2026/08/20More AI Doesn’t Mean More Productivity: Redesign the Workflow, Not Just the Task
Artificial Intelligence and Machine Learning | Appar Technologies Co., Ltd. | 2026/08/20
Artificial intelligence can now write documents, analyze data, generate code, and even autonomously execute a series of tasks for employees. However, companies will soon encounter a counterintuitive phenomenon: while the speed of each task increases, overall operations do not accelerate proportionally. Recent studies by McKinsey, BCG, Deloitte, and the Harvard Business Review all point to the same issue: the true business value of artificial intelligence often does not come from "doing the original work faster," but from re-evaluating which tasks still need to exist, which handoffs can be eliminated, and how humans, AI, and enterprise systems should redistribute tasks.
When Hewlett Packard Enterprise (HPE) introduced agent-based artificial intelligence, they deliberately did not start with "which employee needs an AI assistant the most," but instead chose a complete business process: Operational Performance Review. This task involves organizing operational status from a large amount of corporate data, analyzing anomalies, creating charts, and then compiling the results into actionable information for management. It is a comprehensive process that spans data collection, analysis, and decision support. HPE's CFO, Marie Myers, stated that the team chose this case because they wanted to truly change the end-to-end way of working, not just solve a small pain point.
The system developed by HPE and Deloitte is internally called Alfred. It does not just add a text summarization tool to the original reporting process but involves multiple AI agents handling query decomposition, SQL data analysis, chart creation, and management report compilation, retrieving information from the company's data warehouse. This data environment contains over 300 million data records and is connected to enterprise resource planning systems and customer relationship management systems. The originally static, retrospective reporting process thus begins to transform into a management tool that can continuously interact with data, explore issues, and provide the next steps for analysis.
The noteworthy aspect of this case is not how many AI agents HPE used, but the different "unit" they chose to transform. Many companies, when introducing AI, first break down work into individual tasks, such as writing summaries, creating charts, and searching for data, then ask how many minutes each task can save. HPE, however, started from the complete operational performance review process and then considered which steps should be executed by AI, which information can be directly connected, and what parts managers truly need to participate in. These two approaches may both be called AI implementation, but the organizational effects they produce can be completely different.
This is precisely the most apparent gap in current corporate AI adoption. McKinsey's 2025 Global AI Survey shows that AI usage in companies is already quite common, but the tangible value at the corporate level remains uneven; among the 25 management practices tested by McKinsey, redesigning workflows is the factor most clearly associated with the impact on earnings before interest and taxes brought by generative AI. However, among the surveyed companies that have already used generative AI, only 21% reported having fundamentally redesigned at least part of their workflows.
By 2026, this gap still exists. BCG surveyed 152 CEOs of companies with annual revenues of at least $500 million, and nearly 90% of respondents said they have seen cost or revenue benefits from AI in some areas, but most companies still struggle to expand these results into overall financial impact; companies that perform better are about seven times more likely to redesign workflows and transform business end-to-end than other companies. The real bottleneck of AI is gradually not whether the model can complete tasks, but whether companies are willing to redesign the way work operates.
Partial Efficiency Improvement Does Not Equal Organizational Efficiency Improvement
Suppose a company's marketing team originally needed two hours to write an event proposal. After introducing generative AI, the first draft of the proposal can be completed in just twenty minutes. From an individual's perspective, this is a very obvious productivity boost; if a company measures the benefits of AI by "how much time is saved on a single task," this project will almost certainly be deemed successful.
However, after a proposal is generated, it may still need to be confirmed by the product department, edited by the branding department, reviewed by legal for claims, and budget approved by supervisors before returning to the marketing team for revisions. The original two-hour writing task may be shortened to twenty minutes, but the subsequent waiting, handoffs, and repeated confirmations remain unchanged. In fact, because the cost of AI-generated content is lower, employees may start producing five versions, ten ideas, and more attachments, resulting in an increase in information that needs to be reviewed.
BCG described this phenomenon quite directly in their 2026 study on AI productivity: AI can increase output and work speed, but if existing steps, review levels, roles, and repetitive work still exist, the added productivity is easily absorbed by the original organizational complexity. Companies can see employees working faster, but may not necessarily see costs decrease, processes shorten, or financial performance improve simultaneously.
This is also why "how many times employees use AI daily" or "how many AI-generated contents are produced monthly" can hardly prove AI transformation success on their own. These numbers can only prove that the tool is being used, not that the organization's way of operating has changed. What truly needs to be observed is whether a task that originally took five days to complete now only takes two days, whether a process that originally required six manual handoffs now only requires two, and how many cases that originally needed supervisor confirmation have transformed into handling only exceptions.
Do Not Fit AI into a Workflow Designed for Humans
Business processes are usually not designed all at once but are the result of gradual accumulation over the years. A step may have been added because of a past issue, so an additional review was added; a system may lack an interface, so a person is assigned to copy data; another department needs to be notified, so another email is added. Each step had its reason when it appeared, but after many years, a task that was originally simple may have become a complex process spanning five departments, three systems, and multiple manual confirmations.
The Harvard Business Review discusses AI Agents and job design, using the history of early factories adopting electricity as an analogy. Initially, factories simply replaced central steam power with electric motors, retaining the original plant layout and transmission methods, resulting in limited improvements. It wasn't until companies redesigned factory layouts based on the characteristics of electricity that new production models truly emerged. Artificial intelligence faces a similar issue: if companies merely integrate AI into organizational structures originally designed for human work, they often achieve only partial efficiency improvements.
Deloitte's 2026 study on AI Agents also issues a similar warning: many companies attempt to automate existing processes that were originally designed with human work limitations in mind. Leading companies, however, return to end-to-end processes, re-examining how work should be completed rather than simply looking for "which step can now have an Agent." Deloitte even points out that poorly designed AI Agent projects can add extra work, making processes less efficient.
This presents a significant management issue. When evaluating artificial intelligence, companies often ask, "Can this step be done by AI?" But the truly valuable question should go a step further: "With current AI capabilities, does this step need to exist at all?" The former usually results in faster old processes; the latter may lead to new ways of operation.
Redefining the Responsibilities of Humans, AI, and Systems
Traditional business processes have an often unspoken premise: most information must be read by humans, and most cross-system actions require human operation. Therefore, companies break down work into departments, positions, and operational steps based on human capabilities and limitations, linking them together with forms, emails, meetings, and software interfaces. This is why many company processes involve a lot of "moving information from one place to another."
For example, after receiving a customer request, a salesperson might first organize the email content into a customer relationship management system and then notify technical staff. After estimation, the technical staff enters the results into a project management system. The salesperson then obtains pricing information to create a quote, the manager confirms discounts, and finally, it is sent to the customer. The only parts that truly require human judgment may be clarifying needs, special estimations, and exceptional discounts, while the rest is largely information organization, system operation, and status transmission.
As generative AI and AI Agents begin to understand documents, query enterprise data, call tools, and operate systems, these previously taken-for-granted divisions of labor deserve re-examination. Deloitte's research on AI Agents indicates that leading companies do not simply add tools to existing positions but start thinking from the entire process about which tasks humans and Agents should handle, allowing both to jointly complete the final outcome.
This also means that AI transformation is no longer just the responsibility of the IT department. The IT team can handle models, data, system integration, security, and technical architecture, but it cannot solely decide "which audits are unnecessary," "which customers can be automatically quoted," or "which exceptions require managerial approval." These issues require the involvement of business units truly responsible for operational outcomes, as process redesign is fundamentally a management issue, not merely a technical implementation issue.
Defining a Process Redesign Method Starting from Outcomes
When companies begin redesigning processes, they don't need to redo the entire company from the start. A more practical approach is to select a process with clear outcomes, spanning multiple steps, and currently involving a lot of manual organization, handoffs, or waiting, and re-examine it from start to finish. BCG also recommends that companies choose a complete process to redesign from end to end, rather than looking for scattered AI cases in different departments. Their research indicates that reinventing workflows can have an impact three to four times greater than traditional incremental improvements.
An actual process redesign can sequentially answer the following questions:
- What is the true outcome this process aims to achieve? It's not "producing a report," but "providing managers with information sufficient for decision-making"; it's not "creating a customer service ticket," but "resolving the customer's issue."
- Which steps truly require human judgment? Tasks involving strategy, relationships, exceptions, responsibility, and high-risk decisions usually still require human involvement.
- Which steps are merely information transfer? If employees are just copying data from emails to a customer relationship management system and then retrieving data from another system to paste back, these tasks are the first to be eliminated or integrated.
- Which audits are risk controls, and which are just historical habits? Not every approval should be canceled, but companies must know what risk they are mitigating; otherwise, AI will only speed up documents reaching the next waiting point.
- Which situations can be automated, and which only require exceptions to be escalated? If 90% of cases follow fixed rules, not all cases need to undergo manual review.
- Where should AI appear? Only after answering the previous questions should companies decide to use generative AI, traditional automation, AI Agents, or determine that AI is not needed at all.
This order is crucial. If the first question is "where can AI be placed," teams are likely to see existing processes as unchangeable premises; conversely, by first asking about outcomes, judgment, handoffs, and exceptions, AI becomes one of the tools for redesigning work, rather than the project's purpose itself.
Characteristics of a Good AI Workflow
From the corporate cases and studies of HPE, McKinsey, BCG, and Deloitte, we can see that processes that truly transform AI into operational results share some common characteristics. They do not pursue a 'completely human-free process,' but rather rearrange where humans should appear, how information should flow, and how the capabilities generated by AI can be transformed into real organizational outcomes.
- Focus on outcomes, not tasks: The goal of design is not to make ten tasks 20% faster each, but to significantly change the delivery time, quality, or cost of the entire process.
- Reduce unnecessary handoffs: If the information generated by AI still needs to be repeatedly transferred between people, documents, and systems, productivity can easily be lost during handoffs.
- Place humans at high-value judgment points: Employees do not need to intervene in every task but should focus on handling exceptions, risks, relationships, creativity, and decisions that require accountability.
- Involve enterprise systems directly in the process: If AI can only generate a piece of text that employees then manually input into formal systems, it is difficult to achieve true end-to-end improvement; data, tools, and workflows need to be gradually integrated.
- Simplify before automating: If a process has twelve steps, five of which are no longer necessary, the best automation is not to make all twelve steps faster but to remove the five unnecessary ones first.
- Measure end-to-end outcomes: In addition to saving labor hours, cycle time, manual handoff frequency, exception rates, costs, conversion rates, customer satisfaction, and actual financial impact should also be measured.
BCG describes this as moving from Productivity to Performance. AI can quickly create new work capacity, but only when managers actively redistribute this capacity, remove unnecessary work, and modify roles and decision rights, can productivity be transformed into real cost improvement and performance. Otherwise, employees are just using the new time to complete more of the existing work.
How to Transition from Single-Point AI Implementation to Workflow Redesign
The first step for enterprises is not to stop all current Copilot, ChatGPT, or generative AI projects. Single-point tools are still a good way to learn and can quickly build employees' understanding of AI capabilities. The problem is that companies cannot mistake 'many people starting to use AI' as a transformation that has already occurred; when individual productivity begins to improve, managers should shift their perspective to a higher level to observe where these efficiencies ultimately go.
A better starting point is to select a complete process with clear pain points and business outcomes. This process does not have to be the largest but should have easily measurable results, such as quote cycle time, customer service issue resolution time, procurement processing time, financial closing, project delivery, or quality anomaly handling. HPE chose the operational performance review process because it simultaneously has a large amount of data, clear management value, and cross-step improvement space.
Next, the team should first map out the process that is actually happening, not the process written in the company's institutional documents. Every wait, manual copy, email confirmation, system switch, and supervisor review should be visible. At this point, the team has the opportunity to discover that what really slows down work is often not someone typing too slowly, but rather information not knowing where it is, unclear responsibilities, system disconnections, and all cases being forced through the same control methods.
The third step is to redesign before selecting technology. Some steps may be suitable for large language models, some for traditional rule engines, some only need API connections, and some are suitable for AI Agents to handle autonomously. If a task with fixed rules can be safely completed with a regular program, there is no need to forcefully apply an Agent for 'AI transformation.' Deloitte in 2026 also specifically warns companies that many cases packaged as agent-based AI are actually just traditional automation, and choosing the wrong technology may reduce the return on investment.
Finally, the process must be continuously adjusted after going live. AI capabilities will change, corporate rules will change, and employees will gradually discover new ways to use them, so workflows should no longer be understood as a fixed system designed once and unchanged for years. The Harvard Business Review in 2026 pointed out that after AI reduces information coordination and experimentation costs, companies have the opportunity to transform past large, periodic processes into continuously evolving work methods.
From Usage Metrics to Operational Outcomes
In the early stages of AI implementation, companies can easily take adoption rates as success indicators. For example, how many employees open Copilot, how many people use ChatGPT weekly, how many Agents the company has established. These numbers are valuable in the promotion stage because they can show whether employees are really engaging with new tools; but when AI starts to enter formal operations, companies must gradually change the measurement method to be outcome-oriented.
The Harvard Business Review in 2026 called the phenomenon where companies only achieve individual-level productivity without forming overall business outcomes the 'micro-productivity trap.' Employees can complete emails, presentations, code, and analysis work faster, but if these improvements are not connected to complete business processes, it is difficult to reflect in revenue, cost, and competitiveness.
Therefore, the next stage of management dashboards needs to start showing different indicators. For example, how much the average time from quote request to formal submission has shortened; whether the proportion of customer service cases resolved successfully on the first attempt has increased; how many times a procurement process requires manual intervention on average; how many tasks originally required supervisor approval for each case, now only exceptional cases need handling; whether a process that originally required data exchange between three departments has been integrated into a single information flow.
These indicators, compared to 'how fast the model responds,' may not seem as much like AI indicators, but they are closer to the results that companies truly care about. The value of AI must ultimately fall into at least one of revenue, cost, speed, quality, risk, and customer experience; if all AI usage increases rapidly and these indicators do not change at all, managers should re-examine whether the company is achieving transformation or just acquiring more tools.
The Focus of AI Transformation is Shifting from Tools to Work Itself
Current AI technology is still rapidly developing, and companies will certainly continue to acquire more powerful and cost-effective models and Agents in the coming years. In other words, 'acquiring AI' itself will become increasingly easy. The truly difficult-to-replicate competitive capability, however, may be whether companies understand their own processes, know which decisions are truly important, and can rearrange the division of labor between humans, AI, and enterprise systems.
The Deloitte 2026 Enterprise AI Survey provides a noteworthy figure: more and more companies are seeing efficiency and productivity outcomes from AI, and the number of executives who believe AI has had a transformative impact has increased compared to the previous year. However, only 34% of companies truly state they are 'reimagining their entire business.' This indicates that most companies are still in the phase of integrating AI into existing work rather than redesigning work processes.
BCG's latest research similarly shows that companies that truly perform better are not just deploying more AI but are transforming AI along with processes, roles, decision-making authority, and operating models. When companies only pursue 'how much more work one person can do in a day,' improvements usually remain localized. When companies start asking 'why does this work require so many people, so many handoffs, and such long wait times,' AI can become a force for changing organizational design.
Therefore, now is a good time for companies to re-examine their processes. It's not because every task should be handed over to AI, but because many work methods established due to difficulties in obtaining information, limited manpower, and disconnected systems are no longer the only options. The ability managers truly need to cultivate is not just knowing the functions of the latest models, but being able to reassess what work is worth existing, what should be eliminated, what should be handled by machines, and where human judgment should be most preserved.
Technology itself will not automatically change a company. Producing the same report ten times faster does not mean the company's decision-making speed has increased tenfold; having every employee with an AI assistant does not mean the company's operating model has changed. The true transformative value of AI comes from companies willing to use new technological capabilities to redesign 'how work should be accomplished.'
So, when a company is preparing to procure new AI tools, establish the next AI Agent, or require more employees to start using AI, managers might first ask another question: If we were to design this process from scratch today, knowing AI has the capabilities it does now, would we still design the work the way it is today?
If the answer is no, then what the company truly needs to update next might not be the AI tools, but the work itself.
