Before AI Reorganizes the Workforce, Map the Work
AI workforce planning should begin with evidence about the work—not assumptions attached to job titles.
A title such as buyer, customer-service representative, production planner, marketing specialist, or warehouse lead rarely explains the actual mix of judgment, system activity, customer interaction, exception handling, and physical work inside the role. Automating or reorganizing from the title alone can remove the person who quietly owns the exceptions while leaving the underlying process unchanged.
SAP’s agreement to acquire TechWolf puts task-level workforce data at the center of the AI transformation conversation. The operating lesson is useful well beyond HR software: before deciding what AI should perform, assist, or leave with people, build a validated map of the work and its dependencies.
What changed
Documented fact: On October 6, SAP and TechWolf announced an agreement for SAP to acquire the Belgian workforce-intelligence company. The SAP announcement says TechWolf builds a continuously updated view of tasks, skills, and external labor-market information from connected HR and business systems.
Transaction status: SAP expects the deal to close in the fourth quarter of 2026, subject to customary closing conditions and regulatory approval. Terms were not disclosed. SAP’s current plan is for TechWolf to remain an independent entity and continue serving SAP and non-SAP customers after closing.
Independent reporting: Reuters reported on October 6 that SAP plans to use TechWolf in workforce software for hiring, retraining, and reorganization. Reuters also noted that SAP says the companies already share customers but did not provide customer names or performance figures.
Vendor claims, not independently established outcomes: SAP says the combined data will make its AI assistant more intelligent, reduce token use, and lower the cost of workforce agents. SAP also says joint customers are seeing measurable results. The public announcement does not provide the evaluation method, baseline, sample, or quantified results needed to verify those claims.
Still unresolved: The sources reviewed do not explain how individual task and skill inferences will be validated, how employees can correct inaccurate records, which decisions may be automated, how long detailed work data will be retained, or how customers will manage regional employment and privacy requirements. Those implementation choices will determine whether the system becomes useful operating intelligence or an unreliable layer of employee scoring.
Why job titles are a weak automation plan
Most roles are bundles of unlike tasks. A customer-service representative may answer routine status questions, interpret product fit, recognize fraud, coordinate replacements, preserve a customer relationship, and document a product defect. A production planner may update schedules, resolve material shortages, negotiate priorities, recognize a quality constraint, and coordinate the response to a late supplier.
Some tasks are repetitive and rules-based. Some require judgment. Some are infrequent but operationally critical. Some exist only because two systems do not communicate. That last category should often be removed through business-systems integration instead of being preserved as a permanent human duty or handed to an AI agent.
A task map helps the business separate four different decisions:
- Eliminate: unnecessary work, duplicate entry, avoidable approvals, and broken handoffs.
- Automate: stable, measurable work with clear inputs, rules, and recovery paths.
- Assist: research, drafting, summarization, prioritization, and decision preparation where a qualified person remains responsible.
- Retain: work that depends on accountability, negotiation, safety judgment, physical execution, sensitive relationships, or unresolved exceptions.
This is a constructive path to adoption. It directs investment toward the work AI can improve while protecting the knowledge and ownership the operation still needs.
Controls for an evidence-based workforce map
1. Define the decision before collecting the data
State exactly what the analysis will inform: training priorities, process redesign, internal mobility, staffing forecasts, automation candidates, or role design. Do not collect a broad record of employee activity and decide its purpose later.
The decision statement should identify who may use the result, which actions are prohibited, how long the data remains relevant, and when a separate review is required. Workforce planning is not permission to reuse the same inferences for performance ratings, discipline, compensation, or termination.
2. Map tasks to process outcomes
Start with the process, not the person. For each workflow, document the trigger, inputs, systems, decisions, exceptions, handoffs, output, service level, failure modes, and accountable owner. Then map the tasks people perform inside it.
This matters in operations because the same task can have a different risk depending on where it sits. Copying an order number into a shipping system is not equivalent to approving a substitute product, releasing a work instruction, or authorizing a refund.
3. Preserve data lineage
Every inferred task or skill should point to its source: an approved job profile, system transaction, work item, training record, process document, supervisor validation, employee confirmation, or observed workflow. Record when the source was collected, how the inference was created, and how current it is.
Communication volume, application use, or ticket activity may show that work occurred. It does not automatically establish quality, difficulty, business value, or individual contribution.
4. Give people a correction path
Employees and process owners should be able to review meaningful task and skill records, identify missing work, challenge an incorrect inference, and add context. A person who resolves rare production or customer exceptions may look underutilized in a system that counts only routine transactions.
Corrections should be logged, reviewed by a named owner, and fed into model and taxonomy improvement. Disagreement should remain visible rather than being overwritten by the latest automated classification.
5. Use deterministic validation for important fields
Validate job codes, locations, certifications, required training, system permissions, shift assignments, process ownership, and current employment status against authoritative records. Use controlled task and skill taxonomies where decisions require consistency.
Do not allow a language model’s confidence statement to substitute for a required license, completed training, signed qualification, or approved role assignment.
6. Separate analysis from action
A workforce model may recommend reskilling, reassignment, hiring, or role redesign. It should not make those decisions by writing directly to HR, scheduling, access-control, or payroll systems.
Put human approval gates at consequential boundaries. The reviewer should see the recommendation, source evidence, confidence, disputed data, alternatives, operational impact, and the people who must be consulted.
7. Apply least privilege and data minimization
Task analysis often joins sensitive HR information with operational records. Limit each service account, model, analyst, and manager to the minimum fields and population needed for the approved purpose. Separate identifiable records from aggregate planning data when individual identity is unnecessary.
Exclude unrelated communications, health information, protected leave, personal content, and other sensitive attributes unless there is a documented, reviewed reason to process them. Restrict exports and bulk queries, and monitor unusual access.
8. Keep an auditable decision record
Log the data version, taxonomy, model and prompt version, rules, analyst changes, employee corrections, recommendation, approvers, final decision, and implementation date. Record what the business believed would improve and how it planned to measure the result.
That audit trail supports accountability without pretending the model made the management decision. People remain responsible for the policy, interpretation, approval, communication, and outcome.
9. Assign exception ownership
HR can own workforce policy and employee-record questions. Operations owns the accuracy of process and task maps. Functional leaders own required qualifications. IT and security own access and system integration. Legal or appropriate specialists should review higher-impact employment and privacy questions.
When the system detects an ambiguous role, conflicting skill evidence, or a task with no clear owner, route it to the appropriate function. Do not allow uncertainty to become a default recommendation to remove or reassign work.
10. Review vendor and model risk
A vendor and model risk review should cover source systems, retention, training use, regional processing, subcontractors, model changes, inference explainability, access controls, employee visibility, export, deletion, availability, and contract exit.
Test whether the system performs consistently across job families, locations, shifts, languages, employment types, and work that happens outside well-instrumented software. A clean dashboard can still reflect incomplete coverage.
11. Monitor decisions and downstream effects
Monitor correction rates, stale records, missing roles, disputed inferences, recommendation acceptance, exception volume, training completion, process performance, employee turnover, service quality, safety indicators, and unintended workload transfers.
A faster process is not automatically a better process if defects, escalations, overtime, or customer dissatisfaction move elsewhere.
Rollback and manual fallback
A workforce-data error can spread into training plans, staffing models, access assignments, schedules, and reorganizations. The incident-response process should identify who can suspend recommendations, freeze downstream writes, preserve affected records, locate decisions that relied on the bad data, notify the right owners, and restore the last validated task map.
The manual fallback is the existing management process: validated job and process documentation, supervisor and employee interviews, qualification records, and human review. A system outage or disputed model result should slow the decision—not erase due diligence.
A practical first implementation
- Choose one workflow: select a measurable process with clear pain, such as order entry, quote preparation, customer-service triage, purchasing follow-up, or production scheduling.
- Document the current work: map tasks, systems, decisions, exceptions, and owners with the people who actually perform them.
- Validate the evidence: compare system data with interviews, observation, and authoritative records.
- Classify the opportunity: eliminate, automate, assist, or retain each task—and document why.
- Design the future handoff: specify what AI does, what rules validate it, who reviews it, and who owns exceptions.
- Pilot with limited authority: use narrow data access and require approval before consequential changes.
- Measure the whole process: track cycle time, quality, rework, exceptions, customer impact, employee workload, and recovery.
- Expand only from evidence: scale the pattern after the business can show reliable improvement and a workable fallback.
For manufacturers and industrial distributors, this approach connects workforce planning with industrial automation. It prevents technology planning, staffing, process design, and training from becoming separate projects that optimize against different versions of the work.
The management question
Before using AI to redesign roles, ask: Can we show which tasks make up the process, which evidence supports the map, where judgment and exception ownership live, how people can correct the record, and who approves every consequential change?
If the answer is yes, workforce intelligence can help the business adopt AI with more precision and better training decisions. If the answer is no, the organization is reorganizing around assumptions.
