The Modular Briefing · October 29, 2026
Welcome back to the Modular Briefing. Each week: a short, plain-English read on private AI, governance, and owning your infrastructure. No hype, no vendor spin.
Four new people and one wiki
This week I sat down with four people who had never worked out of a shared knowledge base before and walked them through ours. We spent almost no time on the software. We spent it on habits. Where a decision goes when it's made. What a page looks like when it's done. What you do when you find something wrong on it. You fix it right then. You don't save it for a meeting.
Here's the unglamorous part. I said the same six things over and over. Put the good stuff in. Keep the slop out. One idea per page. Link it. Date it. If you learned it the hard way, write it down before you forget you learned it. Nobody is putting that on a conference slide.
But it's the whole ballgame, and this week's news kept proving it. IBM's CEO told investors that open models are now 90 percent as good as the frontier at a tenth of the price, and the big software companies said they're swapping models more often. Fine. The model was always going to be the part you rent. The knowledge base is the part you own. If yours is clean, you can point a new model at it next quarter and lose nothing. If it's a pile, you'll pay someone to clean the pile first, and the model won't be the problem.
The small version for this week: pick the one place where your business keeps what it knows. Then, every time someone asks you a question you've already answered, put the answer there instead of in a reply. Do that for a week and watch what it does to your inbox. I'm always open to compare notes.
Cale Hollingsworth is the founder of Modular Technology Group and a fractional CTO. He has been future-proofing organizations since 1993, and he works with teams too small for an enterprise AI program and too serious about their data to wing it.
The model got cheaper and easier to swap this week, by the account of the people who sell it. New York's courts reminded lawyers that a model produces text that sounds right, not text that is right. And two regulators, on two sides of the Atlantic, asked financial firms the same thing: how much of what matters to you is running on somebody else's system? Underneath all of it is one question. When you switch, what do you get to keep?
On IBM's October 21 earnings call, CEO Arvind Krishna told investors that open-weight models, the kind a company can download and run on its own hardware, are now about 90 percent as effective as the top frontier models, at "anywhere from 25 percent to 10 percent of the token cost." His estimate for an average large enterprise: "maybe two or three racks." He listed the work that doesn't need the expensive model: "summarizing an email, examining a contract, answering a customer call, composing an email." Constellation Research reported the same week that SAP's CEO said SAP is "switching the models now more often" and shipping a growing share of open-weight models to customers. Disclosure: IBM sells the hardware and software that run these models on premises, so the pitch serves the pitchman. The numbers were still said on an earnings call, where they can be held against him. For a small firm the hardware isn't the point. If the companies with the biggest AI budgets treat the model as interchangeable, the model is a rental. What you own is the knowledge you feed it, and that is only portable if it lives somewhere you control.
IBM Q3 2026 call transcript (primary) · The Register · Constellation Research
On October 26 the New York State court system's Advisory Committee on Artificial Intelligence and the Courts released a seven-page guide, The Risks of Using Artificial Intelligence in Preparing Court Papers: What Every Attorney Should Know. It builds on the model rule the court system adopted earlier this year. Lawyers may use generative AI and do not have to disclose it when they file. They do have to independently verify every AI-generated citation and quotation, and the guide asks them to "document the steps taken to verify material produced with AI assistance." The signature on the filing remains the certification. The guide's own description of the tools is the line to keep: a large language model "is designed to generate text that sounds right, not text that is right." A New York Law School associate dean, quoted by amNewYork, raised a problem she considers more dangerous than invented cases, which she calls misgrounding: the case is real, the citation is correct, and the AI misstates what it says. A quick existence check does not catch that. Reading the case does. The practical move for any firm, legal or not, is a library of what you have already verified. When the AI draws from a source you cleaned, a wrong answer stands out. When it draws from the open internet, you are checking everything, every time.
NY Courts press release (primary) · The guide (PDF, primary) · amNewYork
On October 26 the Bank of England and the Financial Conduct Authority published their fourth joint survey of AI in UK financial services. 86 percent of firms report using AI, up from 75 percent in 2024, and 80 percent are already deploying AI agents, mostly for decision support rather than autonomous action. The figure worth reading twice: third-party AI systems account for 39 percent of all deployments and 53 percent of the systems firms classify as highly material. The report also notes that pre-deployment controls are well established while post-deployment monitoring is less common, and it names concentration on a small number of cloud and compute providers as a dependency risk. Three days earlier, on October 23, the SEC published its examination priorities for fiscal 2027. Examiners will review registrants' policies and procedures for monitoring AI, which the document calls "super intelligence," in fraud prevention, back-office operations, AML and Reg BI compliance, and trading. A British survey and an American exam point an advisory firm at the same question. If the vendor behind your AI tool changed its terms tomorrow or went down for a day, what would you still have? If the client notes and the reasoning behind decisions live in a system you run, you'd be annoyed. If they live in the vendor's tool, you'd be starting over.
Bank of England / FCA report (primary) · Norton Rose Fulbright summary · SEC press release (primary) · FY2027 priorities (PDF)
On October 21 OpenAI and Google Cloud announced that OpenAI's models and its Codex agent workflows are coming to Google's Model Garden, starting in private preview for selected enterprise customers. OpenAI ended its cloud exclusivity with Microsoft and put its models on AWS in April, so this makes the third major cloud. TechTarget's read is that the race is turning toward enterprises that want model choice on whichever cloud they already use. One thing did not change. The models are served through the cloud platform; the weights stay closed and nothing is downloaded. So the same model can now be rented from more counters. For a buyer, that is convenient and it is not portability. Your prompts, your retrieval index, your logs and your permissions still have to live somewhere, and if that somewhere is inside one provider's product, a new storefront for the model does little for you.
TechTarget · Equinix on what has to move when you switch (vendor blog)
The American Society of Employers released its AI in HR Pulse Survey on October 21: 109 Michigan employers, 39 percent of them under 100 people. 78 percent use AI in HR or plan to within a year, up from 24 percent in 2024. Only 4 percent say that use is mostly integrated into formal business processes; 48 percent call it mostly informal. 53 percent have a written policy on employee AI use, 34 percent offer no AI training at all, and only 19 percent are "very confident" employees are using approved tools when they handle company data. The reassuring figure: among employers using AI for employment decisions, 86 percent always require human review. One state, one survey, so hold the exact numbers loosely. The pattern underneath is the one that matters for HR. The informal tool is where your people are doing real work, and it is also where the answers they figure out go to disappear. Whatever an employee learns inside a personal chatbot account leaves with them. A shared, approved place to put it is a retention policy for knowledge, before it is anything else.
ASE press release (primary) · DBusiness
Listen: "What You Keep When You Switch" · 5:59
This week's edition covers the first three stories above: IBM's numbers on open models, New York's guide for attorneys, and what two regulators found when they asked financial firms who their AI really belongs to. Listen to the edition.
Last week's question was whose name is on it when an AI tool gets something wrong. The running count of where that goes wrong in court keeps climbing. Damien Charlotin, an adjunct professor at UC Law San Francisco, maintains a database of court decisions involving AI-fabricated citations or unverified arguments. Straight Arrow News reported on October 28 that it now holds more than 2,000 cases, roughly 70 percent of them in the United States. A federal magistrate judge quoted in the same piece put it bluntly: hallucinations are "a bad lawyer problem, not a bad AI tool problem." The New York guide above is the court system's answer. A verified library of your own is the firm's.
Our question this week: if you switched AI providers next month, what would you lose? Hit reply and tell us. A person reads every reply.
Your data, your rules.