| Last Week in HR AI: Issue #18 | Week of October 26, 2026 |
Your AI can't redesign HR work nobody has written down
One bank had engineers shadow bankers first. Select loans got ready to close in 7 days, not 30 to 60.
|
1 pymnts.com2 jp.ibtimes.com3 beckershospitalreview.com4 Anthropic
Ikona's Take on the past week (October 26, 2026) in HR + AI
|
IO
|
Ian O'Keefe
Co-founder & CEO, Ikona Analytics
|
Most HR functions I talk to are trying to buy AI results without building anything for the AI to stand on. This week's through-line: the organizations getting strong returns from AI did the unglamorous work first, redesigning the workflow, building the data and skills platforms, and setting a real bar for what counts as working. Wharton's fourth annual adoption study names the pattern. Positive returns are now near-universal, strong ones are not, and the strong ones track with redesigned workflows, a standardized platform, and metrics beyond time saved.
The practitioners in this issue show the sequence. Vanderlande's people analytics lead spent two years putting historical HR data on a shared platform before adding AI. Now vacancy data ties to team costs across countries, and an assistant answers peer salary requests for pay transparency work. NTT built a skills platform covering about 165,000 employees before launching internal job scouting. Customers Bank had engineers shadow bankers on real workflows before building agents on the bank's own data. Nebraska Medicine stopped a cancer registry pilot at just over 80% accuracy against a 95% bar. That is the bar doing its job, and I'd want every steering committee able to make that call.
The harder implication comes from Anthropic, which reports agents now logging more hours than its employees, with human attention as the bottleneck. When that reaches your function, the scarce input is the knowledge nobody wrote down: how the process actually runs, where judgment sits, and what good looks like. You can't redesign a workflow or set a bar you can't describe. Start there.
Reply and tell me what you're seeing. I'd welcome the conversation.
| The six to read | The stories that matter for the Office of HR — with our take on each. |
Wharton Human-AI Research and GBK Collective published their fourth annual study of enterprise AI adoption. AI now sits inside core operations at the companies surveyed, and nearly all report some positive return, though far fewer report strong returns. The strong performers share several traits: workflows redesigned around AI, a dedicated chief AI officer, a standardized platform, success metrics that go past time saved, and agents given real authority. This year's edition also compares, for the first time, how leaders and entry- and mid-level employees line up on AI.
Also covered by: SHRM, marketscale.com
Alex Hellemons, who leads people analytics at Vanderlande, told CHRO.nl his team spent two years loading historical HR data onto a shared Azure data platform before adding AI. Vacancy data is now linked to team costs across countries, and recruitment decisions are steered by that view. A Databricks Genie assistant handles peer salary requests for equal pay work connected to the European pay transparency directive, and every delivery ships with its own Genie workspace. Hellemons recommends starting even when the data is imperfect.
NTT Group will launch Job Scout in October 2026, letting 66 group companies contact roughly 100,000 employees directly about internal openings. Matching relies on the skills, experience, and preferences employees register, drawing on a skills data platform NTT adopted in July 2026 that covers about 165,000 people. The new tool runs alongside the existing Job Board, where employees apply on their own initiative. That channel produced about 2,100 transfers from around 4,000 applications over three years.
|
From Ikona
Your AI needs the groundwork nobody wrote down
We diagnose how work and knowledge actually flow through your HR function and capture the judgment that lives in people's heads. Then we structure it into a knowledge asset your AI programs can stand on, and you keep it. Tell us where yours stands → |
Customers Bank reports that a pilot of its multi-agent underwriting process brought select commercial loans to a ready-to-close state in 7 days, compared with a typical 30 to 60. Parent company Customers Bancorp is scaling the approach and doubled its deployed agents in the third quarter. The bank cites 76,000 hours saved, based on employee estimates, and more than $15 million in annualized capacity from roles it did not backfill. Its engineers shadow bankers on live workflows before building agents on the bank's own data.
Nebraska Medicine piloted an AI tool to abstract data for its cancer registry, which handles about 6,000 cases a year at roughly 2.5 hours per case. The tool's accuracy reached just over 80%, short of the 95% threshold set for unsupervised use. A fallback mode that generated summaries performed well, but nurses still went back to the source charts, so little time was saved. The executive AI steering committee ended the pilot after two months, and the project lead noted that summarizing differs from automating a workflow.
Anthropic published two reports. One documents how AI agents have reshaped its own workforce, including a smaller share of junior and technical hires, agents logging more hours than employees by March 2026, and human attention emerging as the bottleneck. The other sets staff expectations about AI against a Morning Consult survey of almost 11,000 US adults. Anthropic concludes the difference stems from how much hands-on experience each group has with AI capability rather than from competing economic views.
Ikona's argument is that agents will not message "Rob in Payroll" to ask what a field means, so the meaning your analysts carried in their heads has to be written down, owned, and kept current. That is knowledge management work as much as data work, and it is the groundwork that lets a Vanderlande-style data platform support AI.
The line I keep coming back to is the headline on Vanderlande's story: "Eerst de basis stevig op orde, dan AI daadwerkelijk inzetten," roughly "first get the foundation solidly in order, then actually put AI to work." What makes it useful is that Alex Hellemons pairs it with advice to start with imperfect data, so the foundation is something you begin this quarter, not a precondition you wait on. For Monday, I would try a small version of what Nebraska Medicine did. Pick the one workflow your team most wants an agent for, and before anyone sees a demo, write down two numbers: the accuracy at which you would let it run unsupervised, and the accuracy at which you would stop.
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.