For most of the history of automation, the guiding question has been how to do the same work with fewer people. AI invites a different question: how can the same people do more? The first question treats people as a cost to be removed. The second treats them as the source of judgment, relationships and expertise that AI cannot replace, and it leads to very different decisions about what to automate and how.
Yet most AI programs are still measured almost entirely on efficiency. Very few organizations track the impact of AI on their people as a KPI, even though research on digital transformation consistently finds that people, skills and process account for far more of the outcome than the technology itself.
The problem with “human in the loop”
Keeping a human in the loop is usually presented as the safeguard that makes automation responsible. Research going back more than four decades suggests it is a much weaker safeguard than it sounds.
In 1983, the psychologist Lisanne Bainbridge published a paper called Ironies of Automation. Her central observation was that organizations tend to automate the easy, routine parts of a job and leave people responsible for the complex exceptions the system cannot handle. Those people then spend most of their time monitoring systems instead of doing the tasks manually, and the skills they need for hard cases fade from lack of practice. When the system eventually fails, the person expected to step in is often the least prepared to do it.
Anyone who has followed GPS directions for years and then tried to navigate a city without it will recognize the effect. The same pattern is now appearing in knowledge work. People who once wrote reports, analyses or customer responses from scratch increasingly review AI drafts instead, and over time the review can drift into approval without real scrutiny.
What the research says about AI and critical thinking
A 2025 study from Microsoft Research and Carnegie Mellon University, The Impact of Generative AI on Critical Thinking, surveyed 319 knowledge workers who use generative AI at least weekly and analyzed 936 real examples of how they used it at work.
The researchers found that workers with higher confidence in AI applied less critical thinking to its output, while workers with higher confidence in their own skills applied more. The nature of the work also shifted. Instead of gathering information and solving problems directly, people spent more of their effort verifying AI output and steering it toward a usable result. Workers found it hardest to evaluate AI in areas where they lacked expertise, which is exactly where unchecked errors are most likely to slip through.
The implication for any business deploying AI at scale is clear. Oversight only works if the people doing it have the skills and confidence to challenge what the system produces, and those skills need to be deliberately maintained.
Start with the people who do the work
Many AI programs are designed by managers and executives looking at processes from the top down. The most useful insight often comes from the people actually doing the job.
When one large telecommunications company brought its frontline customer service staff into an AI planning workshop, the team discovered that agents were spending around 40% of their time chasing internal departments for order and planning information, and another 10% updating a CRM for reports few people read. None of that time was spent helping customers, and none of it had shown up in the management view of where AI should be applied.
That kind of discovery only happens when the people affected by automation are part of designing it.
Separate vicious friction from virtuous friction
A practical way to decide what to automate is to sort work into two kinds of friction.
Vicious friction is effort that adds no value for the customer or the business. Chasing internal updates, re-keying the same data into multiple systems and compiling reports nobody uses all fall into this category. It drains time and attention, and it is the first place AI should be applied.
Virtuous friction is effort where human judgment is the value. A customer service agent who reads a complaint, connects the dots across three previous orders and writes a reply that actually resolves the problem is doing work customers notice and remember. Automating it might save minutes, but it would remove the part of the interaction that builds trust.
The goal is to automate vicious friction aggressively, protect virtuous friction deliberately, and reinvest the time freed up into more of the work only people can do well.
What stays human in commerce operations
Omar Qari, CEO of Logicbroker, argues that AI will change commerce operations one task at a time rather than one role at a time. “The way you need to understand how we’re going to move up this curve is not person by person or by org,” Qari says. “It’s really task by task.”
Some tasks will be fully automated, such as ranking exceptions or drafting a dispute with the supporting evidence attached. Others will become hybrid, with AI preparing the work and people making the call. And some will remain firmly human. “A system could summarize a buyer’s history, but a person still has to absorb some other person’s frustration or manage the relationship,” Qari says. “We believe that relationships and judgment and coaching, those things are still going to be very human.”
Put simply, the work falls into three tiers. Fully automated tasks include actions like ranking exceptions by priority or assembling a dispute with the supporting evidence already attached. Hybrid tasks involve AI preparing the work, for example summarizing a buyer’s history, while a person makes the final call. Fully human tasks include absorbing a frustrated customer’s complaint, managing the account relationship, or coaching a colleague through a hard case.
As businesses delegate more authority to agents, new human responsibilities appear. Qari says people will be needed to triage what actually happened, check for drift in agent behavior, and decide when to change policy boundaries.
Those oversight responsibilities are distinct. Triage involves reviewing what an agent did overnight and deciding what deserves a closer look. Drift monitoring means watching for behavior that is sliding away from how the agent was configured to act, even when no single decision looks wrong on its own. Policy boundary review is about deciding when a boundary itself needs to move, tightening it after a near-miss or loosening it once a pattern has proven reliable.
Those oversight tasks require deep operational knowledge, which brings the challenge full circle. The people best equipped to supervise agents are the ones who understand the work, and that understanding has to come from somewhere, which is why Logicbroker treats autonomous commerce operations as something built around people, not instead of them.
How Logicbroker builds this into its agentic commerce platform
Logicbroker’s agents are designed to resolve well-defined, routine commerce exceptions without a person in the loop, and to escalate ambiguous, high-risk, or unfamiliar cases to a human reviewer. That mirrors the task-by-task argument Qari makes: the tiers of work and the oversight responsibilities describe how Logicbroker expects its platform to be used.
Logicbroker lays out how that balance shifts as a business matures its use of AI on its AI Strategy & Commerce Maturity Curve. The general pattern it describes is one where automation starts by surfacing recommendations for a person to approve, and only takes on more autonomous resolution as trust in the underlying data and the agent’s track record builds. The mechanism that makes a policy boundary enforceable, rather than just a stated intention, is Logicbroker’s Model Context Protocol (MCP) server, which exposes a defined set of commerce actions and data to AI agents through its agentic integrations rather than giving agents open-ended access to every system. The same task-by-task logic also runs through how Logicbroker describes automation for agentic business models, rather than treating AI adoption as one decision applied the same way everywhere.
Protect the pipeline of future experts
If junior staff never learn to do the work manually, and senior staff are too stretched to mentor them, organizations risk losing the expertise they will need to oversee AI in five or ten years.
Protecting that pipeline does not mean refusing to automate entry-level tasks. It means being deliberate about how people build judgment. Junior team members can review agent output alongside experienced colleagues, handle a rotating share of exceptions by hand, or own the virtuous friction in a process while agents take the routine work. The aim is that every person still develops a real understanding of how the operation works, even as less of it is done manually.
Bring people in before the rollout
Technology deployment is usually the easier half of an AI program. Getting people to adopt new ways of working is harder, and agreement at leadership level does not guarantee that a change will be implemented throughout the organization.
One large organization rolling out an AI customer service platform credited its success to treating every affected employee as a stakeholder from the start. Staff helped shape how the system was used, which built trust in it and surfaced a benefit nobody had planned for: the platform helped the team identify upsell opportunities in customer conversations. Involving people early made the rollout faster and the result more valuable.
For more examples of how organizations have approached this in practice, Logicbroker’s case studies cover a range of rollouts across retailers, brands and suppliers.
Four questions to ask before automating a task
Is this vicious or virtuous friction? If the effort adds value the customer or business would miss, think carefully before removing it.
What skill does this task build, and where else will people build it? If the answer is nowhere, design another route for that learning before automating it away.
Who will check the output, and are they equipped to? Oversight is only meaningful when the reviewer has the expertise and confidence to disagree with the system.
What happens if the system goes down? A business that has removed all human capability around a process has no fallback when a model, platform or vendor fails.
Designing AI around your team
The businesses that get the most from AI are those that use agents to clear away low-value work so their teams can spend more time on supplier relationships, customer problems and the judgment calls that keep commerce running. That is the approach Logicbroker takes with its agentic commerce platform: agents handle the routine, and people stay in charge of the decisions that matter.