AI at work: hype, hard costs, and what HR must do next
AI headlines promise frictionless automation. Peter Cappelli, the George W. Taylor professor of management at Wharton, has a simpler take: possibility isn't the same as practicality. We've seen this before with driverless trucks-great demos, messy reality.
That gap between "can" and "will" is where HR earns its keep. The real value isn't in pilots or press releases. It's in job design, workflow mapping, and change that sticks.
What the Ricoh case actually shows
Ricoh tried to automate claims processing-the sort of repetitive, rules-heavy work AI should crush. It took a year, a team of six (including three pricey consultants), and about $500,000 just to stand it up. Early runs? Large language models performed the work at roughly three times the cost of humans doing it manually.
After optimization, ongoing AI fees settled around $200,000 per month-still more than the prior payroll for the task. Headcount moved from 44 to 39, not a mass layoff. Why? Exceptions need judgment, context, and follow-up. Humans still close the loop.
The payoff: that unit is now on track for 3x productivity. Ricoh says the project reached break-even in under a year and delivered around a 15% total cost reduction without relying on big cuts to staff. Work didn't vanish-it shifted to exception handling, quality control, and customer service.
Success stories take time-and a lot of work
Cappelli's Accenture partnership looked at Mastercard, Royal Bank of Scotland, and Jabil. They're winning, but slowly. Productivity improves; headcount doesn't crater. The overlooked factor: the workload to make AI useful is huge.
That's consistent with an influential MIT finding that the vast majority of generative AI pilots haven't shown meaningful returns. The tools are capable. Organizations aren't ready-yet.
Why "AI shame" is warping decisions
Investors love AI narratives. Boards expect movement. The Harris Poll found that most CEOs fear for their jobs if they can't show AI progress, and a large chunk admit to performative adoption that looks good but doesn't deliver real value. Some companies even float "phantom layoff" chatter to juice a stock pop.
As CFOs digest the true cost of standing up AI-data cleanup, workflow redesign, vendor bills, security, compliance-expect a sobering reset. This is organization change, not an app install.
What HR leaders should do this quarter
- Map the work, not the roles: Break jobs into tasks. Label which are rules-based, judgment-heavy, customer-facing, or exception-driven. AI fits tasks, not titles.
- Redesign workflows: Build human-in-the-loop steps for exceptions, risk, and quality. Define escalation paths and SLAs so work doesn't stall.
- Set realistic ROI math: Model total cost of ownership: vendor fees, tokens/API, integration, data prep, consultants, security, and support. Compare to current payroll and error costs.
- Pilot with purpose: Pick one high-volume process with clear ground truth and measurable outcomes. Time-box the pilot. Define success metrics before you start.
- Create new roles and upskill: Workflow designer, data steward, AI product owner, prompt/process engineer, QA lead. Train managers to manage hybrid teams (people + agents).
- Codify human-in-the-loop: Who approves what? What triggers manual review? What gets auto-processed? Put it in policy, not tribal knowledge.
- Own change management: Communicate early. Explain how jobs will change, not just "AI is coming." Offer training paths and credible timelines.
- Build governance: Usage policies, data retention, bias testing, audit trails, model updates, and vendor risk reviews. Document everything.
- Measure what matters: Throughput, error rate, rework, cycle time, customer impact, and cost per transaction. Publish the scorecard monthly.
- Plan redeployment, not cuts: As capacity expands, move people to higher-value work: exceptions, customer care, and continuous improvement.
Hiring and skills: what changes in practice
- Process intelligence: People who can untangle messy workflows and standardize inputs are more valuable than "prompt wizards."
- Data quality and ops: The model is only as good as the inputs. Hire/assign ownership for data definitions, validation, and feedback loops.
- Compliance-first mindset: Privacy, IP, and auditability must be designed in. HR should co-own this with Legal and Security.
- Manager fluency: Train managers to set rules for AI use, review outputs, coach for exception handling, and reset performance expectations.
Budget and timeline reality
Expect a heavy lift in year one: discovery, data prep, vendor selection, integration, and redesign. Costs may exceed the current manual model during ramp. That's normal.
By year two, productivity gains should start compounding-if you've nailed process, governance, and training. Treat "headcount reduction" as a lagging benefit, not the primary lever.
Questions HR should put on the table now
- Which end-to-end process are we standardizing before we automate any of it?
- What's our total cost model-including data, security, and maintenance-not just licenses?
- Where does human approval sit, and who owns exceptions?
- What roles will we redesign, and how will we reskill affected employees?
- What's our success scorecard for quarter one and quarter two?
The bottom line
AI can make teams meaningfully more productive. It rarely drops in as a plug-and-play headcount killer. The hard work is organizational-your work.
If you center task design, human-in-the-loop controls, and credible ROI math, you'll avoid the hype trap and build something that lasts.
Level up your team's AI fluency
Want a structured way to build practical skills across roles? Explore role-based AI learning paths and certifications here:
Your membership also unlocks:
