Author: IRPA AI Senior Analyst, Kieran Gilmurray
Most organisations are still treating AI as a task accelerator rather than a workflow redesign challenge. They use it to draft faster, summarise faster, search faster, respond faster, and produce more output. That is useful, but it is not the same as transformation.
In this article, we explore why AI value depends on redesigning workflows, not simply speeding up individual tasks. The argument is simple: AI does not create durable enterprise value by making broken work faster. It creates value when leaders redesign the flow of work around outcomes, constraints, decisions, handoffs, human judgement, AI roles, and measurable performance.
Why faster tasks do not automatically shorten cycle time
The problem with many AI initiatives is that they start at the task level. A team asks where AI can help and quickly finds useful opportunities: draft the report, summarise the meeting, classify the ticket, generate the response, search the knowledge base, or prepare the analysis.
Those use cases can produce real gains. OECD’s review of experimental studies found task level productivity improvements across areas such as customer support, software development, consulting, writing, and summarisation. The issue is not that AI fails at tasks. The issue is that task speed is not the same as workflow speed.
In most knowledge work, cycle time is not dominated by the time it takes to complete one task. It is dominated by waiting, handoffs, approvals, missing information, rework, exception handling, and coordination. A draft may take five minutes instead of fifty, but if it still waits two days for review, the workflow has not meaningfully improved.
That is why organisations can report high adoption and still struggle to show enterprise impact. McKinsey’s 2025 global survey found that 88% of organisations were using AI in at least one business function, yet only about one third had started scaling, only 39% reported enterprise EBIT impact, and just 6% met the threshold for AI high performers. The difference was not model access. It was operational redesign.
Where knowledge work actually gets stuck
Knowledge work often looks slow because people are slow. In reality, the system around them is often the constraint. Work gets stuck between teams, inside inboxes, across fragmented systems, in unclear approvals, and in the search for the right information.
Microsoft’s 2025 workplace research reported that employees are interrupted every two minutes during the workday, with 275 interruptions a day, while 48% of employees said work felt chaotic and fragmented. That is not just a productivity problem. It is a workflow design problem.
The same pattern has been visible for years. McKinsey’s earlier research found that interaction workers spent large portions of their week on email and searching for information. Harvard Business Review has also documented the growth of collaboration overload, where employees spend more time coordinating work than doing the highest value work itself.
AI can help with some of this, but only if leaders redesign the flow. If AI simply generates more drafts, more messages, and more analysis inside the same fragmented environment, it may increase output while leaving the real constraint untouched.
The danger of automating the mess
Automating the mess happens when AI is added to a broken workflow without changing the workflow itself. The result is usually more activity, not better performance.
A customer service team may generate responses faster, but escalation rules remain unclear. A finance team may produce analysis faster, but decision ownership remains fragmented. A legal team may draft faster, but review and sign off still sit in the same bottleneck. A software team may write code faster, but pull request review, testing, and deployment remain constrained.
In each case, AI improves a local activity while the system constraint remains. Sometimes the workflow even gets worse because faster upstream output creates more downstream review, more rework, and more exception handling.
This is why leaders need to be careful with claims about time saved. A task may be faster, but the organisation only benefits when the end to end workflow improves. The question is not “did AI speed up the task?” The question is “did AI reduce cycle time, improve quality, lower rework, or increase throughput in the actual workflow?”
Start with the outcome, not the tool
A better redesign process starts with the outcome. What business or customer result is the workflow meant to produce? Faster claims resolution? Shorter time to quote? Better forecasting? Faster onboarding? Reduced clinical documentation burden? Higher first contact resolution?
The wrong starting question is “Where can we use AI?” The right starting question is “What outcome are we trying to improve, and where does the workflow currently break down?”
Once the outcome is clear, leaders should map the workflow from trigger to completion. What starts the work? What information is needed? Which decisions occur? Where does the work wait? Who approves it? Where does rework happen? What exceptions break the normal flow?
This is where the real AI opportunity appears. AI may be useful for drafting, but the bigger opportunity may be earlier in the workflow, where information is gathered. It may be useful for summarisation, but the real constraint may be decision latency. It may be useful for automation, but the real value may come from better triage and routing.
The point is to redesign around the constraint. AI should be placed where it removes waiting, improves first pass quality, reduces coordination load, or shortens the path to decision. It should not be placed where it simply makes a visible task cheaper.
Redesign the workflow before adding autonomy
AI autonomy should not be treated as a default setting. It should be designed around risk, reversibility, and consequence.
In many workflows, AI should begin by assisting. It can retrieve information, summarise context, draft responses, classify requests, or prepare evidence for a human. This is useful where interpretation still matters and the cost of a wrong action is meaningful.
In other workflows, AI can recommend. It can suggest a next best action, prioritise cases, flag risk, or propose a decision. The human still approves, but the workflow becomes faster because the decision arrives with better context and structure.
Only in bounded, lower risk, repeatable workflows should AI act with limited autonomy. Even then, it needs clear thresholds, permissions, monitoring, exception routing, and rollback. The goal is not to maximise autonomy. The goal is to match autonomy to the workflow.
That distinction matters because more autonomous AI is not automatically more valuable. Deloitte’s research suggests measurable ROI remains harder for agentic AI than for more established generative AI use cases. Complexity, integration, and control design all matter.
What humans should keep and what AI should do
The most effective workflow designs do not ask whether humans or AI should do the work. They ask which parts of the work require judgement and which parts require speed, scale, memory, or structured execution.
The goal is not to divide work equally between humans and AI. It is to allocate work according to where judgement, accountability, speed, and scale create the most value.
AI is well suited to retrieving context, summarising information, classifying requests, drafting routine outputs, reconciling structured data, routing work, monitoring exceptions, and executing bounded actions where success criteria are clear.
Humans should retain responsibility for goals, judgement under ambiguity, risk appetite, exception resolution, trade offs between customer and commercial outcomes, and approval of high consequence decisions. Humans should also remain accountable for the design and improvement of the workflow itself.
This division of labour prevents two common mistakes. The first is underuse, where AI is limited to low value drafting even though it could remove real workflow friction. The second is overuse, where AI is allowed to act in areas where the organisation has not defined risk, escalation, or accountability.
The metrics that prove real improvement
Usage metrics are not enough. Prompt volume, chatbot activity, licence activation, and user counts can show adoption, but they do not prove workflow value.
Leaders should measure the redesigned workflow through a small set of practical metrics:
Cycle time: how long the workflow takes from trigger to completed outcome
Throughput: how many completed cases, tasks, or outputs the workflow handles
Quality: first pass accuracy, defect rate, or customer accepted output
Rework: how often work loops back for correction or missing information
Exception rate: how many cases leave the normal path and how long they take to resolve
Cost per case: the full cost of completion, including review, correction, and platform cost
Customer or stakeholder outcome: resolution, conversion, satisfaction, time back, or decision quality
Human review effort: how much oversight is still required to make the workflow safe and reliable
These metrics make it harder to confuse activity with performance. A workflow that starts more work but finishes the same amount has not improved. A workflow that drafts faster but increases rework has not improved. A workflow that reduces labour time but increases risk has not yet proved value.
How to pilot one workflow properly
The best way to start is not with a broad AI programme. It is with one workflow that has clear volume, visible friction, measurable outcomes, and a real business owner.
First, select the workflow. Choose an area where delay, rework, or coordination pain is obvious. Then measure the baseline: cycle time, throughput, quality, cost, rework, exceptions, and customer outcome. Without a baseline, the organisation will end up relying on anecdotes.
Next, identify the true bottleneck. Is the constraint intake quality, missing data, expert review, approval, handoff, customer response, system update, or exception handling? The answer determines where AI should sit.
Then redesign the flow. Remove unnecessary approvals, standardise intake, clarify decision rights, define exception paths, assign human and AI roles, and build the controls before scaling. Only after that should the team pilot the AI-enabled workflow.
The pilot should measure the workflow, not just the model. The test is not whether the AI produces a good answer in isolation. The test is whether the redesigned workflow improves speed, quality, cost, risk, and experience compared with the baseline.
What leaders can learn from real examples
Morgan Stanley shows the value of workflow fit. Its OpenAI-supported tools were embedded into advisor knowledge retrieval and meeting follow-up, helping advisers find internal information, generate notes, prepare follow-ups, and update client workflows. The value came from placing AI inside the flow of advisory work, not from offering a generic chatbot.
Intermountain Health shows the same pattern in clinical documentation. Dragon Copilot was embedded directly into the clinical workflow, reducing time spent on notes and helping clinicians spend more attention on the patient encounter. This is a strong example because the workflow problem was clear: documentation burden.
Verizon’s use of AI in customer service also points to the importance of workflow redesign. AI was used to support service agents with knowledge retrieval and live assistance, helping shift the role from issue handling toward resolution and selling. The value came from changing the role economics, not simply reducing call handling effort.
Klarna is the useful cautionary case. Its AI assistant delivered impressive customer service automation results, but later public reporting showed a shift in emphasis away from pure cost reduction and toward growth and quality. That matters because workflow automation can deliver speed while still requiring judgement about customer experience, trust, and human support.
Governance must be designed into the workflow
AI governance fails when it lives in a policy document but not in the work itself. A policy can define intent, but the workflow is where risk appears: a model retrieves the wrong information, a recommendation is accepted too quickly, an exception is missed, or a decision moves faster than accountability.
That is why governance has to be designed at the same time as the workflow. Every AI enabled workflow needs a clear owner, defined decision rights, risk thresholds, approval rules, permission boundaries, audit trails, exception routes, and rollback mechanisms.
As AI moves from assistance to recommendation and then into bounded action, governance becomes part of operational design. Leaders need to decide where human judgement is required, where AI can act within limits, when escalation is triggered, and who owns the outcome if something goes wrong.
The practical test is whether the organisation can explain how the workflow works under normal conditions, how it behaves when something unusual happens, and how accountability is preserved when AI has influenced the outcome. Without that clarity, AI does not remove risk from the workflow. It simply makes the risk harder to see.
Conclusion
AI does not fix broken workflows. It exposes them. Faster drafting, faster summarisation, and faster analysis are useful, but they do not automatically shorten cycle time or create enterprise value. The real unit of AI transformation is the workflow: the outcome, the constraint, the handoffs, the decisions, the human roles, the AI roles, the controls, and the metrics.
The organisations that outperform in the AI era will not be those that automate the most tasks. They will be the ones that redesign workflows faster than competitors can adapt their operating model.
About the Author: Kieran Gilmurray
Kieran is a globally recognized authority on AI, automation, and digital transformation, having authored multiple influential books and hundreds of articles that have earned him prestigious accolades, including being named a Top 50 Global Thought Leader and Influencer on Generative AI in 2024, a Best LinkedIn Influencer for AI and Marketing, Top 50 Global Thought Leaders and Influencers on Manufacturing 2024, Top 14 people to follow in data and one of the World’s Top 200 Business and Technology Innovators.
CLICK HERE TO SCHEDULE AN ANALYST CHAT
Links:
Originally posted on 2025-01-28 in the IRPA AI Network — Announcements & Updates