.webp)
AI adoption is becoming commonplace. Enterprise-scale returns are not. In most cases the difficulty is not what the technology can do — it is where the technology has been placed within the work. This article examines why AI investments underperform, what current research suggests about the capabilities that remain distinctly human, and how to design automation that strengthens the customer and employee experience rather than eroding it.
AI Adoption Is No Longer the Hard Part
Almost every organization is now using artificial intelligence somewhere in the business.
McKinsey's State of AI survey found that 88 percent of respondents report regular AI use in at least one business function, up from 78 percent the year before. The same research found that nearly two-thirds of organizations have not yet begun scaling AI across the enterprise, and that only 39 percent could attribute any enterprise-level earnings impact to it.
Adoption, in other words, has outrun integration.
The customer service function offers the clearest illustration, partly because it invested earliest and most heavily. According to Gartner, service and support leaders committed a median of 12 percent of their 2025 budgets to AI — the highest share among the ten business functions assessed. Only 24 percent of those leaders could demonstrate positive financial returns across their AI use cases.
At least in customer service, that result cannot be explained by technological capability alone. Gartner's own analysis attributed the disappointing impact less to the limitations of the technology and more to a misalignment between what companies deployed and what customers expected from it.
The same pattern appears on the workforce side. Among large enterprises already piloting or deploying autonomous-business capabilities, Gartner found that approximately 80 percent reported workforce reductions. Yet reduction rates were nearly the same among organizations reporting higher returns and those seeing modest or negative outcomes. Cutting cost and creating value turn out to be different exercises.
A capable model can still sit inside a poorly designed workflow. When that happens, the problem is not necessarily what the technology can do. It is what the organization has asked it to do, where it has been placed, and what happens around it.
We have written previously about why most business transformation projects fail. AI produces a particular version of that failure, with its own mechanisms and its own costs. This article is about that version.
Automation and Augmentation Are Different Decisions
Public discussion of AI tends to collapse into a single question: will this replace people or not?
That framing is not precise enough to guide an investment. A more useful distinction comes from research at the MIT Sloan School of Management, where Isabella Loaiza and Roberto Rigobon developed the EPOCH framework.
Their work separates two things most organizations treat as one. Automation transfers a task from a human to a machine. Augmentation uses a machine to increase what a human can accomplish — in that task and in the tasks connected to it. Analyzing roughly 19,000 work activities drawn from the federal O*NET occupational database, the researchers scored each one for human intensity, potential for augmentation and risk of substitution.
EPOCH names five groups of capability where machines remain limited: empathy and emotional intelligence; presence, networking and connectedness; opinion, judgment and ethics; creativity and imagination; and hope, vision and leadership.
What makes this commercially interesting is that these capabilities tracked with employment growth. Tasks scoring high across the five groups were associated with growth over the period studied, and tasks newly added to the database in 2024 scored higher on them than the tasks that had disappeared. The research establishes association rather than cause, but the direction is consistent.
Rigobon has been careful to reject the language usually applied here. His team avoids calling these “soft” skills, observing that a hard skill such as solving a mathematical problem is comparatively easy to teach, while empathy, judgment and hope are considerably harder to develop in a person.
For a business, this converts a philosophical question into an operational one. Before selecting a tool, an organization can ask of each significant task: is this a candidate for substitution, or a candidate for augmentation? Those two answers lead to different designs, different budgets and different risks. One recurring design mistake is treating a task as a candidate for substitution when its greater value lies in augmentation.
Three Ways Businesses Put AI in the Wrong Place
1. In front of the customer instead of behind the employee
The most visible misplacement is at the front door.
The economics can look compelling on a spreadsheet. A U.S. Consumer Financial Protection Bureau report, drawing on industry estimates, cited savings of roughly seventy cents for each chatbot-handled interaction compared with human-agent service. Across a contact centre, that compounds quickly.
A review published in the California Management Review at UC Berkeley documents why the saving frequently fails to appear. It identifies a recurring problem across customer-service research: poorly designed chatbots generate frustration, anger, repeated contacts and unsuccessful handoffs to human agents. Only a small fraction of service issues are fully resolved through self-service. Rather than passing the customer to a person, many systems continue responding with further questions, repetitive information and apologies — a pattern the authors describe as the chatbot loop.
An interaction that ends in a loop has not been deflected. It has been delayed, and it arrives at a human agent later with an angrier customer attached. The seventy cents was never saved.
Customer behaviour is now shifting in response. Gartner's 2026 research found that customers are roughly three times more likely to use third-party generative AI tools during a service journey than company-provided chatbots, and that use of those third-party tools has grown while use of company chatbots has stayed broadly flat. Customers are also increasingly using these tools to complete tasks rather than simply retrieve answers, and they expect a clear route to a person when a company uses AI.
Customers are not necessarily rejecting AI, then. They are increasingly finding their own AI tools more useful than some of the automated experiences companies have built for them. In effect, a growing number are routing around company-owned automation rather than through it.
The response is not to abandon automation in service. It is to move it. Many of the same capabilities can create more value behind the employee than in front of the customer: assembling the full interaction history before a call begins, surfacing unresolved issues, drafting a response for a person to review, flagging accounts at risk, routing a case to someone actually qualified to close it.
Consider a claims adjuster at an insurer. A customer calls because their claim sits outside the normal rule set. AI can retrieve the policy, summarize the prior correspondence, identify the missing documentation and prepare the adjuster before the conversation begins. Whether coverage applies to an unusual circumstance still requires contextual judgment, an explanation the customer can accept, and someone willing to be accountable for the answer. Automating the preparation improves that call. Automating the decision changes what the customer is actually buying.
2. On top of the team instead of inside the workflow
The second misplacement happens internally, and it is harder to see because it generates no complaint tickets.
Researchers at BetterUp Labs, working with the Stanford Social Media Lab, have documented a phenomenon they named workslop: AI-generated work that appears acceptable on the surface but lacks the substance to move a task forward. In a study of 1,150 full-time desk workers, roughly 40 percent reported receiving workslop in the previous month, and recipients spent close to two hours dealing with each instance.
The mechanism inverts the intended benefit. Ordinary automation offloads effort to a machine. Workslop offloads effort to another person — the recipient now has to decode the output, reconstruct the missing context and often redo the work. The work did not disappear. It moved downstream and became harder.
The costs the researchers considered most serious were not measured in hours. Recipients reported viewing the colleagues who sent them workslop as less capable, less reliable, less creative and less trustworthy than before. That is trust being spent, and trust is the operating currency of every collaborative process in a business.
In follow-up work, the authors were careful not to locate the problem solely in individual behaviour. Pressure to adopt AI without equally clear expectations for quality, review and accountability appears to be one contributor among several, alongside culture, incentives and inadequate training. They draw a useful distinction between two postures: people who pilot the tool, steering it toward a defined outcome, and people who ride it as passengers, using it to avoid the work. The same software in the same organization produces very different results depending on which posture the culture rewards.
3. Over human judgment instead of under it
The third misplacement is the slowest to appear.
When an organization routes a decision through an AI system, the immediate effect is speed. The longer-term effect is on the person who used to make that decision unaided.
The International AI Safety Report 2026 surveys emerging evidence on cognitive offloading — the delegation of mental work to external systems. Offloading has real benefits; it frees capacity for higher-order thinking. The report also cites a clinical study in which physicians' ability to detect tumours without AI assistance was approximately six percent lower three months after AI-assisted diagnosis was introduced. The report is explicit that long-term research in this area remains limited and the evidence is still developing.
Related findings point in a similar direction. A Microsoft and Carnegie Mellon study of 319 knowledge workers found that greater confidence in generative AI was associated with lower self-reported critical thinking during work tasks. Survey research published in Societies found a comparable association, with cognitive offloading acting as a mediator. Both measured reported behaviour rather than underlying capability.
None of this supports the claim that AI makes people less intelligent, and Pylet would not make it. The more defensible observation is also the more useful one for business design: human capability is partly maintained through use. When technology changes what people repeatedly practise, an organization should consider not only which capabilities the system adds, but which capabilities the new workflow stops exercising.
The finding raises a harder question than the productivity case usually acknowledges. If people repeatedly delegate part of a judgment process to AI, what happens to their unaided ability to perform that judgment over time? The risk is that an organization becomes progressively more dependent on a system while the human capability required to challenge that system receives less and less practice — which matters most at precisely the moment the system is wrong.
Governance Is Part of the Design
Every scenario above eventually resolves into one question: when the automated system produces the wrong outcome, who is accountable?
The NIST AI Risk Management Framework treats governance as a lifecycle function rather than a final approval gate. Its Generative AI Profile makes the related point that different AI uses can warrant different levels of human oversight, review, tracking and documentation, and it emphasizes documentation as a mechanism for transparency and accountability rather than as paperwork.
That framing matters commercially, because it changes when governance enters the process. Legal and governance requirements are not constraints to check after a solution has been selected. They are design requirements that can change which solution should be built in the first place.
Agentic systems sharpen the point. When a system can take actions rather than only produce text — invoking tools, executing transactions, initiating communications — the risk surface extends beyond what it might say to what it might do.
Any organization deploying AI into a customer-facing or decision-bearing process should be able to answer six questions:
- Which decisions may the system take autonomously, and which require human authorization?
- Who reviews its outputs, how often, and against what standard?
- What is disclosed to customers about where AI is being used?
- What data enters the system, where does it reside, and on what terms does the vendor hold it?
- What contractual and regulatory obligations apply in each market where it operates?
- When an error occurs, who owns remediation — and is that ownership documented?
If an organization cannot answer all six, its AI governance design is not finished.
Because AI regulation differs by jurisdiction, sector and use case, and continues to change, the specific obligations will vary. The design principle does not. Strategy, systems, people and legal obligation move together, or the architecture has a hole in it.
The Pylet Principle: Automate the Load. Protect the Relationship.
Organizations that get returns from AI tend to have made the same underlying distinction, whether or not they have named it.
Every process contains load — the retrieval, preparation, reconciliation, administration and routine drafting required for work to proceed. Every process also contains relationship value — the trust, ambiguity, judgment, reassurance, negotiation, accountability and recovery that a customer or colleague is actually relying on a person for.
The dividing line is not human contact versus machine contact. Plenty of contact automates well and improves in the process: a password reset, a shipping status request, a change to a reservation time. Customers generally prefer those handled instantly. The line falls between work that carries relationship value and work that does not.
Automate the load
- Retrieving records and interaction history
- Summarizing long documents and threads
- Drafting first versions for human review
- Triage and intelligent routing
- Reconciling data across systems
- Monitoring and anomaly detection
- Scheduling and administrative follow-up
Protect the relationship
- Onboarding and first impressions
- Delivering bad news
- Apologies and service recovery
- Negotiation and pricing exceptions
- Ambiguous or precedent-setting judgment
- Decisions with legal or ethical weight
- Situations where emotion is part of the problem
The test for the second list is not whether AI is technically capable of handling the moment. Increasingly it is. The test is where the value sits. When a meaningful part of the value resides in the relationship itself, automating that moment can remove part of what the customer came to the business to receive.
A second example makes the boundary concrete. A salesperson preparing for a renewal conversation can have AI assemble the account history, identify unresolved service issues, summarize purchasing patterns and flag the three things most likely to come up. The negotiation itself — what to concede, what to hold, how much latitude the relationship can absorb — stays with the person. The preparation is load. The conversation is not.
There is an overlooked workforce consequence here, and it deserves more attention than it usually receives. The goal is not to preserve every existing task or role exactly as it stands today; some roles will shrink or be reorganized as capability expands. The goal is to identify which human capabilities continue to create value as the composition of the work changes. When routine work disappears, what remains tends to concentrate in exceptions, judgment, conflict and emotionally demanding interactions. That is harder work, not easier. Role design, training, workload and compensation may need to change with it.
A Six-Step Approach to Human-Centred AI Transformation
1. Map before you buy. Score the work before selecting the tool. For each significant task, assess risk of substitution and potential for augmentation separately. The output is a map of what to automate, what to support and what to leave alone — and it is the document that prevents a capable platform from being aimed at the wrong problem.
2. Fund the organizational half of the change. McKinsey's research on agentic adoption describes a 1:3:5 pattern in successful AI transformations: for every dollar invested in the technology, roughly three go to process redesign and five to capability building and adoption. Most organizations invert it. Whatever the precise ratio in a given business, the underlying point holds — the technology is often the least expensive part of meaningful transformation. In the same research, leaders identified change management and siloed ways of working as larger obstacles to scaling AI than technology infrastructure.
3. Design the handoff before you scale the automation. Decide in advance how a case moves from system to person, what context travels with it and how quickly. Customers will forgive an AI that could not solve their problem. They will not forgive being trapped, or being asked to repeat what they have already explained. The escalation path is a core feature, not a fallback, and it should be built first.
4. Set the norms so people pilot rather than ride. Publish explicit expectations about where AI may be used, what review is required before work passes to a colleague, and what quality standard applies regardless of how the output was produced. Pressure to adopt without guidance on using the tools well is a documented contributor to poor internal outcomes.
5. Keep human judgment in practice, deliberately. Where a system supports a consequential decision, build in the conditions that keep the underlying human capability exercised — periodic unaided review, second opinions on a sample of cases, structured disagreement with the system's recommendation. This is not ceremony. It is how an organization retains the ability to notice when the system is wrong.
6. Assign ownership of outputs, not just uptime. Someone must remain accountable for what the system produces after launch: reviewing performance, updating guardrails, retiring use cases that are not working, answering for errors. Without a named owner, an AI deployment becomes another legacy system — still running, no longer trusted, quietly worked around.
Measure What the Automation Changes
Deflection rate is the metric most likely to make a failing deployment look successful. A tool that removes 80 percent of tickets from the queue and returns 30 percent of those customers as complaints has improved nothing except the report.
A serious measurement framework tracks four dimensions at once:
- Business. Cost, cycle time, throughput, accuracy, revenue.
- Customer. First-contact resolution, customer effort, repeat contact rate, satisfaction, retention.
- Employee. Rework volume, adoption, workload composition, autonomy, confidence in the system.
- Risk. Error rates, override frequency, escalation accuracy, incidents, audit findings.
The four are read together, not separately. A deployment that improves the first dimension while degrading the other three has not created value; it has relocated cost into places the finance report does not show. This is why weak AI deployments can look successful before they look expensive: the efficiency appears in one metric while the cost reappears somewhere the original business case was not designed to measure.
Building a More Capable Business, Not Simply a More Automated One
The case for keeping people in the design does not rest on sentiment.
The available evidence points the same way from several directions. Replacement-oriented deployments are producing weak returns, and in customer service a measurable share of customers are working around them. The capabilities AI has not absorbed — empathy, presence, judgment, creativity, and the ability to set a direction others will follow — are associated with the work that is growing rather than shrinking. Organizations that define when outputs require human validation appear to capture more value than those leaving review informal.
Technology does not build companies. People do. AI should change what those people spend their days doing; no organization benefits from capable staff performing retrieval and reformatting. But when automation crosses from the load into the relationship, the arithmetic reverses, and savings become churn, rework and eroded trust that no dashboard attributes back to the decision that caused it.
At Pylet, we approach AI as an operating-model decision before we treat it as a technology decision. The work begins with the business problem, the workflow, the people affected, the customer experience and the risks involved. Only then can an organization decide what should be automated, what should be augmented, and what should remain deliberately human.
Explore how Pylet approaches AI and business transformation.
