The AI Transformation Brief—September 22, 2026
The AI Transformation Brief
// Today’s Signal
The enterprise AI market is moving from model selection to company-wide decisions about how AI runs. OpenAI is retiring a familiar customization layer and moving customers toward plugins, which are connected AI packages with explicit skills, tools, files, and permissions. BigID and askelie are making security, compliance, and business operations callable from many AI app or chat interfaces. TypeSafe is putting a fast decision model into live business-use decision processes while warning that injected instructions can change its verdict. IBM's latest workforce study says the human ability to supervise and override AI is becoming the scarce capability, while LittleHorse is turning work into executable processes. The common signal is clear: the competitive asset is no longer access to an AI model. It is the management layer that makes AI work repeatable, permissioned, measurable, and accountable.
// Top Stories
OpenAI says custom GPTs are scheduled to stop running and leave the GPT directory on December 11, 2026.
The planned Enterprise migration experience is targeted for September 22, new custom GPT creation is scheduled to end in affected Enterprise workspaces on October 26, and the migration carries instructions, reference files, and connected apps into a plugin while leaving selected models, sharing settings, conversations, and custom actions behind. OpenAI Help Center The migration creates an inventory problem before it creates a technical one. OpenAI says a GPT must be published for the planned Enterprise migration workflow, the latest published version is migrated, and existing users do not automatically receive access to the replacement plugin. OpenAI Help Center
This is a management layer reset disguised as a product migration. A saved prompt is no longer the durable unit of enterprise AI; the durable unit is a governed package of instructions, data, tools, identity, and approvals. That changes who owns the asset. A business leader who treated a custom GPT as a lightweight app now needs an owner, a test set, a permission map, and a retirement path. The platform vendor is also asserting the right to redefine the boundary between a user customization and a live business use capability. Keep a vendor-neutral master list of AI work instructions and integrations so a provider migration does not become an operating-model migration. OpenAI Help Center
BigID launched AgentIQ on September 21 as a way to operate data security and compliance from inside BigID or directly through Claude, Copilot, GPT, or Gemini.
The company says customers can use custom prompts or pre-built BigID AI software workers to retrieve, report, respond to, and remediate data and AI security and compliance violations. PR Newswire The product move pushes security work toward a callable service rather than a destination console. The release does not provide pricing, customer names, performance metrics, or deployment counts, so the strategic signal is the interface pattern, not a proven return claim. PR Newswire
The security vendor that remains trapped inside its own console becomes a database for someone else's AI software worker. BigID is trying to keep the policy and evidence layer while letting the user's preferred interface handle the request. That is an important market boundary: the winning security product will be the one whose controls can travel with the work, not the one with the most screens. Buyers should demand an explicit action log, approval policy, and reverse path for every AI-initiated remediation. Otherwise a convenient interface will turn a compliance control into an unreviewed side effect. PR Newswire
VentureBeat reports that TypeSafe's Jev returns structured choices, scores, probabilities, and confidence instead of a written explanation.
The service launched September 15, cleared 140,000 waitlist users within 36 hours, and was reported at $0.042 per million input tokens with 70 to 500 millisecond end-to-end latency. VentureBeat The same report says an Octomind test moved from a 0.76 block probability and 0.64 confidence to a 0.48 block probability and 0.22 confidence after a hidden instruction inside tool output claimed that a destructive command had been pre-approved. VentureBeat also reports that only 34 of 107 enterprise respondents in June, and 57 of 116 in July, gave every AI software worker its own scoped identity; only 11 of those 57 also isolated the workers from one another. VentureBeat
Typed decisions look safer because they are easier to log and route, but a clean structured format does not make an untrusted input safe. A fast decision tool can move risk deeper into the system by making the wrong decision faster and more cheaply. The control requirement is a record of where information came from: the system must record the input source, tool state, model version, policy version, confidence, and the human or system that approved the action. Treat every decision model as a policy component that needs adversarial testing, version pinning, and a safe default. VentureBeat
IBM's September 21 announcement says its study surveyed 1,500 CHROs and equivalent senior executives plus 8,800 full-time employees across 21 CHRO geographies and 28 employee countries between April and June 2026.
IBM reports that 71% of CHROs identify supervising, validating, and overriding AI output as the most essential workforce skill, while only 29% of employees rank judgment as important and 60% worry AI is eroding their skills. IBM Institute for Business Value The announcement also reports that 46% of organizations do not involve the CHRO when AI strategy is defined, 28% of CHROs report a joint roadmap with IT backed by a shared operating cadence, and 43% of employees say that when AI goes wrong, the blame falls on them. These are IBM's reported survey findings, not independent peer-reviewed estimates. IBM Institute for Business Value
The workforce constraint is shifting from access to judgment. Enterprises are distributing AI capability faster than they are distributing authority, training, and safe escalation. That creates a dangerous asymmetry: the system can act at machine speed while the person accountable for the outcome still lacks the right to question it. Make a person's ability to stop or reverse AI a designed role with time, data, and escalation rights, not a sentence in a policy document. The CEO should measure whether people can safely stop an AI business process, not only whether they use it. IBM Institute for Business Value
askelie announced an Operational AI Platform on September 21 and said it was available the same day.
The platform is designed for managed service providers, software vendors, and business users to configure and compose reusable AI-powered capabilities across data, knowledge, business processes, enterprise integrations, security, governance, human review and approval of important actions, and records that let people review what happened. PR Newswire The company describes a progression from assistants and automations to digital workers and autonomous applications, with model choice and consumption-based pricing included in the platform. The release provides no customer names, quantified performance results, or pricing amounts, so this is an ecosystem and packaging signal rather than proof of deployment scale. PR Newswire
This is the platform layer trying to escape the services trap. Reusable capabilities can create leverage if the underlying controls, data connections, and audit records survive from one customer or business process to the next. If every deployment still requires bespoke integration and human cleanup, the product is a consulting wrapper. The buyer question is simple: which components are truly reusable, which permissions are customer-specific, and who owns the operational evidence when an automated action causes harm? Price the platform on completed business work, not on the number of AI software workers configured. PR Newswire
LittleHorse announced Saddle Command Center 1.3 on September 21.
The release adds AI software worker creation, pre-built task workers and skills, a JavaScript SDK, and a free serverless offering, while supporting development in Java, Python, Go, and C#. GlobeNewswire LittleHorse says its software can coordinate, stream, connect, and govern AI software workers, small software services, and events. The release cites a 3x to 4x speed increase reported by Sejal Learning Systems without new hardware, while also saying the company has been in live business use with enterprise customers for more than 18 months. GlobeNewswire
The important product is not the AI software worker. It is the executable definition of the business process around that worker. This is where enterprise value can move from a model vendor to the system that coordinates events, permissions, services, and what happens when the normal process fails. The risk is creating a new layer of process code that nobody can explain six months later. Require a business owner, a versioned business process, a agreed limit for errors and downtime, and a human escalation path before an AI business process touches a financial, customer, or compliance record. GlobeNewswire
// Shelly Palmer Pulse
Shelly Palmer reports that Amazon blocked Meta's Muse AI software worker from shopping on Amazon, citing its Conditions of Use, while Meta says Muse stores credentials securely and asks before purchases.
Palmer's enterprise lesson is that AI software worker access will be negotiated through identity, contract, and transaction rules, not assumed from a human user's browser session. The consumer retail dispute is an early test of the same control boundary every enterprise will face when an AI software worker crosses into a partner's systems. Shelly Palmer
// What It Means For Your Business
The strategic asset is shifting from access to a model toward ownership of the work system that gives AI identity, context, permission, and accountability.
OpenAI's migration shows that a vendor can change the packaging of enterprise AI inside one quarter, while BigID, askelie, and LittleHorse show that control and execution are becoming callable layers across interfaces. OpenAI Help Center BigID askelie LittleHorse This quarter, name the 10 business processes where AI can own a measurable unit of work, assign a business owner to each, and require ability to switch models, ability to move permissions when systems change, and ability to carry records of what happened in every vendor contract.
The market is re-forming around the layer between AI app or chat interfaces and enterprise systems.
Security and operations vendors are exposing their capabilities to Claude, Copilot, GPT, Gemini, and partner channels, while business process platforms are turning business process definitions into executable infrastructure. BigID askelie LittleHorse Pricing power will migrate toward the vendor that owns the policy, business process, and evidence layer, unless enterprises force open interfaces and maintain their own control plane.
Your customer experience will increasingly be assembled through a customer's AI app or chat interface, while partners will ask your systems to prove identity and permission before allowing an AI software worker to act.
Amazon's blocking of Muse illustrates the coming channel conflict between a company's direct relationship and an AI software worker acting on a customer's behalf. Shelly Palmer
which data, offers, actions, and service commitments can be exposed to outside AI software workers, under what identity, with what compensation and audit record.
The operating unit is moving from a person completing a task to a human and AI system completing a verified outcome.
IBM reports that 43% of employees say the blame falls on them when AI goes wrong, while LittleHorse describes business processes that coordinate AI software workers, small software services, and events. IBM Institute for Business Value LittleHorse Redesign one end-to-end process before adding more AI tools, define the a person's ability to stop or reverse AI role, and measure completed outcomes, exception rates, and rework instead of logins or AI-worker counts.
Jev is reported at $0.042 per million input tokens, and LittleHorse cites a 3x to 4x speed increase from Sejal Learning Systems without new hardware, but neither figure is a substitute for a verified unit-economics model.
VentureBeat GlobeNewswire Fund AI work only when the business case includes quality, what happens when the normal process fails, human review, and the cost of a wrong action, then price the completed outcome rather than raw number of AI requests.
The minimum architecture now includes portable instructions, tool permissions, identity, record of where information came from, policies with tracked versions, and a tested way to reverse a bad action.
OpenAI's migration leaves custom actions and selected models behind, while VentureBeat reports that injected tool output changed Jev's decision probability in a test. OpenAI Help Center VentureBeat Create a master list of AI software workers and decision tools, pin versions, test untrusted inputs, and block any live business use action that lacks a recorded authority chain.
The line that determines who is responsible is becoming a strategic control, not an IT detail.
IBM reports that 71% of CHROs see supervising, validating, and overriding AI output as essential, while only 28% report a joint HR and IT roadmap with a shared operating cadence. IBM Institute for Business Value Require management to show which AI software workers can act, who can stop them, what evidence is retained, and which executive owns the consequence before approving material deployment.
The most consequential shift is the rise of the enterprise management layer. OpenAI is changing the unit of customization, BigID and askelie are exposing operating capabilities through many AI app or chat interfaces, and LittleHorse is turning process logic into executable workflows. OpenAI Help Center BigID askelie LittleHorse
The broken assumption is that the model or chat interface is the product. The enterprise product is the permissioned system around the model, and its value rises with the quality of the business process, data, identity, and evidence.
The decision is whether to let each vendor define that management layer for you or to own the portable policy and outcome layer yourself. Are you building a business that can change models without changing its operating identity, or are you quietly outsourcing the way your company organizes work to the next migration notice?
The Transformation Brief is written daily by Les Ottolenghi. Delivered every morning at 6:00 AM MT, a 7-minute read on the AI shifts that matter to operators and boards.
**Build the system that lets AI tools work together. Price the outcomes. Redesign the org.**
Read On
The AI Transformation Brief—October 14, 2026
Banks, card networks, and regulators are deciding who answers for AI that acts for customers, and companies that wait will inherit their ...
Read the brief →The AI Transformation Brief—October 11, 2026
Gatekeeping decides who benefits from AI, and this weekend shows four different gates. Microsoft's CEO argues the limits on what a model ...
Read the brief →The AI Transformation Brief—October 10, 2026
Permission is becoming the operating constraint on AI. Anthropic is cutting live-internet access for internal AI tests until its monitoring ...
Read the brief →