Reporting period: October 22, 2026, at 09:00 through October 29 at 09:00, Beijing time. Only deployment plans, business disclosures, research, and product developments officially made public during this period count as new developments in this issue. Customer cases already in operation are explicitly presented as background, rather than repackaged as new announcements.
Introduction
Enterprise AI has entered a more challenging but more valuable phase: moving from “able to write and answer” to “actually integrated into organizational workflows.” This week, Uber announced that it is expanding its 1,000-person pilot into a company-wide ChatGPT Work rollout covering more than 30,000 employees; ServiceNow reported year-over-year changes in agent usage in production; and SAP shared customer practices in finance and procurement. Meanwhile, a survey based on companies’ own reports highlighted unexpected costs and governance issues arising from agents, while Google, IBM, and Akamai strengthened the underlying engineering foundations through execution isolation, identity credentials, and access controls, respectively.[1][2][3][4][6][7][8]
Two misleading metrics warrant caution: provisioned accounts do not equal sustained employee use, and vendor-reported savings do not equal independently verified returns on investment. This article distinguishes between “rollout underway,” “already in use,” “production consumption,” and “product beta” to avoid conflating developments at different stages of maturity.
Top Story | Uber Expands ChatGPT Work from a 1,000-Person Pilot to More Than 30,000 Employees
On October 22, Uber officially announced that it would make ChatGPT Work available to more than 30,000 employees worldwide as an agent platform for enterprise knowledge work, expanding beyond its previous pilot of roughly 1,000 people. This involves more than simply providing access to a chat interface: according to Uber, the platform connects to everyday tools such as Google Drive, Gmail, Google Calendar, and Slack, allowing employees to ask it to gather information across applications, prepare meeting briefs, organize presentations, and analyze data. Software engineering teams will continue to use Codex, with the two products serving different purposes.[1]
The most useful lesson for other companies from this deployment is the organizational groundwork for rollout. Uber also highlighted an internal AI Skills library, approved connectors, and dedicated employee training. Without these supporting measures, different departments using the same model may adopt different data transfer channels, sharing permissions, and operating procedures, making auditing harder as adoption scales. Expanding from a 1,000-person pilot to a company-wide rollout indicates that procurement, permissions, training, and cross-functional workflows have been incorporated into the project, making it more than a performance test.[1]
The announcement does not, however, report that “30,000 people are already active daily,” nor does it provide average work hours saved, financial benefits, or per-person usage costs. Uber said it would share adoption figures and measurable results later. The appropriate interpretation is: the company-wide rollout is confirmed; evidence of widespread actual use and realized ROI is still pending. Other companies can assess the benefits of scaling through ticket completion rates, accuracy on cross-system tasks, time spent on human review, the number of permission violations, and total cost per task, rather than treating registration counts as a measure of success.[1]
Top Story 1 | ServiceNow Reports Production Agent Consumption Up More Than 13-Fold Year Over Year for the First Time
In its third-quarter business update on October 28, ServiceNow said that, by its platform metrics, customers’ year-to-date consumption of agents in production environments grew more than 13-fold year over year. This is different from counting agent demos or forecasting how many licenses will sell next year; at a minimum, it describes growth in production usage observed by the platform provider. However, the earnings release did not disclose a standardized unit for this “consumption,” its distribution across customers, or individual task effectiveness. It therefore does not establish that revenue or efficiency also grew 13-fold.[2]
The same announcement noted that HPE selected ServiceNow to transform customer support and field service for its cloud and AI businesses, while Stagwell chose to consolidate four service desks into a governance platform spanning more than 70 agencies. These were purchasing decisions and implementation plans announced that day, and should not be described as fully delivered deployments. The Bundesliga, Thrive, and INRY customer stories revisited in the announcement had already been announced separately, so this issue does not treat them as new deployments that occurred on October 28. Enterprises should shift their focus from “Which agents do we have?” to “How much are they actually consuming, how many tasks are they completing, and who is accountable when something goes wrong?”[2]
Top Story 2 | SAP Reports 45,000 Cloud Customers Using AI, but Reported Results Still Need Process-Level Analysis
For finance and supply chain teams, a meaningful controlled comparison should hold business volume, exception rates, and service levels constant, while bringing automated issue detection, final confirmation by finance staff, system rework, and model costs into a single report. The 45,000 figure counts customers adopting its AI products, not 45,000 fully autonomous agent projects.[3]
Top Story 3 | Redwood Survey Offers a Reminder: Scalable Agents Are Not Necessarily Governable Agents
On October 22, Redwood released survey results from 300 practitioners and managers working in technology, risk, and AI, reporting that 77% of respondents had experienced negative consequences caused by ungoverned agents. Unexpected cloud costs were the most common consequence, followed by operational disruptions and compliance issues. The survey also found that 42% reported having a centralized registry of deployed agents, while 51% said they had real-time activity monitoring. These figures measure respondents’ self-reported experiences and organizational capabilities; they do not mean that 77% of enterprises worldwide have experienced confirmed incidents.[4]
The survey was commissioned by Redwood and centered on its automation business, so its sample size, respondent composition, and potential selection bias must be considered. Still, it raises a practical issue: if enterprises allow agents to operate across accounting systems, CRM, support tickets, and code repositories without an accessible record of accountable owners, a tool inventory, a pause control, and task-based cost tracking, investigating incidents afterward becomes difficult. Establishing asset inventories, approval rules, and cost monitoring early offers more dependable benefits than simply adding more agents.[4]
Top Story 4 | Finance Agents Begin Connecting Invoice Receipt, Approval, and Payment, but Permission Boundaries Remain Essential
On October 22, Tipalti announced AP Agent, positioning it to coordinate invoice receipt, account classification, approvals, payments, and reconciliation within a single workflow, while making financial data available to authorized enterprise AI tools through MCP. Its design emphasizes field-level explanations, traceable action logs, the ability for finance staff to challenge or override agent decisions, and enterprise control over which steps can run automatically.[5]
These are a vendor’s descriptions of new products and capabilities; the source material does not provide customer payment success rates or long-term cost curves that can be independently verified. Before deployment, enterprises should distinguish between “checking invoice status,” “suggesting accounting codes,” “preparing payment instructions,” and “actually disbursing funds”: the last requires stricter approval based on the amount, recipient, and permissions. Introducing MCP does not mean existing ERP permissions, bank dual-approval controls, or payment audits can be bypassed.[5]
Top Story 5 | Google, IBM, and Akamai Address Three Layers of Production Readiness for Agents
Execution environment. On October 22, Google Cloud introduced Agent Compute Engine: it provides agents with hardware-isolated microVMs, short-lived identities, and persistent workspaces, and pauses idle instances while waiting for a model response or human approval to control resource consumption. This infrastructure already supports Google’s internal Gemini agent workloads, but a preview of Agent Compute Engine for external customers is still expected to arrive in the coming months. Internal usage experience should not be presented as a generally available service that all enterprise customers can already purchase.[6]
Identity and authorization. On October 27, IBM announced the use of HashiCorp Vault and Agent Identity for agent identities, short-lived credentials, and on-demand access revocation. watsonx Orchestrate also emphasizes cross-platform discovery, isolation, evaluation, and shutdown controls. Some discovery and isolation capabilities remain labeled as technology previews. Compared with the traditional practice of issuing long-lived administrator keys to bot accounts, this design focuses on “who owns the agent, what it can access, and when its access should be revoked”—questions enterprises must answer before connecting agents to production systems.[7]
Runtime boundaries. On October 27, Akamai released beta versions of Akamai Inference and Agent Sandbox: the former provides managed model invocation, while the latter uses the Spin ecosystem and WebAssembly to isolate agent-generated code and restrict compute, memory, external connections, and system capabilities. Both are in beta and should not be described as generally available for commercial production use. The three companies’ offerings are not fully interchangeable: microVMs provide relatively strong isolation, Vault-style components manage identities and secrets, and Wasm sandboxes restrict the capabilities of executable code. Deployment decisions must also weigh latency, observability, and operational complexity.[6][7][8]
Numbers at a Glance | Check the Denominator First
- 1,000 → 30,000+: The pilot participant count and company-wide rollout coverage reported by Uber—not daily active users or hours saved.[1]
- More than 13×: ServiceNow’s disclosed year-over-year increase in year-to-date production agent consumption, with no business denominator provided that would allow a direct conversion to ROI.[2]
- 45,000: The number of cloud customers SAP says are using its AI solutions—not the number of distinct agent projects in production.[3]
- 300 respondents, 77%: Redwood’s survey sample size and the share of respondents reporting negative consequences—not an industry-wide incident rate.[4]
- 0 images: This issue does not use vendor logos, decorative illustrations, or promotional photos of unclear origin as evidence of production results.
Engineering Take | From Model Demos to Sustainable Delivery: At Least Five Gates to Clear
The first gate is scope. Start with workflows that have clearly defined inputs, permitted actions, and acceptance criteria, such as preliminary reviews of financial anomalies, internal knowledge retrieval, customer service categorization, or IT ticket triage. Set separate permission levels for retrieval, recommendations, submissions, and payments. Just because an agent can generate accurate answers does not mean it should be allowed to directly update production financial records.[5][7]
The second is identity and secrets. Assign each agent an accountable owner, short-lived credentials, available tools, and least-privilege permissions. Keep browsers and code execution environments separate from long-lived keys wherever possible; route high-risk actions through a gateway with audit logging. This is the common control layer that this week’s offerings from IBM, Google, and Akamai point toward, though availability varies by vendor.[6][7][8]
The third is human review and fallback handling. Maintain a path for human takeover when tasks do not match rules, produce low-confidence outputs, exceed budgets, or encounter external system failures. Account for time spent reviewing, undoing actions, and reworking results, rather than equating the speed of a single model inference with efficiency gains across an entire workflow.[4][5]
The fourth is a cost ledger. At a minimum, track token, tool call, environment runtime, storage, logging, retry, and labor costs for each completed task. As agents become more complex, idle environments and repeated calls are more likely to erode the benefits of low per-token prices; establish a per-task baseline using a small batch of real-world samples before purchasing.[4][6]
The fifth is business acceptance. Track separately the number of employees granted access, weekly active users, completed tasks, human corrections, end-customer satisfaction, and financial results. Uber and SAP have reported indicators of reach and adoption, while ServiceNow has reported indicators of production use on its platform. These are all valuable signs of progress, but task-level outcome data is needed to calculate a verifiable return on investment.[1][2][3]
What to Watch Next
Watch whether Uber discloses actual user counts, the quality of completed tasks, and cross-application security incidents; whether ServiceNow further explains how it measures its 13-fold increase in production consumption and the resulting business benefits; whether SAP pilots yield customer reviews with baselines that account for labor costs separately; and the requirements and pricing for Google Agent Compute Engine and Akamai Sandbox to move from preview and testing to general availability for production use.[1][2][3][6][8]
Sourcing and deduplication notes: This issue draws on original company pages first published October 22–28 and checks each topic against the previous issue of AI Deployment Weekly for duplicates. Access to NIUX /blog/rss.xml failed during this review; the public blog listing was a cached version and could not cover all drafts and published pages from the past 14 days, so a complete sitewide deduplication check covering the past 14 days was not completed. All vendor operating metrics are explicitly labeled as self-reported. No original images from credible sources directly related to this issue’s evidence were obtained: 0 images in the body and 0 cover images.
References
- Uber brings ChatGPT Work to all employees with OpenAI — Uber, accessed October 29, 2026
- ServiceNow Reports Third Quarter 2026 Financial Results — ServiceNow, accessed October 29, 2026
- AI Release Highlights Q3 2026 — SAP News Center, accessed October 29, 2026
- 77% of Enterprises Have Paid a Price for Ungoverned AI Agents — Redwood Software, accessed October 29, 2026
- Tipalti Advances Agentic AI with New AP Agent to Bring Greater Autonomy to Finance Operations — Tipalti, accessed October 29, 2026
- Introducing Agent Compute Engine and expanded agent governance — Google Cloud, accessed October 29, 2026
- IBM Introduces Identity and Control for AI Agents, Securing Frontier Intelligence for the Enterprise — IBM Newsroom, accessed October 29, 2026
- Akamai Launches AI Inference Service and Agent Execution Environment — Akamai, accessed October 29, 2026
