problem
Work requiring input from multiple experts is integrated into a coherent output. AI can synthesize what those experts would likely have said. It saves time. But the cafeteria is also where collaboration happens, where relationships are formed and where shared understanding is built. When AI replaces it, the task gets done, and the team quietly becomes a collection of individuals who no longer need each other in the same way. Morgan Stanley designed around this deliberately. Their AI assistant gives 16,000 advisors access to 100,000 internal documents. Adoption reached 98%, but the client relationship stayed entirely human. Should the AI have the client meeting? The answer is no.

3. The learning curve problem
This is the most seductive pattern and the most dangerous. AI helps novices perform beyond their experience level. Brynjolfsson, Li and Raymond's landmark study finds a 34% productivity lift for novices in the deployment of 5,000 customer service agents. But MIT adds the question that the productivity data doesn't answer: Is the novice developing expertise or simulating it?

When the answer is simulation, the problem has not been solved; it has been deferred. The evidence is clear. A Harvard and BCG study finds that while AI improves consultants’ quality 40% and increases their speed by 25% overall, it creates a negative productivity gap when those consultants work in other domains. Consultants using AI on tasks outside their designated capabilities perform 19% worse than those without it — they trust the output without being able to judge its quality. Ropes & Gray took that risk seriously by letting first-year associates count up to 400 hours toward AI training rather than client work, building judgment alongside the tools, not behind them. It is consistent with what Matt Beane finds across more than 30 professions in his research for The Skill Code: When intelligent machines remove novice involvement from the workflow, skills stop developing even as the expert becomes more productive.

Each problem carries a different risk. Get the bottleneck right and you free people. Get the cafeteria wrong, and you quietly hollow out the team. Get the learning curve wrong, and you defer a problem that compounds silently.

6 principles for producing results

The MIT working group distilled three years of observation into six principles that distinguish the organizations producing durable results. 

1. 1. Minimize manual work, but not the work people value.
      Deloitte found that 56% of organizations design AI solely for business outcomes, while only 40% design it for both business and human outcomes. The compounding cost of that gap is what Deloitte calls “culture debt.” The most defensible deployments target tasks both workers and organizations agree are burdensome. The riskiest ones target tasks that build relationships or develop expertise. 
      
      
   2. Design for learning, not just performance.
      The risk is that workers complete tasks without retaining them, developing apparent competence while the underlying capability atrophies. For high-stakes tasks, this means deliberately scheduling work that keeps the human skill alive, not just the output flowing.
      
      
   3. Protect teamwork even when AI makes it optional.
      What AI cannot replace is what happens when people work through a problem together. Several organizations are now explicitly preserving collaborative processes they could have automated because the collaboration is doing something the AI can not.
      
      
   4. Control the interface.
      Most organizations now buy AI rather than build it. But they still control how the tool is presented to their people; what it shows, what it hides, how it signals uncertainty. Interface design shapes supervisory behavior. The organizations that have thought carefully about this are getting different results from the same underlying models.
      
      
   5. Respect domain expertise as irreplaceable.
      Organizations reducing expert headcount in response to AI-enabled novice performance are simultaneously degrading the quality assurance on which their outputs depend. The right question is not how to do something faster; it is how to do it better.
      
      
   6. Build accountability into the workflow.
      Effective design begins with understanding how AI works and what it is architecturally driven to do. AI drives toward consensus by default. The organizations getting this right treat this drive for consensus as a risk to be managed. They build deliberate challenges into the process: first, defining how truth is verified in a given context; second, testing the AI’s output against varied human perspectives and domain expertise; and only then moving toward a final consensus. In this model, audit mechanisms are not bureaucratic; they are the professional accountability that the technology’s confidence would otherwise displace.
      
      

humans in the loop: a design discipline 

The phrase “human in the loop” has been reduced in most corporate conversations to a governance requirement. A risk mitigation. A box to tick before deployment. But the research suggests it is something far more expansive than that.

Keeping humans genuinely in the loop is not just about catching errors. It is about designing work so that humans continue to develop and retain their capabilities. This is critical to ensure that the judgment, relationships and contextual understanding that only come from doing the work are actively maintained rather than quietly eroded. It’s also important because people find more joy and meaning in what they do when AI is handling what burdens them, not what defines them. 

When AI takes on more of the execution, humans move into the role of supervisor. That transition has happened before, in aviation, nuclear operations and manufacturing automation. Supervisory control is a skill. It requires training, situational awareness and the maintained ability to take over when automated systems fail. Crucially, it is also about wisdom, the kind of judgment that stems from lived and learned experiences that AI simply does not have. These human insights are vital because systems tend to fail in the highest-risk moments, exactly when that supervisory wisdom is most needed.

The consequences of getting this wrong are already visible. A database maintained by Damien Charlotin at HEC Paris shows more than 486 AI hallucination filings worldwide as of late October 2025. A Stanford study finds AI legal tools hallucinate in at least one of six queries — and, in some tools, as often as one in three times, even when those tools are purpose-built for legal research. And when Air Canada's chatbot hallucinated a bereavement fare policy, the court held the airline fully responsible; the AI's error was the organization’s liability. 

But the design question goes beyond liability. It is what humans need, in terms of training, maintained capability, genuine accountability and meaningful work to perform the role rather than just occupy it. Most organizations have answered the compliance question. They have not yet answered the design one.

The cases that best illustrate positive-sum augmentation, where AI and human capability genuinely compound each other, produce a new capability that is greater than the sum of its parts. This structure is worth naming explicitly.

• • • A Harvard and NBER field experiment at Procter & Gamble with 776 professionals found that individuals working with AI matched the output quality of two-person teams working without it, and that AI broke down functional silos, with R&D and commercial professionals producing more balanced solutions together than before.
    • JPMorgan’s LLM Suite, available to approximately 250,000 employees, was designed around this architecture. The firm projects $2 billion in AI-related upside. The design kept the humans in the decisions. It is refreshed every eight weeks. And it is positioned not as an automation tool but as an intelligence amplifier through: 
      • • Research synthesis
        • Document drafting
        • Data querying handled at speed
        • Freeing advisors and analysts to do work that specifically requires them
    • Moderna gives every employee access to a suite of custom GPTs, more than 750 built internally within two months of launching ChatGPT Enterprise. The legal team achieved 100% adoption. But adoption in this case did not mean replacement. It meant that lawyers were spending less time on drafting and more time on judgment, strategy and client relationship, the work that required them. 
    • Siemens Industrial Copilot generates control-system code that requires around 20% adaptation by engineers. It produces panel visualizations in roughly 30 seconds that would previously have taken hours. The engineers’ attention shifted from mechanical code generation to the creative, judgmental work of adaptation, validation and improvement. 

We have our own internal example that illustrates the positive-sum argument. When using AI to enrich our Notebook outputs, if we let it go too far, the output becomes smooth and generic — vanilla. It loses the specific human observations, the nuanced details, the edges that make our analysis distinctive and valuable. Our team actively pushes back against that because clients can produce vanilla themselves if they have the capacity. What they come to Randstad Advisory for is the judgment and texture that only comes from a human working alongside the tool, not behind it. 

In one recent work redesign engagement, when we mapped which tasks to augment and which to protect, the answer came from the people doing the work. They wanted AI to handle the time-consuming research — bid reviews, executive profiling, sector-specific drafting, proposal summaries — so they could spend more of their time on what actually moves client relationships forward. Automation was welcome; the relationship was non-negotiable. 

That shift is the positive-sum outcome. Not fewer but different and better-used people.

which questions should you start with?

The research, the case studies and years of organizational experimentation converge on a set of questions that any leader designing work around AI should be asking. Not once, at the point of deployment, but consistently, as a discipline.

What follows is the first layer, the questions that structure everything else. The complete framework requires a conversation about your specific work, your specific people and the specific decisions you are navigating. But these are the ones worth starting with.

Before you deploy, the questions are about design.
What problem are you actually solving, and does your answer account for what the task produces beyond its output? The expertise it builds, the relationships it maintains, the judgment it develops? Whether AI could do this is rarely the hard question. Whether it should, and whether now is the right time, almost always is.

When you deploy, the questions are about control.
Not an oversight, but genuine supervisory capability. The ability to catch failure, validate output and take over when something goes wrong. Verification must be real, not theatre. How is the interface designed? What is the human actually being asked to verify? What is the audit mechanism? Most organizations have not asked these questions at the level of specificity they require.

As it scales, the questions are about what you cannot yet see.
Whether you are building one AI-enabled workforce or many, because personality, thinking style and working preference are shaping AI adoption in ways that are almost certainly not uniform across your people. What is happening to your entry-level pipeline. And what signals your workforce is generating that have not yet reached you. If you are not hearing the uncomfortable ones, it does not mean they are not there.

At scale, two quieter risks compound: institutional amnesia, where rapid automation erodes the organization’s ability to understand and fix its own systems, and agency decay, where decisions are made but nobody can explain why, and organizational knowledge hollows out invisibly. Neither will appear in any performance dashboard. Both will be expensive to reverse.

The organizations that will define what good looks like are the ones treating AI deployment as a design discipline rather than a procurement decision. McKinsey’s The State of AI Report is clear: Workflow redesign, not the technology stack, is the largest driver of value. Only 20% of organizations have done it. 

The ones that automate without redesigning are accumulating structural debt. In capability, talent pipeline, institutional knowledge. This will not be visible until it is expensive to reverse.

the big lessons

• • The organizations pulling ahead are not asking what AI can do. They are asking what work should become and making deliberate choices about how to build it.
  • Transformation stalls in the gap between deploying a tool and redesigning the work around it. Most organizations are still in that gap.
  • Not all AI work design problems are the same. The problem you are solving determines the design, the risk and what failure looks like. Misidentify it, and the consequences compound quietly.
  • The organizations producing durable results protect what makes work human: the expertise that builds over time, the relationships that hold teams together, the collaboration that generates thinking no individual could produce alone.
  • Human involvement is not a safety net. It is a capability that requires investment, practice and deliberate design. When it atrophies, it fails at exactly the moment it matters most.
  • The real goal is not automation. It is work where humans are doing more of what requires them and less of what does not.
  • The risks that will matter most in three years are not visible today: talent pipelines thinning, institutional knowledge hollowing out and decisions being made that nobody can explain. None of these appear in any dashboard until they are already expensive.
  • The largest driver of AI value is redesigning how work gets done. It is also the thing most organizations have not yet done.
    
    

from asking what AI can do to designing what work should become

When work is designed well, it isn't just more productive; it is more meaningful. That is what we are building toward with our clients. If you want to work through what this means for your organization, we can help you decide whether work redesign is necessary, where to start, what to protect and what to watch for that you cannot yet see.

Let's build the architecture together.