Quick answer: To redesign workflows for AI, pick a process that passes through several teams, such as employee onboarding (HR, then the hiring manager, then IT). Choose one with repeated, brainstorming, or writing tasks. Record each step's owner and wait time, and remove steps that serve no purpose. Assign each task to AI, to AI with a person, or to a person, based on your tools, information, and reviewers. Rebuild with fewer handoffs and one owner. Then compare six signals with your starting numbers: AI use, handoffs and waits, review workload, contact between teams, rework, and business results.
A common approach is to add AI to the way work already runs. A 2025 Faros AI study of more than 10,000 developers found that teams using AI heavily merged 98% more pull requests (proposed code changes). The time those changes spent in review rose 91%. Faros says slow steps later in the process, such as review, were absorbing the gains.
The alternative is to redesign the work itself. Deloitte's 2026 research found that organizations that redesign work are twice as likely to exceed the return on investment they expected from AI. It also describes a European telecom that gained about 5% in productivity by adding AI without redesign. It gained about 30% when roughly 90% of its rollout effort went into redesigning how people and AI share the work.
Below are five steps, then a six-signal scorecard to check the result. It is written for organization leaders, including engineering leaders and HR directors.
A workflow is the series of tasks that turns a request into a finished result, such as a new employee ready to start. People, software, and AI do the tasks, and the work passes from one to the next. Each pass is a handoff. To redesign a workflow for AI is to change which steps exist, who or what does each one, and where work waits.
Workflow redesign changes which steps the work goes through and who does them. Restructuring changes how the organization is set up: teams, management layers, and roles. Reorganizing moves people between teams or changes who reports to whom. Deloitte's guidance is to redesign the work before you redesign the workforce, and this guide follows that order. Headcount is a separate decision from all three.
AI speeds up the person doing a task, not the person who has to review, approve, or accept the result. Faros also found that the average size of a pull request grew 154%, which makes each one slower to review. The same thing happens anywhere one team produces faster than another can handle it, such as HR opening onboarding requests faster than IT can work through them. Worklytics describes this in its piece on the AI productivity paradox: the slow point moves to the review and approval steps.
Nick Colisto, CIO of Avery Dennison, argues in a CIO.com column that every workflow reflects past decisions. AI can run an outdated process faster without making it simpler. In his example, AI can route purchase approvals, but it cannot decide whether six approvals are still necessary.
Start with work that passes through more than one team. Employee onboarding is one example: it goes from HR to the hiring manager to IT. Every handoff from one team to the next is a place for work to sit. In each workflow, list the tasks that suit AI. Step 3 decides which of them AI does alone.
Before you redesign, find out who already uses AI and for what, including tools people brought themselves. In the 2024 Work Trend Index, a Microsoft and LinkedIn survey of 31,000 knowledge workers, 78% of AI users said they bring their own AI tools to work. Check four places:
Admin reports show one tool at a time. A people analytics platform such as Worklytics combines usage data from tools like Microsoft Copilot, ChatGPT, Claude, Gemini, and Cursor in one AI adoption dashboard. Usage data means how often each person or team uses each tool, and you can split it by team, role, department, and manager. If you find little use, redesigning around AI may be premature. Help people adopt it first.
For each workflow you picked, write down every step, who is responsible, what goes in, what comes out, and who receives it next. Record two times for each step: how long the item waits before anyone starts, and how long the work itself takes. AI usually shortens the work time, but waits often stay the same unless you change the steps. Ask the people who do the work, not only their managers. Here is an example for employee onboarding:
Keep these numbers as your starting numbers, also called a baseline, and compare against them after the change. Use the median wait (the middle value) rather than the average, because a few very long waits distort an average.
Work waits at handoffs and approvals. Ticketing, finance, and workflow software record when an item entered and left each stage. Engineering's code review tools show how long changes wait for review and how many are waiting.
Worklytics's organizational network analysis adds the people side. It uses data about who contacts whom, how often, and when, from email, calendar, chat, and meeting tools, not what is said. It shows which teams work together, which have little contact, and which people link two teams. It shows contact, not waiting time, so pair it with the timestamps above. Meeting data can also hint at approval levels, such as how often a team's meetings include several levels of management.
Colisto's four moves, in order, are simplify, standardize, redesign, then automate. Do the first two before you decide what AI does. Simplify by removing approvals, reports, and handoffs that no longer serve a purpose. Standardize by agreeing on one way of doing each remaining step. His test for the first move: if we were designing this process today, would we build it the same way? Keep approvals that guard against costly mistakes, such as access to sensitive systems.
Deloitte's guidance is to break your most important workflows into tasks and decide which to automate, augment, or elevate. In plain terms, each task goes to one of three places: AI does it alone, AI helps a person do it, or a person does it. These four questions sort each task:
Worklytics sorts AI activity into task types such as writing code, research, analysis, summaries, and email, using usage data and activity signals, not the text of prompts. Treat the result as a starting point and check it with the teams who do the work.
The four questions show which tasks suit AI. What you have decides how many you can hand over now. Check these four resources for each task AI would do or help with:
For example, answering common policy questions suits AI. If the policies sit in scattered files the tool cannot read, a person keeps answering until they are in one place. The tools you found in Step 1 show what you have today. As you add tools, information, and reviewers, move more tasks to AI.
Fewer, clearer handoffs are better than many small ones. Write down where a person must approve and what sends a case to a senior person (an escalation). Then move each approval to the person closest to the work who has enough information to decide. Name one owner for the redesigned workflow and one for each AI step. The AI step owner looks after the tool, the instructions given to the AI (prompts), and the quality checks. Decide in advance what happens when AI output is wrong or a case does not fit the rules.
If your engineers will submit far more pull requests, reviewing them needs real people and time. Try smaller pull requests and a rotating schedule so reviewers share the load. Add an AI first pass followed by a human sign-off, and name one person responsible for how long changes wait for review. Worklytics connects to engineering tools such as GitHub and Jira. Check how long changes wait for review every week for the first three months.
Redesign one workflow first, and pick one that runs often enough to compare before and after within a few weeks. Compare its results with your starting numbers before you apply the same changes elsewhere. Once it works, check whether your teams are still organized the right way, because roles change when AI takes over tasks. You may need to restructure or reorganize if managers now have too many or too few direct reports. A 2024 Worklytics analysis found that managers with more than 7 direct reports were more likely to work 10 to 13 hour workdays. With too few, teams often suffer from over-management: the manager gets too involved in the details. The post suggests 4 to 7 as the best range, depending on the role and type of work. It names no sample size, so test the range against your own numbers.
People are less likely to adopt a redesign they do not understand. Bain surveyed 976 executives, managers, and employees whose roles were affected by a new organizational structure (fielded in late 2025). It found that AI-related changes underperformed other kinds of organizational change. Fewer than 40% said the scope and reason for the AI change were clear. Targeted support and coaching reached 59% of AI transformations, against 70% of change efforts in general.
Involve the people who do the work in the mapping and the redesign. Give them a one-page explanation of what is changing, why now, what stays the same, and what it means for their role. Then say when you will share the results.
Compare the results with the baseline you saved in Step 2, and track outcomes, not AI activity. Colisto makes the same point: measure cycle time (how long work takes from start to finish), quality, customer satisfaction, and cost, not the number of prompts submitted.
Do not read signal 1 alone. If many people use AI but the handoffs and waits are the same as before, the old steps are still in place. Worklytics describes a four-layer efficiency map: who uses AI, how deeply, what productivity changes, and what the recovered time is worth.
Check at 30 days, 90 days, and six months. Early checks cover behavior: AI use, handoffs, and items waiting for review. Later checks cover results: rework, time from start to finish, and satisfaction. A redesign disrupts work at first, so do not judge it on the first month. If a similar team or workflow did not change, compare against it, so a busy season does not look like a result.
These are common waiting points. Check them first.
Usage and collaboration data show who uses AI and who works with whom, not whether the work was good. No single tool covers everything. Ticketing and workflow software holds approval times. Engineering platforms such as DX, LinearB, and Faros cover developer productivity and delivery. Microsoft Viva Insights covers Microsoft 365 collaboration data. People analytics platforms such as Worklytics combine AI usage and collaboration data from many connected tools. Whichever you pick, know the limits:
Pick a workflow that passes through several teams, such as employee onboarding. Write down each step's owner and wait, remove steps that serve no purpose, and assign each task to AI, to AI with a person, or to a person. Rebuild with fewer handoffs and one named owner, then compare with your starting numbers.
Automation applies technology to the steps you already have. Redesign changes the steps first: it removes, merges, or reorders them before you decide what to automate. CIO's Nick Colisto puts it as redesign first, automate second.
AI does or helps with individual tasks, so the slow part shifts to the steps where work is handed on, reviewed, and approved. In a 2025 Faros AI study, teams using AI heavily merged 98% more pull requests (proposed code changes), and review time rose 91%.
Redesign the workflow first, as Deloitte advises. Once it works, check roles and how many people report to each manager, and treat headcount as a separate decision.
Four kinds help: AI use by team and role, who works with whom, when work enters and leaves each step, and quality and business results. A platform such as Worklytics can supply the first two, but not approval waits or quality.
If you want to see which teams use AI and which teams work together before you redesign anything, you can book a demo. Worklytics shows this from your own data at team level, without reading message content. If you already have these numbers from other tools, the six signals work just as well.
