| Last Week in HR AI: Issue #14 | Week of September 28, 2026 |
Your agent problem isn't the model. Amazon shelved 9 in 10.
AWS says nearly 90% of Amazon's early agent prototypes never reached production. The fixes were structural.
|
1 thenextweb.com2 thenextweb.com3 thenextweb.com
Ikona's Take on the past week (September 28, 2026) in HR + AI
|
IO
|
Ian O'Keefe
Co-founder & CEO, Ikona Analytics
|
Amazon says nearly 90% of its early AI agent prototypes never reached production. Industry-wide, 17% of organizations have agents deployed and 7% are measuring what those agents returned. None of that is a model problem. What separates the agents that ship from the ones that die in a demo is four foundations underneath them: captured tacit knowledge, governed data, process fixed before AI is layered on top, and an integrated stack rather than another purchase.
Rockwell's maintenance copilot is the clearest proof this week. The win (33% less downtime, onboarding cut from nine months to three) came from encoding what veteran engineers knew how to do, not from the Azure endpoint behind it. Compare that with Booking.com, which built AI tooling it now regrets because the frontier labs shipped the same capability months later, and Diageo's blunt read that AI layered onto unchanged process simply failed. Meanwhile the average company runs about 101 apps, and HR tech dissatisfaction jumped from 28% to 39% in a year. Deloitte's warning lands: agents pointed at an unintegrated stack just process the mess faster.
So the question I would put to your team is not which agent to pilot. It is who owns the capture of what your best HR people know, and whether you can name that person. And keep Gartner's Australian numbers in view: no agent roadmap survives managers your people are leaving. Reply with where your foundation is thinnest. We read every one.
| The six to read | The stories that matter for the Office of HR — with our take on each. |
Speaking at HumanX Amsterdam, AWS vice president Swami Sivasubramanian said close to 90% of Amazon's earliest AI agent prototypes never made it to production. He cited industry figures showing 17% of organizations have successfully deployed agents and only 7% are measuring return. Amazon's response was to consolidate onto Bedrock and AgentCore, add deterministic guardrails through the Strands framework, and let internally grown tools such as Kiro spread once measurement caught their traction.
Rockwell Automation's Singapore factory has used a GenAI maintenance copilot since October 2025 to help technicians diagnose machine errors, drawing on veteran engineers' tacit troubleshooting knowledge, manuals, and software data via Azure OpenAI. The company reports 33% less machine downtime, roughly 25% lower servicing costs, and onboarding time cut from nine months to three. The site holds a World Economic Forum lighthouse designation and plans to extend the tool to Ohio, Poland, and Mexico plants.
Also covered by: industryweek.com
Also at HumanX Amsterdam, Booking.com's James Waters said the company built AI orchestration tooling it now wishes it had bought, because frontier model providers shipped equivalent capability off the shelf shortly after. Diageo's Susan Jones said her early AI efforts failed when layered onto unchanged processes, and that process and data work had to come first. Waters pointed to Booking.com research showing 89% of travelers want AI for research while only 6% trust it to make the decision.
|
From Ikona
Your best HR people know things no system has captured
We diagnose how work and knowledge actually move through your HR function, then capture the judgment sitting in people's heads. What comes out is a structured knowledge store an AI system can stand on. Get in touch → |
Capital One's VP of enterprise AI, Rashmi Shetty, described building the bank's agentic capability on a governed data foundation with a platform-first architecture. She argues agents should be treated as end-to-end systems that need runtime controls, observability, evaluations, and human review on high-risk actions rather than as standalone models. Deployed use cases include a multiagent Chat Concierge for car buying, live for two years, plus internal customer-service agents.
Also covered by: VentureBeat
Gartner's Global Talent Monitor found active job-seeking among Australian workers rose to 26.5% in Q2 2026 from 19.4% at the end of 2025, while intent to stay dropped to 34%. Scaled nationally, that equates to roughly 3.8 million workers exploring new roles. Gartner analyst Neal Woolrich characterizes the movement as a market recalibration rather than a boom and advises employers to shift from passive, fear-based retention toward manager quality, career pathways, and flexibility.
Eagle Hill Consulting's HR Technology and Service Delivery Survey, fielded by Ipsos among 200 US HR professionals, finds that most respondents believe AI and automation improve service delivery. Even so, more than half still spend at least half of their week on routine administrative work. Nearly all say they work across multiple systems to finish a single task, many re-enter the same data, and most depend on manual workarounds. The report concludes that adopting technology has not resolved fragmented HR workflows, unclear ownership, or gaps in AI governance.
The argument is that HR keeps absorbing operational work because of how the function is designed, not because its people lack capability, and that the fix is an operating model change rather than another tool.
Of everything this week, the figure I keep returning to is Rockwell's: onboarding down from nine months to three. That did not come from a model. It came from writing down what the veteran technicians on that floor already knew, while they were still there to ask. If your agent roadmap has nobody assigned to that job, it is a roadmap to a prototype.
If your team is wrestling with workforce intelligence, AI readiness, or a transformation that has to land, we'd love to help. Reach out and tell us what you're working on.
Get in touch →Last Week in HR AI
Practitioner perspectives and diagnostic frameworks, delivered the week they publish. Be the first to know when new research drops.
No spam, ever. Unsubscribe anytime.