A CFO's AI bill runs four times over budget, OpenAI and Anthropic both go down on September 29, the White House gets a voluntary AI safety pact, the UK gets a n
By Avik Ghosh, Managing Founder, illuminis
Five stories from the past week that matter to anyone buying, running or regulating AI, with our take on each.
On September 29, CFO Dive reported that finance software company Esker has been spending about four times its AI budget this year. CFO Scott McDermott said the company has gone over its AI token caps again and again in recent months, as use grew across the business and model providers moved toward usage-based pricing. Esker now counts AI spend as part of each employee's cost, next to salary, bonus and payroll taxes. A survey Esker released the week before found that 72% of finance leaders spent more than planned on AI over the past year, and 65% of CFOs found it hard to connect AI use to business outcomes. The article also cites Futurum Research, which found that agentic AI can multiply token use per task by up to 100 times. Source.
Our take: Tracking AI spend per employee shows where the money goes. It does not show whether each task was done on a model priced for the job. When agents multiply token use, the choice of model for each step, and the rework that a weak first answer causes, decide the bill. Budgets should be built on the cost to finish the work, not on token caps someone keeps raising. That is the case for CompletionPrism™.
On September 29, OpenAI reported elevated errors across ChatGPT, Codex and its API, including the Agents API. The incident listed 12 affected API components, among them Chat Completions, Responses, Batch and Embeddings, and users saw failed requests, failed logins and tasks that did not complete. Just over five hours passed between OpenAI's first update and its resolution notice, and OpenAI said a root cause analysis would follow within five business days. Source. The same day, Anthropic reported elevated errors on Claude.ai, its desktop and mobile apps, Claude Code, Claude Cowork and the Claude API from 14:00 to 14:59 UTC. A second issue blocked single sign-on and new sessions, and Anthropic said some messages sent during that hour may not have been saved. Source.
Our take: Two of the largest providers had outages on the same day, one of them lasting most of an afternoon. Batch jobs, agents and customer-facing features built on a single provider stopped with it. A software vendor cannot tell its customers to wait for a status page to turn green. Failover between models has to be part of the product. See the Embedded License.
On September 29, President Trump announced the Joint Commitment on Frontier Responsibilities, a voluntary AI safety agreement signed by the leaders of Google, Anthropic, Meta, OpenAI, xAI and Nvidia. According to The Verge, each company commits to four steps: internal controls to monitor model capabilities and alignment, an internal team to make sure those controls work, an independent external auditor or evaluator, and an independent board committee to oversee the results. The document does not mention penalties for breaking the agreement, and says that over time it may make sense to write these steps into law or regulation. Source.
Our take: The pact has no penalties, but it sets a clear expectation: AI controls should be checked by someone independent. Auditors and board committees can only work with evidence, and the same will be asked of companies using AI in hiring, credit and healthcare decisions, whether the question comes from a regulator, a customer or their own board. Records that prove what a system decided, and why, are cheaper to keep now than to rebuild later. See InferTrust™.
On September 30, the Information Commission replaced the Information Commissioner as the UK's data protection regulator, under the Data (Use and Access) Act 2025. Powers that sat with one person now sit with a corporate body overseen by executive and non-executive board members. The regulator keeps the ICO name and its existing powers and guidance, and has moved its headquarters to Manchester. A new corporate strategy is coming, with AI, cyber resilience, children's privacy and public services named as areas of focus. Source.
Our take: The rules have not changed, and neither have the powers. A new board writing a new strategy is still a good moment for marketing teams selling into the UK to check their contact data: where each record came from, on what basis it is used and how long it is kept. A database built from contacts that employees chose to share, with stale records removed, holds up much better under that review. That is how RelationLens works.
On September 29, Supply Chain Dive reported on comments from Procter & Gamble, Colgate-Palmolive and Kimberly-Clark at a Barclays conference in early September. Kimberly-Clark expects $30 million to $40 million in extra costs this quarter, partly from a tight North American freight market, and plans to cut logistics and procurement costs by combining operations after its pending acquisition of Kenvue. Colgate-Palmolive said rising oil prices could raise material costs late in the fourth quarter. P&G said its Supply Chain 3.0 program, which has brought digital tools into quality measurement and inventory management, will take the next 24 months to scale. Source.
Our take: Freight and material costs move faster than annual supplier reviews. Procurement teams need to see which suppliers and lanes are exposed while there is still time to act, not after the quarter closes. The merger point matters as well: procurement savings promised in a deal are only worth what the supplier data behind them can prove. See SourceMind™, and for testing deal savings before signing, Deal Intelligence.
In the News comes out every week. Subscribe from the Insights page to get it by email.