.png)
The HRIS team saved 1,000 hours, but they did not use generative AI. They simply documented scattered work and automated repetitive tasks first. The lesson is straightforward. Before choosing an AI tool, you need to see what keeps repeating and why.
The real meaning of 1,000 hours is not “AI ROI”
First, read the number accurately. In an HRBench podcast, HRIS leader Dani DeMaio explained that, at a previous employer, the team automated manual work and tracked the time saved. At performance-review time, this let her report to leadership that “the HRIS team saved 1,000 productivity hours.” However, the company size, measurement period, list of automations, and formula were not disclosed. So 1,000 hours should be seen not as a generalizable benchmark, but as one practitioner's self-reported tally.
Do not use 1,000 hours directly as ROI
Time saved is not automatically cash saved. You also need to confirm whether that freed-up time was actually reassigned to other work and how many hours went into automation maintenance and exception handling. The important part of this example is not the size of the number, but that it recorded time before and after automation.
The starting point was not “Which AI should we buy?” The team documented in detail what work people performed, how many people were involved, and how long it took, then identified the biggest time drains. They did not add technology simply because one person complained often; they checked whether the work recurred across multiple people.
This approach also fits how HR systems are structured. Workday defines a business process as a collection of tasks that people initiate, process, and complete to achieve a goal. Each step has assigned roles, security rules, conditions, and completion criteria. In other words, the target for automation is not a vague “recruiting task,” but a specific flow from input → validation → approval → notification → completion.
If a workflow exists only in memory or verbal explanations, AI will struggle to work reliably too. Process mining, as Microsoft describes it, uses event data from system records to visualize the actual path and find bottlenecks, rework, automation opportunities, and KPIs. Because the process a person believes exists may differ from the path actually executed, the first step is creating observable records.
Automation and AI solve different problems
For work where the correct answer is fixed by rules, conventional automation comes first. AI can be considered where input formats keep changing or the meaning of text must be interpreted. Without distinguishing the two, you may use expensive AI for a simple conditional statement or hand a low-error-tolerance decision to a probabilistic model.
| Decision question | Best suited to rule-based automation | Work where AI may be considered |
|---|---|---|
| Is the input consistent? | Arrives through fixed fields and codes | Formats vary, as with emails, documents, and free text |
| Can the correct answer be written as rules? | Approval routing by work location, deadline reminders | Classifying inquiry intent, summarizing documents, drafting |
| Can errors be spotted immediately? | Required values, formats, and duplicates can be validated | Quality judgments vary by context |
| What is the impact of failure? | Rerun it or route the exception to the responsible person | Strong review safeguards are needed when it significantly affects people, as in hiring or evaluations |
For example, once new-hire information is approved, a flow that creates an IT ticket, notifies the equipment owner, and records completion status has clear conditions and owners. Workday processes can also be designed to route tasks by role, apply security and business rules, and run steps based on conditions. IBM likewise describes HR automation as connecting HRIS, payroll, benefits, and applicant tracking systems to synchronize data flows.
By contrast, tasks such as classifying employee inquiries by topic or drafting policy documents may benefit from AI because the wording varies every time. Even then, that does not mean AI should make the final decision. The NIST AI Risk Management Framework recommends documenting the specific task AI will support, the model's limitations, how people will oversee its results, the expected benefits, and the cost of errors.
Work that directly affects people's opportunities, especially hiring, calls for even more care. The UK government's responsible AI in recruitment guidance identifies risks such as bias, digital exclusion, and discriminatory advertising, and advises addressing impact assessments, data-protection review, transparency, and appeal mechanisms. Legal requirements differ by country, but the practical principle still holds: a single statement that a human approves AI recommendations does not by itself create responsible operations.
If you count only time saved, you will discover failure late
Your scorecard needs to include quality, exceptions, and the human experience alongside speed. Even if automation reduces average handling time, it is hard to call it an improvement if incorrect routing or missed access revocations increase.
In particular, you need a fixed baseline to compare “time saved.” For two weeks before automation, record volume and actual handling time for each task, then measure using the same definitions after rollout. Publishing a formula in advance, such as time saved = number of cases × reduction in time per case − exception-handling and maintenance time, leaves less room to inflate the numbers at performance review time.
At the AI stage, model fees and review time must be included as costs too. In a January–February 2026 survey of 1,342 HR and business leaders, CIPD found that organizations more confident in their AI-adoption capability reported more performance improvement, but this is an association in self-reported data, not causation. The more practical message from the survey is that tool access alone is not enough; established capabilities such as work redesign and employee involvement must be put to use in real change.
A four-step plan to start this week
- Step 1: Record the actual path of one task for two weeks
Choose just one high-volume process among onboarding, employee changes, leave, or offboarding. In a spreadsheet, record start time, completion time, owner, source system, approval wait, reason for rework, and whether an exception occurred. If you use Workday, you can check current steps and roles in the Business Process Definitions report and View Diagram. In a Microsoft environment, you can review event-log-based flows in Power Automate under Process mining > Processes. - Step 2: Automate one segment that can be expressed as a rule
Choose a step you can test in one sentence, such as “If condition A is met, route it to owner B.” Examples include required-value validation, approval routing, deadline reminders, and ticket creation. Before putting it into operation, test normal, missing, duplicate, canceled, and unavailable-owner cases, and designate whose queue it returns to if it fails. - Step 3: Experiment with AI only for variable segments automation cannot solve
Start with work where inputs vary but people can review the output, such as classifying employee inquiries or drafting policies. Put the data used, prohibited data, reviewer, approval criteria, log-retention period, and stop conditions on one page. It is safer not to run hiring, evaluation, or compensation decisions automatically without impact assessments and appeal procedures. - Step 4: Decide whether to expand, revise, or stop after 30 days
Compare against the baseline for handling time, error rate, exception rate, and manual-intervention time. If speed improved but errors or review burden increased, narrow the scope. Even when results are confirmed, do not connect the next tasks all at once; expand one step at a time using the same scorecard.
Explain to employees first that this is not “documentation to eliminate jobs”
In the seed example, a team member asked whether the request to record work was an attempt to eliminate their job with technology. Share the purpose of the records, the criteria for choosing automation candidates, and how the changed time will be used, then involve the people doing the work in design and review.
If you want to dig deeper
Automate Before You Even Think About AI: An HR Leader's Playbook — The original article covering the 1,000-hour example, work mapping, and the limits of AI use in HR. hrbench.com
What Is HR Automation? — An overview of HR automation and how HRIS, payroll, and benefits systems are integrated. ibm.com
Overview of process mining — Explains how to use event logs to identify actual work paths, bottlenecks, rework, and KPIs. learn.microsoft.com
Concept: Business Processes — Shows how Workday processes are structured using steps, responsible roles, conditions, and security policies. doc.workday.com
AI RMF Core — A NIST framework for designing AI use cases, error costs, human oversight, and operational accountability. airc.nist.gov
Responsible AI in Recruitment — UK government guidance that frames bias, data protection, impact assessments, and appeals in recruitment AI as practical questions. gov.uk
