The most common question I get from leaders considering AI is some version of the same thing: Where do we start? They have heard the case for productivity gains, watched competitors announce pilot programs, and feel pressure to act. What they often lack is a method for deciding where AI will actually create value in their organization — and where it will simply create motion.
My answer is almost always the same. Start with the work, not the tool. Specifically, start with task analysis: a structured look at what people actually do, how long it takes, and where the friction lives. Done well, this exercise reframes the entire conversation from “deploying AI” to something more useful — redesigning how work gets done.
This Is Organizational Change, Not a Technology Project
The framing matters enormously. When AI is positioned as a technology rollout, it follows the familiar arc of vendor selection, pilot, and deployment, with success measured by adoption rates and seat counts. When it is positioned as work optimization, it follows a different arc: understanding the job, identifying friction, redesigning the task, and supporting the people doing it. The same tools may end up in play, but the outcomes are not comparable.
The Job Demands–Resources (JD-R) model, originally introduced by Demerouti, Bakker, Nachreiner, and Schaufeli in 2001 and further developed by Bakker and Demerouti in subsequent years, offers a useful lens here. The model holds that every role consists of demands (the cognitive, emotional, and physical effort required) and resources (the supports that help employees meet those demands and stay engaged). When demands chronically outstrip resources, you get burnout and disengagement. When resources are well-matched to demands, you get motivation and performance. AI, properly applied, is a new category of job resource — one that can absorb high-volume, low-judgment demands and free human capacity for the work that requires discernment, relationship, and creativity.
That framing changes how managers should think about implementation. The question is not “what can this tool do?” It is “what demands are wearing my people down, and where can a resource — human, procedural, or AI — relieve them?” When leaders skip this step and lead with technology, they tend to automate the wrong things, miss the friction that actually matters, and leave employees feeling surveilled rather than supported.
This is fundamentally an organizational change problem. The work itself is being redesigned. Roles will shift. Skills will need to develop. Trust must be built. None of that is a software deployment.
What Task Analysis Is — and What It Reveals
Task analysis is a method for decomposing a job into its constituent activities so they can be examined individually. It originates in the human factors work of John Annett and Keith Duncan, whose 1967 paper on Hierarchical Task Analysis (HTA) established a structured way to break a role’s overall goal into sub-goals, operations, and the plans that govern their sequence. HTA was developed to improve training and ergonomics, but the underlying logic — that you cannot improve what you have not described — applies cleanly to AI redesign.
A modern complement to HTA comes from Erik Brynjolfsson and Tom Mitchell, who developed a Suitability for Machine Learning (SML) rubric in a 2017 paper and, with Daniel Rock, applied it to 18,156 tasks in the U.S. Department of Labor’s O*NET database in their 2018 study. Their finding is the single most important data point I share with leaders entering this conversation: most occupations contain at least some tasks suitable for machine learning, but few occupations are fully automatable, and realizing the value of AI almost always requires redesigning task content within a role rather than eliminating the role itself (Brynjolfsson, Mitchell, and Rock, 2018).
That conclusion should orient the entire effort. We are not looking for jobs to replace. We are looking for tasks to redesign.
A practical task analysis blends both traditions. From HTA, take the discipline of decomposing a role into a hierarchy of goals and observable activities. From the Brynjolfsson and Mitchell rubric, take the discipline of scoring each task against criteria that predict whether AI can meaningfully help: Is the task well-defined? Are inputs and outputs clear? Is there sufficient data or precedent? Does it require tacit judgment, physical presence, or relational nuance that machines handle poorly?
What the analysis reveals is almost always more nuanced than leaders expect. Roles that seem ripe for automation turn out to contain a few critical judgment-intensive tasks that anchor their value. Roles that seem untouchable turn out to contain hours of weekly administrative drag that AI can absorb almost immediately. The map is rarely what the executive team predicted.
How to Actually Do It
A serviceable task analysis does not require a consulting engagement. It requires structure, honesty, and a few hours with the people who do the work. Here is the approach I use:
Start with the role’s purpose. Write a single sentence describing what the role exists to accomplish. This becomes the goal at the top of the hierarchy and the standard against which every task is judged.
Decompose into tasks and sub-tasks. Working with the person in the role, list the activities they perform across a typical week or month. Push for specificity — “respond to client questions” is not a task; “triage incoming client emails, draft initial responses, and route complex issues to specialists” is three. Aim for 20 to 40 discrete tasks per role.
Add time and effort estimates. For each task, estimate hours per week (or per cycle) and rate the cognitive effort involved on a simple scale. This is the data layer that will drive prioritization. Self-reported estimates are imperfect but directionally accurate; if precision matters, sample with a time-tracking exercise for two weeks.
Score for AI suitability. For each task, ask the questions drawn from the Brynjolfsson and Mitchell rubric. Is the input structured? Is the output verifiable? Does the task draw on clear precedent? Does it require empathy, physical presence, or judgment under ambiguity? Tasks that score high on suitability and high on time consumption are your priority candidates.
Decide between automation and augmentation. Not every suitable task should be fully automated. Some are better augmented — AI drafts, human reviews. Compliance-sensitive work in healthcare and financial services almost always falls into this category. The decision should be explicit and documented.
Consider how this plays out in a community health center. A typical care coordinator role contains client intake documentation, insurance verification, appointment scheduling, follow-up outreach, case notes, referral coordination, and direct patient conversations. A task analysis often reveals that 40 to 50 percent of the role is documentation and administrative coordination — exactly the work that pulls attention away from patients. AI-drafted intake summaries, automated insurance verification, and structured note-taking assistance can return meaningful hours to direct patient care without touching the relational core of the role. The same logic applies in nonprofits, where program managers often spend more time on grant reporting and donor administration than on program design. The pattern is consistent: the highest-value tasks are usually under-resourced, and the highest-volume tasks are usually under-examined.
Bringing Employees In as Contributors
The final piece is the one most often mishandled. Task analysis works only when the people doing the work are full participants in it — not subjects of it.
This is partly practical. Managers do not actually know how the work gets done at the level of detail required. The person in the role does. Skipping their participation produces a bad map.
It is also partly cultural. When employees experience task analysis as something done to them, they reasonably suspect it is the prelude to a layoff. When they experience it as something done with them, with the explicit goal of relieving demands and freeing capacity for higher-value work, they tend to surface the friction faster and more honestly than any external assessment ever could. They know which reports nobody reads. They know which approvals are rubber-stamped. They know where the system has been quietly failing for years.
The framing I use with teams is direct. We are not doing this to find out who to replace. We are doing this to find out what to fix. Your job is to help us see the work clearly. The tasks that drain you are the ones we want to redesign first. That promise has to be backed by behavior — by actually redesigning the work the team flagged, by reinvesting freed capacity into work the team finds meaningful, and by being transparent about what changes when AI enters the workflow.
Done this way, AI implementation stops looking like a technology rollout and starts looking like what it actually is: a once-in-a-generation opportunity to reexamine how work is structured and to build roles that are more sustainable, more valuable, and more human. The starting point is not the tool. It is the task.
References
Annett, J., & Duncan, K. D. (1967). Task analysis and training design. “Occupational Psychology”, 41, 211–221.
Brynjolfsson, E., & Mitchell, T. (2017). What can machine learning do? Workforce implications. “Science”, 358(6370), 1530–1534.
Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What can machines learn, and what does it mean for occupations and the economy? “AEA Papers and Proceedings”, 108, 43–47.
Demerouti, E., Bakker, A. B., Nachreiner, F., & Schaufeli, W. B. (2001). The job demands-resources model of burnout. “Journal of Applied Psychology”, 86(3), 499–512.