.png)
You need to propose AI adoption, but can’t see what to change first
Which HR task should get AI first: onboarding, employee changes, leave, or offboarding? Ask an HRIS owner and they all look repetitive and urgent. The step each person considers most painful will differ, too.
If you choose a chatbot or generative AI in this state, it’s easy to use an expensive model for simple conditionals or hand a probabilistic system a decision where a wrong answer would cause trouble. The first thing to do is record where work actually stops and how many times people touch it.
An HRIS leader featured by HRBench also surveyed the time each team spent and the number of people affected, then automated manual work first. They kept logging returned time in a spreadsheet and reported saving 1,000 productivity hours at performance review time—but AI was not used in that process.
1,000 hours is not a target; it is an example of a measurement habit
Do not treat this number as your team’s projected time savings or AI ROI. The source does not disclose company size, measurement period, the list of automated tasks, baseline time per item, or whether time for exception handling and maintenance was deducted. It is a self-reported example from one practitioner, so there is also insufficient basis for converting it into cash savings.
What you can take from it is the order of investigation. The practitioner did not only listen to one person’s repeated complaints; they looked at the time each task required alongside the number of people affected. They also explain that storing owners’ work documentation in a repository helped preserve operating knowledge tied to individuals when staff left or during corporate transactions. That effect was not quantified separately, however.
Time saved and cost saved are different numbers.
Even if processing time falls, it is not cash savings unless the returned time is actually reassigned to other work. You also need to subtract time spent again on error correction, exception handling, review, manual intervention, and maintenance.
Your first run: turn one completed onboarding case into a process map
You do not need to investigate company-wide HR processes from the start. Choose one recently completed recurring task and begin with a small CSV that removes or minimizes personal data. The values and task names below are hypothetical examples meant to show the input format, not product performance or actual processing times.
- Decide on the log type first.
For an event log, where one row represents one event, prepare Case ID, Activity Name, and one occurrence timestamp. For an activity log that also analyzes each step’s duration, include start and end timestamps separately. Microsoft documentation requires Case ID and Activity Name, while distinguishing the timestamp structures of event and activity logs. - Save the hypothetical example below as a CSV.
case_id,activity_name,event_timestamp
ONB-024,New hire information received,2026-09-01T09:00:00+09:00
ONB-024,Manager approval,2026-09-01T13:40:00+09:00
ONB-024,IT equipment request,2026-09-01T14:05:00+09:00
In real work, reconcile HRIS status changes with approval histories and ticket records, and do not include personal data such as names or email addresses when it is unnecessary for analysis. A table that stores only the current status may not contain past activity history. - Check the environment and permissions, then sign in in your browser.
To start Process Mining, you need an appropriate license or trial, a Power Platform environment with a Microsoft Dataverse database configured, and the Environment Maker role in that environment. If you cannot select an environment to use or there is no Dataverse database, ask an administrator to configure the environment and assign the role. Once the conditions are met, sign in to Power Automate with your organization account and enter a process name under Start here on the Process Mining home page. No separate local installation is required. - Connect the CSV and map the columns.
Go to Choose data source → All categories → Text/CSV → Browse OneDrive, authenticate with OneDrive for Business, and open the CSV. Keep only the required period and activities, then map Case ID, Activity, and the timestamp columns appropriate for the log type on the next screen, and select Save and analyze. If OneDrive for Business is unavailable, you can review the Blank table path in the official documentation. - Confirm the first success through activity nodes and paths.
If the process map in Analytics shows the three activity nodes New hire information received, Manager approval, IT equipment request, the transition New hire information received → Manager approval → IT equipment request, and a total case count of 1, the hypothetical log has been analyzed. The process map aggregates and displays paths between activities; it is not a screen for expanding events under an individual Case ID. If you need to verify ONB-024 itself, find it in a separate case list or detail view. If analysis returns an error, first check that Case ID and Activity Name are Text fields and timestamps are Date/Time fields.
Process mining is an option for finding actual flows, bottlenecks, rework, and automation candidates from recorded event data. It does not replace initial work documentation. If you do not have stable Case IDs or historical events, return to collecting completed cases in a spreadsheet rather than adding more tool configuration.
Divide the work found on the map into three stages
Automate the portions of the recorded path that can be fixed as rules first, then pass only the portions that still require interpretation to AI candidates. IBM’s general HR automation process also describes the sequence as investigating and documenting manual work, selecting tools, and connecting existing HRIS, payroll, benefits, and applicant tracking systems.
- Stage 1: Fix the actual path and baseline.
Across several completed cases, record volumes, elapsed time between steps, manual intervention time, rework reasons, and exceptions using the same definitions. When an owner’s memory and system records differ, retain both and determine why the difference occurred. - Stage 2: Automate portions you can test in one rule sentence.
Write the input, condition, outcome, and failure queue, such as: “If an approved new-hire record has all required equipment fields, create a ticket in the designated IT queue and record the processing status.” Define expected results for normal cases, missing required values, duplicates, cancellations, and unavailable owners, then implement them in the current HRIS or automation tool. - Stage 3: Evaluate only the portions AI cannot fix as rules.
Start with work where input wording varies but a person can review the result, such as classifying the intent of inquiry messages or drafting policy documents. Document the tasks AI will support, knowledge limits, scope of use, review owner, stop conditions, expected benefits, and cost of errors. The NIST AI RMF also guides organizations to define and document such items.
| Decision question | When rule-based automation fits | When AI may be considered |
|---|---|---|
| Input | Defined fields, codes, and status values | Email, documents, and free text |
| Correct answer | Conditions and outcomes can be listed in advance | Classification, summaries, and drafts vary by context |
| Check | Immediately verify required values, duplicates, and format | A person reviews context and evidence |
| Example | Routing approvals, notifications, ticket creation | Classifying inquiry intent, drafting policies |
| Failure response | Rerun or route to an owner queue | Hold approval, human review, or stop use |
AI that affects people’s opportunities, such as recruitment AI, should be separated from general productivity automation. UK government guidance recommends defining purpose and functionality before purchase; reviewing data protection, effective human oversight, accessibility, and bias risks; and piloting with actual users and affected groups. Because this guidance is for the UK context, actual deployment must also separately check applicable labor, anti-discrimination, and privacy laws.
Calculate success using both speed and the cost of failure
The fact that automation ran once does not by itself mean it succeeded. With the same scope, measurement period, and start-and-completion definitions, processing and wait times must decrease while errors, rework, exceptions, and manual intervention remain within acceptable limits.
For example, you can define volume processed × reduction in processing time per item − exception handling, review, and maintenance time as net returned time. This formula is not an official standard; it is a conservative example a team can agree on before applying it. Rather than making the HRBench example’s 1,000 hours a target, it is more important to make your organization’s definitions and source data recheckable.
Share the purpose first when asking for work records.
In the HRBench case, team members also asked whether documentation was intended to eliminate their jobs. Explain first what will be considered for automation, where returned time will go, and how the people doing the work will participate in design and review; only then can you capture hidden exceptions as well.
If you want to go deeper
Automate Before You Even Think About AI: An HR Leader's Playbook — The original article for understanding the context of the 1,000-hour example and the sequence of work investigation and documentation. hrbench.com
Prepare processes and data — Learn about the required columns for event and activity logs, CSV connection, and the procedure for starting analysis. learn.microsoft.com
Overview of process mining and task mining in Power Automate — Check prerequisites before getting started, including licensing, an environment with a Dataverse database, and the Environment Maker role. learn.microsoft.com
AI RMF Core — A voluntary framework to consult when documenting AI task scope, limits, human oversight, benefits, and the cost of errors. airc.nist.gov
