| Last Week in HR AI: Issue #13 | Week of September 21, 2026 |
72% blame the process, not the model, when AI fails
Average loss: $1.55M per company, and most of them bolted AI onto workflows that were already broken.
|
1 fst.net.au2 camunda.com3 hrtechedition.com4 UNLEASH
Ikona's Take on the past week (September 21, 2026) in HR + AI
|
IO
|
Ian O'Keefe
Co-founder & CEO, Ikona Analytics
|
Westpac now has 95% of its 35,000 employees using AI at least monthly, up from 69% a year ago. The number worth studying is not 95. It is 285: the systems the bank pulled into a single data platform before usage went vertical. That is the week's through-line. Adoption has outrun the plumbing almost everywhere, and the organizations reporting real gains are the ones that rebuilt the layer underneath first.
Camunda priced the alternative. 72% of organizations say broken processes caused AI initiatives to fail, at an average of $1.55 million in losses each, mostly because they bolted AI onto workflows that were already bad. Meta's agent design points at the fix HR keeps skipping: separate institutional knowledge from the reasoning that uses it, keep the knowledge in versioned files, and put a human checkpoint where judgment is genuinely ambiguous. Assessment time went from days to minutes with near-zero regressions. That is knowledge architecture, not model selection, and it is the most copyable thing in this issue.
Then the question stops being about tools and becomes a question about our function. Deloitte's 25,000-worker survey found 46% of UK workers on unsanctioned free tools and £958 million in AI your people bought for themselves. Tighter policy does not touch that. Faster procurement does, and procurement speed is a CHRO decision. Sharon Doherty at Lloyds gives the honest version: she would have stood up a dedicated AI team inside the People function a year earlier than she did.
What would you do with that year back? Reply and tell me, I read all of them.
| The six to read | The stories that matter for the Office of HR — with our take on each. |
Westpac reported that more than 70% of its 35,000 employees now use AI daily and 95% use it at least monthly, up from 69% monthly usage a year earlier. Executives cited productivity gains across banking operations, software development, and mortgage processing, along with faster customer migration off legacy platforms. The bank also described Adapt, an Azure-based data platform that unifies 285 source systems, and framed trusted data as the precondition for running AI at scale. Leaders discussed internal debates on AI risk and reported favorable token-cost economics on one migration project.
Camunda commissioned a survey of 1,000 process leaders and 5,000 employees across the US, UK, Germany, and France on the state of AI deployment. Some 72% of organizations said process-related problems have caused AI initiatives to fail, with average reported losses of $1.55 million per company. Respondents described bolting AI onto legacy workflows instead of redesigning those workflows first. The study also found governance gaps, employee workarounds, and a gap between how leaders and staff perceive productivity gains.
Meta published details of an internal agent architecture built to preserve expert judgment in a compliance domain, arguing the pattern generalizes to security, finance, and similar fields. The design keeps institutional knowledge in versioned text files, separate from the reasoning "recipes" that use it, and inserts human checkpoints where cases are genuinely ambiguous. Rather than retraining models, the system improves through verified edits that are regression-tested each cycle. Meta reports assessment time dropping from days to minutes with near-zero regressions.
|
From Ikona
The plumbing under your AI is undocumented
We diagnose how work, data, and decisions actually move through your HR function, then capture the judgment that lives in people's heads. The output is a structured knowledge store your AI can stand on, not a slide deck. Talk to us about yours → |
Sharon Doherty, chief people and places officer at Lloyds Banking Group, describes the bank's shift from AI experimentation toward proving customer and shareholder value. She outlines three cross-functional forums run with IT and finance, covering AI strategy, ethics, and employee experience, alongside continuous testing of multiple AI tools. Doherty says openly that she would have created a dedicated AI transformation team inside the People function roughly a year earlier than she did.
An HRTech Edition opinion piece draws on Deloitte UK's GenAI Workforce Survey of 25,000 respondents to argue that HR has lost control of AI rollout. The cited figures include 63% of UK workers using generative AI, 46% relying on unsanctioned free tools, and 17% paying for AI tools out of their own pockets, amounting to an estimated £958 million in worker-funded spend. The author contends that stricter policy will not change this behavior, and that only faster procurement of trained, licensed tools will.
Also covered by: fairplaytalks.com, HR Brew
At a panel during AGNTCon and MCPCon Amsterdam, protocol maintainers from OpenAI, Anthropic, Google, and AWS challenged the claim that AI is making core engineering skills obsolete. They argued that Deloitte and World Economic Forum findings on skills depreciation are being misread to justify expensive certification programs, while fundamentals such as systems thinking, testing, and protocol design remain essential. AWS's Clare Liguori likened today's agentic tooling gap to the early days of serverless computing, expecting shared frameworks to reduce the cognitive load developers currently carry.
The piece argues that HR AI initiatives stall because the foundations they depend on, data quality, process maturity, and governance, get defunded by the AI spending itself, and it offers a sequencing framework for building the base before building on top.
The tooling gap will close on its own. Clare Liguori's comparison to early serverless is the right read: shared frameworks will absorb most of today's complexity, which makes a large certification bill a payment against a problem with an expiry date. What does not close on its own is the 285-systems problem, or the year Doherty says she lost before putting a dedicated AI team inside People. One of those is a multi-year data program. The other is a decision you can make on a calendar this week.
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.