Last updated November 15, 2026
In a six-month randomized field experiment across 66 large firms, researchers at Microsoft Research and Harvard Business School gave half of 7,137 knowledge workers access to Microsoft 365 Copilot. Workers who used it spent about two fewer hours a week on email than the control group, while their time in meetings did not change significantly.
Knowledge workers change their workflow for AI when the team process their work runs through changes with it. At The Decision Lab, our AI adoption work usually starts after a rollout has stalled, and at a Fortune 500 HR technology company with a 20,000-person workforce, two of the five barriers we found were poor workflow integration and unclear usage guidelines, neither of which an individual employee can fix alone.
The five parts of AI workflow redesign
AI workflow redesign means changing the agreements a team works by once AI is part of the work. Start with five parts of the team operating process:
- Handoffs: who passes what to whom, and in what state.
- Approvals: which AI-assisted work needs sign-off, from whom, and how fast.
- Meeting rhythms: what meetings are for once AI can produce the status update.
- Quality checks: how output gets reviewed when AI drafted part of it.
- Decision rights: what AI may be used for on its own, and who owns the result.
Why AI rollouts stall without workflow redesign
Most AI rollouts change the tool and leave the process alone. In the Microsoft experiment, the tool sped up work people could change on their own, such as email, while shared work such as meetings carried on as before. Workers with more close colleagues who also had the tool showed few additional changes, and the researchers suggested that larger shifts in responsibilities need time and broad institutional effort.
Survey data points the same way. In McKinsey's 2025 State of AI survey of 1,993 respondents, the roughly 1 in 16 organizations that reported significant value from AI and attributed at least 5% of their operating earnings to it were nearly three times as likely as the rest to say they had fundamentally redesigned individual workflows. A survey like this shows an association and cannot show that redesign caused the results.
In the Fortune 500 case, the tools worked and employees knew about them, but people could not see where AI fit in their actual job, and the five sections below start from that gap.
1. Handoffs
What changes: a handoff is any point where work moves from one person or team to another. With AI in the workflow, a first draft often arrives partly written by a model, and some handoffs exist mainly to broaden an early draft with a second specialist's view.
Why the old process fails: handoffs designed for fully human drafts send specialists the same routine drafting work, even when AI could supply a useful first pass. In a preregistered randomized field experiment at Procter & Gamble, published in Organization Science, Fabrizio Dell'Acqua and colleagues assigned 791 professionals to work on real product development problems alone or in two-person teams, with or without AI. Individuals working with AI produced solutions of similar quality to two-person teams working without it, and their proposals balanced technical and commercial angles whatever their own background.
What to change:
- Map the handoffs in one recurring workflow as it actually runs, including the informal ones that happen over chat.
- For each handoff, write down the state the work is in when it arrives and whether AI produced any of it.
- Reassess handoffs that exist mainly to broaden an early draft. Where AI can provide a useful first-pass perspective, shift specialists from routine drafting to targeted review and final judgment.
- Keep specialists in the loop wherever their domain expertise or accountability is the point of the handoff.
Decision rule: if a handoff exists to add expertise or accountability, keep it and change what arrives; if it exists only to widen a first draft, test moving it later in the process.
2. Approvals
What changes: AI-assisted work raises new questions at the approval step, such as whether AI use was allowed and how much checking the output needs.
Why the old process fails: an employee who is unsure whether AI-assisted work is allowed, or who expects it to be judged harshly, will either avoid AI or use it without saying so. In four preregistered experiments on the social evaluation penalty for using AI, with more than 4,400 participants, Jessica Reif, Richard Larrick and Jack Soll found that people described as using AI at work were rated as less competent and less motivated than people using other tools. Unclear usage guidelines were also one of the five barriers in our Fortune 500 case.
What to change:
- Sort recurring outputs into tiers by the cost of an error, from internal working notes up to anything with legal or financial consequences.
- Give each tier a default level of review, and make the lowest tier need none.
- Make AI-use disclosure a routine field in approvals, while defining which work requires additional source validation or legal review.
- Judge AI-assisted work on the quality of the output and the judgment behind it, and stop rewarding hours spent.
- Remove approval steps that existed only because work used to take a long time to redo.
Decision rule: the review an output gets should depend on what an error would cost, and AI involvement on its own should neither add nor remove a review step. Our guide on how to get employees to follow AI guardrails covers how to write usage rules people remember at the moment of use.
3. Meeting rhythms
What changes: a meeting rhythm is a team agreement about how often the team meets and what each meeting is for. Once AI can summarize a project's status from documents and messages, a meeting spent reading out updates has lost most of its purpose.
Why the old process fails: meetings are shared work, and no one person can change them. In the Microsoft experiment, time in meetings did not shift, even though meetings were where workers were most likely to use the tool.
What to change:
- List every recurring meeting and label its main job: sharing information, making decisions, solving problems together, or keeping relationships strong.
- Move information-sharing to a written, AI-assisted update that people read beforehand, and name who checks it for accuracy.
- Shorten or cancel meetings whose only job was status, and protect the ones where the team decides or works through problems together.
- Have the manager model the new rhythm by using the written update as the starting point, rather than requesting a verbal walkthrough.
- Agree who confirms the decisions and owners in AI-generated meeting notes before they are sent.
Decision rule: a meeting earns its slot if the team needs to decide or solve something together in it.
4. Quality checks
What changes: when AI drafts part of an output, the errors shift toward the kinds AI makes, such as invented sources, wrong figures, outdated policy or a tone that does not fit the reader.
Why the old process fails: the usual check is "a person reviews it," and on its own that check is weak. In a review of experimental research on how people use automated systems, Raja Parasuraman and Dietrich Manzey found that automation bias, the tendency to accept an automated system's output without checking it properly, appears in experts as well as novices, and in teams as well as individuals. Training and instructions alone do not reliably prevent it. In McKinsey's survey, the highest-performing organizations were more likely than others to have defined processes for deciding how and when model outputs need human validation.
What to change:
- For each AI-assisted output, list the errors that would matter most to its reader.
- Turn that list into a short checklist the reviewer completes, so that review means checking specific things.
- Name one reviewer per output, so it is clear who signed off.
- Log errors found after release and use them to update the checklist each month.
Decision rule: if a reviewer cannot say which specific errors they checked for, the output has not been reviewed.
5. Decision rights
What changes: decision rights say who can decide what. Once AI enters a workflow, someone has to decide which tasks AI may complete without review and who can change those rules as the tools improve.
Why the old process fails: when nobody owns these questions, employees answer them individually and inconsistently, and cautious employees default to not using AI at all.
What to change:
- For each recurring workflow, write down what AI may do on its own and what it may only draft for a person to finish.
- List the tasks AI should not be used for, and say why, so the line is easy to apply to new cases.
- Keep accountability with a person: whoever sends the work owns it, whether or not AI drafted it.
- Name one person on the team who can update these rules, and review them every quarter.
Decision rule: every task in the workflow should have a named human owner, including the tasks AI completes.
A six-step sequence for redesigning a team process around AI
- Pick one recurring workflow the whole team touches, such as a weekly client report or a hiring loop.
- Map how the work actually runs today, including informal handoffs and the time each step takes.
- Mark where AI changes the cost of each step, and where it changes who needs to be involved.
- Rewrite the five parts above for that workflow, with the people who do the work.
- Pilot the new process for four to six weeks, ideally alongside a similar team that keeps the old process, so you have a comparison.
- Measure process outcomes such as turnaround time and rework, alongside usage.
The comparison in step 5 matters because people misjudge their own AI-assisted speed. In a randomized trial run by the research group METR, 16 experienced open-source developers took 19% longer to complete real tasks from their own projects when they were allowed to use AI tools than when they were not. Afterward, the same developers estimated that AI had made them about 20% faster.
How The Decision Lab redesigns AI workflows
Our AI adoption work treats low usage as a behavioral problem that usually sits in the process, and we design and test changes to that process with the people who run it.
In our case study with the Fortune 500 HR technology company, we built a map of how employees moved from first exposure to everyday AI use, drawing on a 48-source literature review, workforce interviews, surveys and expert consultation. We then designed 20 candidate programs, each aimed at a specific barrier for a specific group of employees, and piloted eight of them for one month with more than 100 employees against a control group. The programs ranged from hands-on tasks embedded in employees' real workflows to a challenge series run in Slack. Confidence in exploring AI tools, measured before and after the pilot, rose 41% among participants and fell 26% among control-group colleagues over the same period. The programs are now scaling to the full 20,000-person workforce, with the protocols and guides the client's own teams need to run them. The full story is in our case study on increasing AI adoption across a Fortune 500 workforce.
We diagnose where a team's process is blocking AI use with our SPROUT framework for AI adoption, validated with data from more than 20,000 employees. Two of its five dimensions, Role Fit and Organizational Enablement, cover most of what this guide describes: whether AI fits the work people actually do, and whether policies, leadership, incentives and culture support using it. For the barriers that sit with individuals, see our guide to behavioral barriers to enterprise AI adoption, and for the wider adoption program, our seven evidence-based steps to increase AI adoption across your workforce.
Frequently asked questions
What is AI workflow redesign?
AI workflow redesign means changing how a team's process runs once AI is part of it: who hands work to whom, what needs approval, what meetings are for, how output is checked and who owns decisions. AI adoption training teaches individuals to use tools inside a process that stays the same. Redesign changes the process those individuals work within.
Why don't employees change their workflow after getting AI tools?
Most knowledge work is shared, so one person changing their workflow runs into reviewers and meetings that still expect the old one. In a randomized field experiment with 7,137 knowledge workers, AI access cut time spent on email without changing time spent in meetings. Shared work needs team-level agreements that no single employee can make.
Who should own AI workflow redesign in a team?
The manager of the team doing the work should own it, with the people who run the workflow each day doing most of the redesign. Central AI and IT teams can supply tools and guardrails. They rarely know a team's informal handoffs and workarounds well enough to redesign that team's process on its behalf.
How do you measure whether a workflow redesign for AI worked?
Compare the redesigned team with a similar team still using the old process for at least a month. Track turnaround time, rework, errors found after release and time spent in meetings. Treat self-reported time savings with caution: in one randomized trial, developers believed AI made them faster when it had made them slower.
How long does it take to redesign a team process for AI?
For one recurring workflow, mapping and rewriting the process usually takes a few working sessions, followed by a pilot of four to six weeks. Expanding across a department takes longer, because each team's handoffs and approvals differ. Starting with one workflow gives the team evidence to show before asking other teams to change.
Start with one workflow
Most teams we work with already have AI access and training, and the gap is in how their work is organized around it. If your AI rollout has stalled at that point, The Decision Lab can identify the behavioral and organizational barriers standing in the way of everyday use, and you can start today with our free AI adoption diagnostic for your team.

