Why did KPMG retract its AI report?

KPMG pulled a published report on AI usage after it appeared to contain hallucinations, meaning AI-generated content that was factually unreliable or fabricated made it into the final document. It was covered by TechCrunch as a notable example of AI being used to produce analysis about AI without sufficient verification. The incident highlights the need for human review of any AI-generated output before it is published or shared professionally.

What does the Ardmore administration mean for UK construction subcontractors?

When a main contractor like Ardmore enters administration, subcontractors and suppliers with outstanding invoices typically become unsecured creditors, which means recovery is uncertain and often partial. The Mackoy administration in the same week signals continued financial fragility in the UK construction supply chain. Smaller businesses working in this sector should monitor payment timelines closely and not allow slow-pay situations to drift.

Do AI agents actually need clean data to work, or is that an oversell?

Clean, accessible, well-structured data is a genuine prerequisite for AI agents that perform reliably, not a sales pitch. As Xebia's global CTO explained this week, agentic AI scales on data quality. An agent built on scattered spreadsheets and inconsistent records will produce inconsistent results. The practical implication for any SME considering AI automation is that assessing your data foundation before committing to a build is not optional preparation, it is part of the build itself.

AI This Week: KPMG Pulled Its Own AI Report for Hallucinating

Key Takeaways

KPMG published and then withdrew a report on AI usage after it appeared to contain AI hallucinations, raising serious questions about content verification in professional firms.
Ardmore Construction Group and Hampshire-based groundworks firm Mackoy have both entered administration, adding to the picture of a cash-strained UK construction sector.
ONS data for April 2026 shows UK construction output growing at a near-standstill month-on-month.
AI agents fail in practice when they are not built on clean, accessible data, according to Xebia's global CTO, confirming what anyone who has tried to build one already knows.
Visa has connected its payment infrastructure to ChatGPT, allowing AI agents to complete retail purchases autonomously without human sign-off.

KPMG retracted its own AI report over hallucinations

This one landed awkwardly. KPMG produced a report covering AI usage, and then pulled it because it apparently contained hallucinations, meaning the AI fabricated or distorted information that made it into a published, professional document. TechCrunch covered it as a straightforward irony: a firm producing AI analysis that itself could not be trusted.

The practical lesson here is not that AI is useless for research or content. It is that AI-generated output needs a verification layer before it becomes anything official. For SMEs, that matters in two directions. First, if you are using AI to draft proposals, reports, or any document a client will see, someone with domain knowledge needs to check it. Second, if you are evaluating an AI vendor or tool and they hand you impressive-looking output with no clear process behind it, ask how it was verified.

We build systems with human checkpoints for exactly this reason. Automation handles the volume; people handle the judgement calls. If you want to understand how that looks in practice for your business, the AI automation audit checklist at /checklist is a good place to start.

Two construction firms hit administration in the same week

Ardmore Construction Group filed for administration following legal battles and remediation disputes, while Hampshire-based Mackoy Groundworks and Civil Engineering, a business turning over around £23 million, was sold through a pre-pack administration deal on 9 June. Insolvency firm Quantuma cited labour and material costs as primary causes of Mackoy's difficulties.

These are not isolated events. They fit into a broader pattern in UK construction where margin pressure, retentions, remediation liability, and slow payment cycles are combining to collapse businesses that looked stable from the outside. Crest Nicholson moved to secure a £15.7 million debt against Ardmore-linked property assets via interim charging orders before the administration filing, which tells you something about how quickly these situations can accelerate once legal action begins.

For smaller contractors, subcontractors, and trades businesses working in the supply chain, this matters directly. When a main contractor fails, outstanding invoices become unsecured debt. Cash flow monitoring and fast response to payment delays are not optional housekeeping. For businesses that rely on prompt follow-up and quote conversion to keep work moving through the pipeline, having gaps in your enquiry and lead handling is a risk you cannot afford right now. Our construction industry page covers how firms in this sector are approaching that problem.

ONS April 2026 data: construction output barely moved

The ONS figures for April 2026 show monthly construction output growth that PBC Today described as glacial. There is some longer-term optimism buried in the data, but month-on-month the sector is effectively flat.

That context matters when you are thinking about where to spend operational budget. When growth is slow, the margin you protect by reducing internal waste matters more than any incremental revenue gain. Missed enquiries, slow quote turnaround, manual admin eating into billable hours. These are not growth problems, they are leakage problems, and they compound in a flat market.

Why AI agents fail without a proper data foundation

Xebia's global CTO, Niels Zeilemaker, made a point this week that is worth taking seriously: AI agents only work as well as the data they can access. If the data is siloed, inconsistent, or not structured for AI consumption, the agent fails regardless of how good the underlying model is.

This is something we encounter constantly when we start working with new clients. The question is never just "what do you want the AI to do?" It is always "what does your data actually look like, and where does it live?" A quoting system connected to a well-maintained job history will outperform a more technically impressive system built on top of scattered spreadsheets and email threads. Getting the foundation right is not a prerequisite to avoid, it is most of the work. You can read more about the types of systems we build at /what-we-deploy.

Visa connects to ChatGPT to let AI agents complete purchases

This is the macro story of the week. Visa has integrated its payment infrastructure directly with ChatGPT, allowing AI agents to not just recommend products but complete transactions autonomously, without a human confirming at checkout. The system processes user prompts, evaluates merchant catalogues, and executes payment through Visa's rails at supported retailers.

This is significant less because it changes anything for most SMEs today, and more because it marks a clear direction of travel. Agentic AI, meaning AI that can take action rather than just generate text, is moving into financial transactions. For now the immediate implications are in retail. But the broader principle, that AI agents will increasingly be expected to close loops rather than just inform humans who then act, is relevant to anyone building or buying automation. It raises the bar on what "good" looks like for enquiry handling, lead follow-up, and purchase workflows.

If you want to pressure-test where your own business is leaking time and money before any of this becomes your problem, start with the free checklist at /checklist.