The Agenda — 2026-09

The Model You Audit Is Not the Model You Ship

Tech Policy Press · Tier B

The fact. New work benchmarking quantization-induced bias by Tech Policy Press found that while standard safety checks pass, compressed models volunteer stereotypes in open-ended answers in roughly one in four answers across eight languages.

What it changes for you. Your AI model evaluations, if they do not state the numerical precision of the deployed model, are describing an artifact that may not be the one in your users' hands.

The question to ask on Monday. "Do our AI model certifications state the deployed configuration, including the level of quantization, and do our evaluations include multilingual open-ended generation?"

Worth quoting. A certification that does not state the precision it was performed at is describing an artifact that may not be the one in users' hands.

Ten Cents a Call: AI Voice Agents Have Turned Identity Verification Into a Commodity Attack

Artificial Intelligence on Medium · Tier A

The fact. The AnonyMousKIT phishing platform spent $19.24 to run 200 AI-driven calls to steal iPhone passcodes, at a cost of roughly 9.6 cents per attempt, according to analysis by the SOCRadar Threat Research Unit.

What it changes for you. Your assumptions about attack selectivity are now obsolete: the marginal cost of a voice call social engineering attempt is now 9.6 cents, removing economic pressure to target only large entities.

The question to ask on Monday. "Which of our customer or employee authentication processes rely on verbal proof of identity, and what is their unit cost?"

Worth quoting. “Harden the technical control and the attacker relocates to the trust process.”

Figures are from an analysis by the SOCRadar Threat Research Unit, based on production logs exposed by the platform's operators.

Owner.com Did an AI Rebuild to Accelerate Past $100M ARR. The 7 Top Lessons, and What It Takes to Copy Them

SaaStrAI · Tier A

The fact. Owner.com rebuilt its acquisition path around AI, leading to over 83% of new customers starting their journey inside an AI product and growing past $100M ARR.

What it changes for you. Your assumption that customer engagement is measured by logins and direct activity is challenged by Owner.com's inversion of this metric, where every login to fix software is counted as a software failure.

The question to ask on Monday. "What are our product success metrics that do not rely on user logins or direct activity?"

Worth quoting. If a restaurant owner is logging into the website builder to manually fix how the software set up their business, the software failed and the customer is cleaning up after it.

Growth and AI adoption figures are self-reported by the company's CEO, and the article's author is an investor and board member.

Z.ai’s Models Found 2,436 Vulnerabilities.

Towards AI - Medium · Tier B

The fact. On August 14, 2026, Z.ai published a ledger of 2,436 software vulnerabilities its models found across 269 open-source projects, with 2,383 still under embargo.

What it changes for you. Your team faces a bottleneck in vulnerability remediation, not discovery, as 98% of Z.ai's identified flaws remain in a queue for fixes.

The question to ask on Monday. "What is the average remediation time for critical vulnerabilities in our most-depended-on open-source projects?"

Worth quoting. “The scarce resource in software security has flipped from finding vulnerabilities to fixing them.”

Cyber figures are Z.ai’s own, with no independent replication.

My Routing Table Is Now a Baseline — Measuring Microsoft Foundry’s Model Router, Three Times, With…

AI Advances - Medium · Tier B

The fact. A comparative test measured Microsoft Foundry’s model router against a hand-built routing table, using a golden set and reconciling costs against Azure invoices. The router outperformed the manual table on both cost and quality.

What it changes for you. Your current model routing table, even if optimized, is a recurring decision with maintenance costs. A managed model router can reduce these costs and improve performance without manual intervention.

The question to ask on Monday. "What are the hidden maintenance costs of our current model routing table, and how do we measure them?"

Worth quoting. "There is only one honest way to find out."

Legal AI Works, Why Is the Return So Hard to Find?

Artificial Lawyer · Tier B

The fact. Law firms retain roughly 75% of what they thought they sold after utilisation, realisation, and collection take their cut, with little of the missing quarter lost in drafting.

What it changes for you. Your team may measure AI efficiency gains in billable tasks, but these gains do not translate directly into net profit for your firm this year, as they make the core product cheaper without changing pricing.

The question to ask on Monday. "Which non-billable processes, crossing multiple systems, have we never measured and could yield immediate cost savings if automated?"

Worth quoting. “Automating a non-billable hour carries none of that ambiguity. There is no client to renegotiate with, no realisation impact, no privilege question, no malpractice exposure, because the work never touched legal judgment in the first place. The hour was pure overhead. Removing it removes cost. That is a number, and it arrives immediately.”

Revenue retention figures are self-reported by the CEO of a legal tech consulting firm.

Rule of 40 Is Half Dead: Growth Is All That Matters, Margins Above 25% Don’t Help, and Category Beats Both. The Latest From Kroll

SaaStrAI · Tier A

The fact. Annualized 2026 software M&A volume is roughly 2,672 transactions, the second-highest count ever according to Kroll. However, deal value, excluding the $60 billion SpaceX acquisition of Cursor, is near a decade low at approximately $120 billion.

What it changes for you. Your assumptions about software company valuations must now account for high transaction volume but low median deal value. The market is paying a high price for a few deals, and a low price for the majority.

The question to ask on Monday. "Do our valuation projections account for the concentration of capital in a small number of deals, and how does this affect our acquisition targets or exit plans?"

Worth quoting. More companies are getting acquired than in almost any year on record, and the aggregate price paid for all of them is near a decade low.

After Rippling blew millions on AI in months, it built an employee ROI tool

AI News & Artificial Intelligence | TechCrunch · Tier B

The fact. Rippling, an HR software provider, reduced its AI token spend from 40% to 15% of its R&D headcount budget, while maintaining usage, after developing an AI Spend Console.

What it changes for you. Your contracts with AI model providers might be renegotiated to include spending caps, or you might consider less expensive models for specific tasks.

The question to ask on Monday. "Which 10–15% of our employees are driving 60% of our total AI spend, and for what output?"

Worth quoting. "The truth is that the inference providers, like Anthropic and OpenAI, have absolutely no incentives to help you control your spend."

Spending reduction figures are self-reported by Rippling.

Backstory Retiered Its Entire Customer Base in 3 Days. The Same Exercise Used to Take Five Teams a Quarter.

SaaStrAI · Tier A

The fact. Backstory, a software company, re-tiered its entire customer base of 141 accounts in three to four days using AI, a process that previously took five teams a full quarter.

What it changes for you. Your team can now re-tier your entire customer base in days, not months, by integrating unstructured signals like conversation history and internal Slack channels.

The question to ask on Monday. "What unmeasured customer signals could we integrate into a tiering analysis to get a more accurate picture of their value?"

Worth quoting. The Slack input is the one most teams could copy tomorrow. An account team’s internal dialogue is usually the earliest and most candid read on a customer, and it almost never makes it into any structured system.

Figures are self-reported by the company involved.

AI Training: Why Selling Fear of the AI Act Undermines the Acculturation That Should Be Built

JDN : Derniers contenus · Tier B

The fact. On July 27, 2026, the European Digital Omnibus postponed the application of AI Act obligations for high-risk AI systems in Annex III, including HR tools, until December 2, 2027.

What it changes for you. Your team faces a potential fine of up to 15 million euros or 3% of your global turnover for failing to map AI uses and train personnel, an obligation in force since February 2, 2025, and sanctionable since August 2, 2026.

The question to ask on Monday. "What will my team be able to do on their own in six months?"

Worth quoting. "A session built around threat produces a predictable result: participants retain fear, not competence."

Fine figures are legal maximums, not measured amounts.

AI Act: The Obligation You Might Have Missed on August 2

JDN : Derniers contenus · Tier B

The fact. Since August 2, 2026, the European AI Act requires published texts that inform the public on matters of public interest and have not been human-reviewed or are not under editorial responsibility to indicate they were written with AI.

What it changes for you. Your online publications since August 2, if they inform the public on matters of public interest and were produced with AI, must be clearly flagged, unless a human has substantively reviewed them and bears editorial responsibility.

The question to ask on Monday. "Do our legal notices designate a publication director for our online content?"

Worth quoting. "What the Commission expects is a single line: 'the identity and contact details of the legal entity, natural person or functio'."