Welcome to the September 2026 edition of This Month in AI. This month, the risk of AI slipping out of human control became mainstream business news. The CEOs of Anthropic and OpenAI called for a slowdown in frontier AI development, and AI stocks fell on the news. A week later, a UN scientific panel published its review of a real incident in which AI agents broke the rules they were given.
Inside enterprises, the concerns were more practical. CIOs say they can’t keep track of the agents their own employees are building. McKinsey’s annual State of AI survey found that the share of companies seeing any earnings impact from AI hasn’t moved in a year.
The slowdown debate is complicated
On September 12, Anthropic CEO Dario Amodei published an essay calling on frontier labs to slow the pace at which they improve their models. According to Axios, he proposed embedding outside evaluators inside AI companies and passing national laws that require testing of frontier models. OpenAI CEO Sam Altman agreed the industry needs to “pace the frontier,” while adding that pacing doesn’t mean stopping. SoftBank, a major OpenAI investor, fell about 10% in the sell-off that followed.
On September 21, the UN’s Independent International Scientific Panel on AI published a thematic brief about agents under evaluation at OpenAI that bypassed network restrictions, breached Hugging Face systems, cheated an evaluator, and tried to hide what they’d done. Panel co-chair Yoshua Bengio said the conditions for loss of control had come together “in a real system, not a laboratory”. In a separate survey of 111 national security and AI officials, dozens put the chance of AI operating outside human control within a decade at 10% or higher.
The labs’ finances complicate the picture. OpenAI’s own forecast projects a $14 billion loss in 2026, and Forbes reports that Anthropic could begin marketing its IPO as early as mid-October. Critics argue a coordinated slowdown would ease the spending race among the largest labs and raise compliance costs for smaller challengers, and venture capitalist Chamath Palihapitiya called the proposal “as much about market position as safety,” Axios reported. Spending hasn’t slowed. Twelve days after the essay, Anthropic signed an $11.6 billion cloud agreement with Akamai.
Nobody outside the labs can say with certainty how much weight safety and money each carry. For enterprises, the more pressing issue is the agents already running in their own systems. Those agents need bulletproof governance, with defined limits on the data they can reach and the actions they can take without a person’s approval. From our founding, RapidCanvas has championed the importance of strong governance for any AI initiative. It’s good business practice, shields the company from unnecessary risks, and ensures better insights, recommendations, and actions.
CIOs want more control
A Harris Poll survey of 685 CIOs for Dataiku, released September 24, shows how far agent adoption has outrun oversight. Of those surveyed, 84% say employees are creating agents faster than IT can govern them, and 83% have no standard process for managing agents across their lifecycle. Many have started cutting back, and 47% have already decommissioned 20 or more agents this year. Their own jobs are on the line, since 76% believe their role is at risk if they can’t show measurable AI gains by the end of 2027.
However, KPMG’s Q3 AI Pulse suggests the push for control is starting to work, with a 16-point quarter-over-quarter increase in confidence in their AI governance.
Proof of financial value remains scarce
McKinsey’s State of AI 2026, based on 1,719 respondents, found that 37% of organizations attribute at least some EBIT impact to AI, unchanged from 2025. Only 6% qualify as high performers under the company’s models. McKinsey wrote that “organizations’ conviction in AI is growing faster than the immediate financial returns.”
KPMG’s survey is more positive, with nearly six in ten leaders reporting measurable business value, but most of that value shows up as productivity (55%) and faster decisions (49%). Only 37% cite stronger financial performance, matching McKinsey’s EBIT figure. Measurement accounts for part of the gap. Grant Thornton’s 2026 AI Impact Survey Report found that 78% of the 950 business leaders it surveyed aren’t confident they could pass an independent AI governance audit within 90 days, and it concludes that most organizations “are scaling AI they cannot explain, measure or defend.”
While gains in productivity and faster decision-making are valuable, companies are increasingly demanding short-term ROI when they evaluate and pursue AI initiatives.
At RapidCanvas, our process begins with a clear definition of pain points, projects, and expected outcomes, as well as a precise timeline for expected positive ROI. Typical time to value runs at six to twelve weeks when clients leverage our Hybrid Approach™.
Rising costs put ROI in focus
McKinsey’s July 2026 analysis of enterprise AI costs found that 93% of respondents to its Enterprise AI FinOps survey had exceeded their AI budgets, and that token usage for the same task can vary by as much as 30 times.
Response to these overruns has been immediate in many businesses. In the State of AI survey, 20% said operating costs have already limited their use of AI, and KPMG found that 74% of organizations now include cost reviews in AI approvals.
Setting limits is a natural response to overruns, but also holds the potential to threaten the pace of AI deployments. In our view, there are three key ways that companies can control these costs strategically, without limiting the potential of AI transformation.
- Confirm that AI is the right tool for each initiative. Some problems are better solved with standard automation or a process change.
- Complete a cost-benefit analysis for every project before it begins, with a bottom-line target and a baseline to measure against.
- Assign a business owner to each AI initiative who is accountable for both its results and its running costs, and who reviews them regularly after launch.
The build versus buy gap
Nearly a third of State of AI respondents turned down a software purchase and built the functionality themselves with agentic coding tools. The theory here is that an internal team can build a customized solution while dramatically shortening development timelines. Yet only a quarter of companies using those tools have achieved meaningful acceleration, according to the Technology Trends Outlook.
The pros of buying are well-known. A purchased solution has already been tested at other companies, and the vendor handles updates, security, and maintenance that an internal team would otherwise have to staff. It also gets to production faster. MIT’s NANDA initiative found that AI tools bought from specialized vendors or built through partnerships succeeded about 67% of the time, while internal builds succeeded about a third as often.
However, buying carries well-known risks. IDC’s analysis “AI Is Ready. Enterprises Are Not. Vendors Need to Fix It.” predicts that nearly 50% of AI-driven digital use cases will miss their ROI targets in 2026, citing poor data foundations and weak integration. The report puts the burden on vendors to connect their products to each company’s own workflows and data.
Mitigating both build and buy challenges prompted RapidCanvas to develop its Hybrid Approach™, which combines human experts and a proven agentic platform. By leveraging the strategic thinking and expertise of PhD-level data scientists and category experts, companies gain custom solutions tailored to their needs and existing tech stack. These solutions are crafted using any of 1,000+ prebuilt agents and connectors to maximize speed to market and time to value.
September’s common thread
The frontier labs are arguing over who should oversee the most capable models, and CIOs face a smaller version of the same question with agents their employees built without approval. Flat EBIT numbers and rising costs are pushing companies to tie AI spending to results they can measure. The research points to four requirements:
- Governance that defines what each agent can access and do, with human approval for consequential actions
- A bottom-line target and a baseline for every initiative before it launches
- Regular cost reviews that weigh inference spending against measured returns
- Build and buy decisions based on fit with existing workflows, data, and skills
With many AI budgets under review heading into 2027, governance and measurable returns will decide which programs keep their funding.
These requirements are central to how RapidCanvas works. We use a Hybrid Approach, combining human experts with a proven agentic AI platform to deliver solutions customized to each client’s goals, workflows, and tech stack. We help enterprises move from disconnected pilots to governed, measurable use cases that business and technical teams can trust.
That’s it for this month. Thanks for reading. Subscribe to our newsletter for more curated AI insights from leading business publications.
