The Neural Brink Report · edition 04 · November 2026
Your software bill no longer depends on how many people you have. It depends on how much the machine works.
And no organization we have looked at has an owner for that meter. Here is what to ask for in writing before the January renewal.
This report comes out once a month. It does not cover vendor news. It covers what is left without an owner inside a company when it adopts what just changed.
This month there is one idea. The price of AI stopped being a fixed figure and became a meter. You no longer pay for how many people use the system. You pay for how much the system works. What follows are four jobs left without an owner, a read by type of operation and what you can ignore this month without losing anything.
01 · The thesis
Who can tell you today what one conversation costs
Renewals arrive in January. This time the price per person will not tell the whole story.
Salesforce now charges three times for the same assistant. It charges for the Sales Cloud seat, between 195 and 550 dollars a month. It charges 5 dollars for the Agentforce licence. And it charges for use: 500 dollars per 100,000 credits, or 2.00 dollars per conversation. All three charges are on its official page, checked on November 1.
Microsoft left the Copilot seat alone at 30 dollars all year. It raised the price through use instead, in Copilot Credits that carry no public price per credit.
And Zendesk writes it into the contract, in the True Forward clause. If you go over the agreed usage, the renewal moves up to that level. One quarter's peak becomes your floor from then on.
Why it is going up. Microsoft, Amazon and Alphabet together reported close to 660 billion dollars in data centre leases already signed and not yet started. It is in their quarterly filings as of September 30 and it does not show up as debt. That money has to come back as revenue. It comes back as usage.
Who can say today what one conversation with your assistant costs, and who approves it when that price goes up?
02 · left without an owner
Who shuts the agent down at two in the morning
If one of your agents starts doing something nobody asked for, someone has to stop it. That is exactly the piece that is not ready yet.
On October 27 IBM announced two things on the same day, in the same release. Its agent identity registry became available to everyone. Containment, which is the shutdown, stayed in technical preview. SailPoint added the switch now and left the piece that disables an identity for a year from now.
The industry learned to name the agent before it learned to stop it. Meanwhile the Federal Reserve, the OCC and the FDIC are already asking in writing how it gets shut down.
If you have to stop an agent at two in the morning, who shuts it down, with what tool, and who signed off that it was tested?
03 · left without an owner
Who maintains the agent the vendor built and left
Anthropic is training ten thousand engineers to install agents in companies like yours. It put up 100 million dollars and a date, the end of 2027. Accenture, Bain, Capgemini, Deloitte and McKinsey are in the first cohort. Its programme ends at handover.
After handover there is no contract. We spent the month looking for an offer to maintain agents built by a third party: with a name, a scope and a price. As of November 1 we found none.
What does exist is something else, and it is worth not confusing the two. The AI Gateway IBM announced on October 27 discovers and watches other people's agents, including competitors'. Watching is not answering when one fails.
Of the agents running in your operation today, how many does someone on your payroll understand?
04 · left without an owner
Who answers when your agent is cut off
This month's new risk is not that your agent fails. It is that the company on the other side throws it out, with nothing having failed.
On October 27 the Royal Bank of Canada revoked access it had already granted. Its customers' personal agents stopped working overnight. The next day other Canadian banks did the same.
On October 6 more of the same was documented. Amazon blocked Muse, Meta's agent. Delta stated that no outside agent is allowed to book its flights. United invokes its terms against robots. Yelp, eBay and Zillow restrict too.
This year's bottleneck is not what the agent can do. It is being let in. We found no precedent for this, but the method is old: it is the same third party access file as always. With one difference. Here the other side can close the door in twenty four hours.
If your agent is cut off tomorrow, what process stops, and what replaces it?
05
What you can ignore this month
Five figures that will circulate this month and are no use for deciding.
The 47 million euro European fines. There is no record of them. The article that made them popular does not state in its body that any were imposed: it lists the maximums under articles 99 and 101, with no date, no authority and no case number. And today they could not be imposed, for two reasons. The breaches it mentions are not enforceable until December 2, 2027. And the European AI Office has no power to fine in those cases, because that falls to the national authorities.
The 79% of companies with agents in production. It is SailPoint's, with 340 self selected responses, with no countries and no field dates.
The 150,000 agents per company by 2028. It is a Gartner projection and publishes no sample.
The October layoff figures going around, 153,074 jobs with 31,039 attributed to AI. They are from the Challenger report of October 2025.
The gap between open and closed models measured in points. It cannot be verified, because the index loads its values on the client side. What can be calculated from official prices is the ratio: the top closed model costs about 25 times what the best open one costs.
06
If you run
Ten operations where one failure is expensive. For each one: what moved this month, and the question worth asking on Monday.
Adobe Analytics, October 29: traffic arriving from assistants grew 71% in a year and brings 83% more revenue per visit. Nobody owns the catalogue the machine reads.
Who answers for what an assistant tells a customer about our products?
Article 50 of European Regulation 2026/1744 requires, from December 2, 2026, that AI made content carry a mark a machine can read. The record of who applies it is missing.
Who signs off that a piece carries the mark, and proves it six months later?
On October 22 the staff of the Florida Public Service Commission recommended denying Duke Energy's data centre proposal. It said the plan does not ensure each large customer pays its own way.
Who pays for the new load, and what happens to the residential rate?
The GAO reported on October 14 that the US Department of Defense responsible AI council had not met in nine months, still has no charter, and in February had 3 of 10 posts filled.
How many seats on the body that governs our AI are filled today?
The Bank of England and the FCA surveyed 196 firms on October 26: third parties are 53% of high materiality systems, and three clouds hold 79% of the relationships.
What share of our critical systems sits with a third party?
A JAMA study of October 15 tracked 952,008 clinicians using data from the OpenEvidence platform. Some 51.9% of queries to the model are for a clinical decision, and they happen outside the patient record.
Do we know what the doctor asked, and which model, before deciding?
Gartner, October 28, 303 organizations: 47% of supply chains cut staff because of AI, by 10.5% on average, and two thirds of the cut fell on entry level.
Who is going to know our exceptions five years from now?
Datarails, a finance software vendor, October 6, 270 organizations: finance teams spend 26% of the week checking what AI produced. And 75% point to the missing audit trail.
How much of my team's week goes to checking what the machine prepared, and who pays for it?
BCG, October 19, 398 revenue leaders: 90% already use AI and 91% expect quotas to rise. Yet only 8% report any improvement in quota attainment.
If the quota goes up because of AI, which of our own numbers justifies it?
In the EDUCAUSE list, with 882 leaders surveyed, an AI issue tops the sector's priorities for the first time.
Who measured whether what we bought last year worked, and with what data?
The list price stopped describing the cost. What decides your January bill is a meter nobody in your company reads.
07
This month's four questions
- Who can say what one conversation with our assistant costs?
- Who shuts an agent down at two in the morning, and who signed off it was tested?
- How many of the agents running today does someone on our payroll understand?
- If we are cut off tomorrow, what process stops and what replaces it?
How this report is put together
Each month we review what changed at the technology frontier and discard almost all of it. What stays is what leaves new work inside a company where one failure is expensive. When a vendor published the figure, we say so. Every figure carries its date and its sample size. That way you can go and check it yourself.
We always prefer the source that is not selling the remedy. Vendor figures are marked as such.
The sources for this edition, one by one
Price and spend. Microsoft, Amazon and Alphabet 10-Q filings as of Sep 30, 2026 (364,600, 164,922 and 130,500 million dollars in signed, not yet commenced data centre leases). Salesforce official pricing page, checked Nov 1, 2026 (Sales Cloud 195, 395 and 550 per user per month; 5 Agentforce licence; 500 per 100,000 Flex Credits; 2.00 per conversation). Microsoft 365 Copilot official pricing (30 dollars per seat) and Copilot Credits documentation. Zendesk True Forward terms. Gartner, CIO Agenda 2027, n=2,002, fielded May to July, Oct 19, 2026 (51% and 58%). KPMG, n=2,131 across 20 jurisdictions, Oct 14, 2026 (89% see value, 12% measure it against cost). KPMG US AI Pulse, n=314 (43% with token budgets). OpenAI, Anthropic and DeepSeek official pricing pages, checked Nov 1, 2026 (ratio of about 25 to 1).
Agents. IBM release of Oct 27, 2026 (Agent Identity generally available; Discover and Contain in technical preview) and watsonx Orchestrate AI Gateway of Oct 27, 2026. SailPoint announcement on Agentic Fabric and the deferral of A-ISPM (vendor). Federal Reserve, OCC and FDIC enquiries of Oct 13, 2026. Anthropic Claude Frontier Academy, Oct 2, 2026 (100 million dollars, 10,000 engineers by end of 2027; cohort with Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk). Royal Bank of Canada notice of Oct 27, 2026 and TechCrunch report of Oct 6, 2026 (Amazon and Meta's Muse; Delta, United, Yelp, eBay and Zillow).
By sector. Adobe Analytics, Oct 29, 2026, vendor with observed data (assistant traffic, +71% and 83% more revenue per visit). Regulation (EU) 2026/1744, published Jul 24 and in force since Jul 27, 2026 (article 50, machine readable marking from Dec 2, 2026). Florida Public Service Commission staff recommendation on Duke Energy, docket 07479-2026, Oct 22, 2026. GAO-27-107837 report, Oct 14, 2026 (Department of Defense Responsible AI Working Council; 3 of 10 posts as of Feb 2026). Bank of England and FCA, n=196, Oct 26, 2026 (third parties 53% of high materiality systems; three clouds 79% of relationships). JAMA, Oct 15, 2026, n=952,008 clinicians (51.9% of queries in clinical decision making; telemetry from the OpenEvidence platform, does not measure patient outcomes). Gartner, n=303, Oct 28, 2026, fielded May and June with declared selection bias (47% cut staff, 10.5% on average, 66% at entry level). Datarails, finance software vendor, n=270, Oct 6, 2026 (26% of the week verifying, 75% cite missing auditability). BCG, n=398, Oct 19, 2026 (90% use, 91% expect higher quotas, 8% improvement in attainment). EDUCAUSE Top 10, n=882, Oct 1, 2026.
What we discarded. aipolicydesk.com and arturmarkus.com (the 47 million euro fines). SailPoint, n=340 self selected (79%). Gartner, no published sample (150,000 agents per company by 2028). Challenger report of October 2025 (153,074 and 31,039). Artificial Analysis (the gap in points).
Neural Brink
The editions
One edition a month. Each at its own address, with its date.
Whoever is subscribed gets it the day it comes out. The site publishes it a week later and keeps it online for twelve months, date in view, so that what the firm said can be checked.
- Edition 04 · November 2026 · Your software bill no longer depends on how many people you have. It depends on how much the machine works. · current
- Edition 03 · October 2026 · In January your software bill goes up. The increase is AI you never approved.
- Edition 02 · September 2026 · Your company already adopted AI. The result is not showing up.
Next edition: December 2026. One topic a month, with its mechanism and its window. No commercial follow up and no calls.
Subscribe to the Report
One edition a month, the day it comes out, a week before the site. No commercial follow up.
One step is missing to complete the subscription. Check the email address and try again; you can also write to hola@neuralbrink.com.
How long does your organization take to answer the four questions above?