The price of delegated work

01 Executive signal

September made delegated work more concrete. Microsoft put subscriptions and metered execution into the same commercial story. GitHub strengthened enterprise control over agent operations while announcing a change to feature defaults. AWS and Google made managed execution environments more capable. Enterprise software vendors moved business logic and context closer to whichever agent interface employees choose. These developments widen the feasible set of work. They also make it easier for activity, permissions and expenditure to expand before an organization has defined the result it wants to buy. S01MicrosoftIntroducing the new Copilot with Home, Code, and Autopilot2026-09-25 · vendor announcement S03Microsoft LearnSeptember 2026 Partner Center announcements2026-09-16 · official product documentation S07GitHubEnterprise-managed permissions for GitHub Copilot agent operations2026-09-09 · official release notes S08GitHubDefault enablement of Copilot features for Copilot Business and Enterprise2026-09-24 · official release notes S10AWSNew AgentCore Runtime generally available2026-09-18 · official release notes S14Google CloudGemini Enterprise Agent Platform release notes2026-09 · official product documentation S15SalesforceSalesforce unveils AIforce2026-09-15 · vendor announcement

August's edition, The control plane meets reality, argued that inventories and runtime controls were becoming basic infrastructure. September adds a sharper economic question: can the buyer connect a particular permission, execution trace and invoice to an accepted business outcome? An inventory answers what exists. A production operating model must also answer who can act, under which conditions, with whose budget, and with what evidence of completion.

The strongest evidence for improving enterprise work is narrower than the product narrative. A randomized patent-drafting study reports better work and heterogeneous learning effects. A synthetic enterprise benchmark finds that curated context improves evidence access and task completion. Customer cases show credible deployment and some measured outcomes, but rarely establish incremental value against a stable counterfactual. The sensible conclusion is selective expansion with better measurement, rather than either universal autonomy or a blanket productivity verdict. S17SmarshSmarsh scales Agentforce after strong customer support results2026-09-03 · customer primary case S25Autor et al., MIT/NBERDoes AI assistance enhance or erode expertise?2026-09 · research working paper S26WorkWorlds authors / arXivWorkWorlds: evaluating agents in realistic enterprise environments, v22026-09-23 · research preprint

October's immediate decision: attach a named business owner, an authorization boundary and a cost-per-accepted-outcome measure to every workflow proposed for wider deployment. Review forthcoming defaults and model retirements before their effective dates. Preserve a matched baseline long enough to discover whether human review, exceptions and failed attempts absorb the apparent saving.

02 The five developments that change priorities

The scores below are September judgments. Signal scores measure the strength of a decision implication; CIO relevance measures proximity and breadth of the decision. Neither is an estimate of financial return. Confidence concerns the stated interpretation, not the vendor's entire product.

Figure 1 asks which September judgments require action. Ordinal scores keep relevance separate from evidence maturity. The apparent precision of two numbers should not conceal the narrower source behind each judgment.

03 Three overrated interpretations

A unified agent interface means an autonomous enterprise is ready. Home, Code and Autopilot are a product direction, with distinct rollout states. The work still crosses identities, data entitlements, business rules and budgets. A private preview cannot support an assumption of general tenant availability, and an available tool cannot establish reliable process completion. The operational question is whether the particular workflow can complete, fail safely and leave sufficient evidence. S01MicrosoftIntroducing the new Copilot with Home, Code, and Autopilot2026-09-25 · vendor announcement

More activated agents, employees in scope or customer interactions establish productivity. Such counts are useful for understanding distribution. They do not identify incremental output, labor released for other work, or net cost. A large rollout can coexist with shallow use; a small workflow can generate substantial value. The denominator must follow the decision. The adoption cases below retain each source's denominator rather than combine unlike counts into a league table. S15SalesforceSalesforce unveils AIforce2026-09-15 · vendor announcement S17SmarshSmarsh scales Agentforce after strong customer support results2026-09-03 · customer primary case S18Adecco GroupAdecco Group and Salesforce expand Agentforce to more than 40 countries2026-09-15 · customer primary case

A runtime speedup can be applied to the business case. AWS's published cold-start results are material infrastructure evidence. They describe one latency component under a specified benchmark. They do not measure the time a customer waits for a resolved dispute or the time an engineer spends checking a change. End-to-end gains depend on how much of the original cycle cold starts occupied and on what follows execution. S10AWSNew AgentCore Runtime generally available2026-09-18 · official release notes S11AWSThe new AgentCore Runtime: elastic, optimized, and consistently fast starts2026-09-18 · vendor technical benchmark

04 Three underappreciated developments

Default policies are part of architecture. A team may review a feature at purchase and never revisit the unconfigured policy that later admits it. GitHub's 22 October change creates a concrete review window. Explicit decisions should remain explicit, with an owner and a recorded reason. An organization's architecture review loses force if product defaults can silently become its operating policy. S08GitHubDefault enablement of Copilot features for Copilot Business and Enterprise2026-09-24 · official release notes

Ownership transfer is a production feature. Google's September release notes include shared-agent ownership transfer. This is less conspicuous than a new model, but it addresses an everyday operational problem: a useful workflow outlives its creator. Offboarding, credential renewal, budget ownership and incident response should continue when that employee changes role. S13Google CloudGemini Enterprise release notes2026-09 · official product documentation

A narrow predictive model may be the better AI investment. SAP's availability of TabPFN-3.5-Plus in AI Core broadens access to structured-data prediction. That does not establish superiority on a buyer's data. It does justify a conventional evaluation for a defined forecasting, classification or prioritization task, where outcomes and error costs may be easier to observe than in an open-ended assistant. Benchmark it against the current statistical model before adding conversational orchestration. S20SAPTabPFN-3.5-Plus now available in SAP AI Core2026-09-15 · official product announcement

05 The month in numbers

Publication in September does not mean that the underlying activity happened in September. These are distinct observations, with their own units and populations.

06 CIO signal map and August continuity

The ten signal IDs below continue August's map. August's impact, momentum and evidence scores remain frozen in the accompanying data. September adds the current profile's signal and relevance fields; it does not retrofit the earlier edition or subtract unlike scores. There is no defensible composite CIO index in this series, so none is manufactured here.

The reclassification of SIG09 narrows the claim rather than declares sovereignty unimportant. The evidence now supports a specific procurement option more clearly than a general shift in strategic independence. SIG04 is carried forward with lower confidence in the monthly update because selected customer cases cannot reveal a population distribution.

August's forecast also stays frozen: by 31 August 2028, at least 60% of large enterprises with ten or more production agents would maintain a cross-platform inventory, while fewer than 25% of high-consequence workflows would permit ungated irreversible external actions. September's releases are consistent with demand for controls. They neither verify the adoption threshold nor identify its eventual denominator. Status: unresolved.

07 Enterprise AI: define completion before expanding execution

A useful enterprise unit is a workflow with an observable destination: a correctly resolved support case, a valid quotation, an accepted code change or a reconciled invoice. An agent run is an input to that process. Treating it as the output produces an attractive dashboard even when the organization is paying people to repair the result.

Each proposed deployment should therefore define completion, exceptions and reversibility in advance. A draft can be checked before release. A refund, supplier change or external message may create consequences that survive a later correction. Approval should follow the consequence and the remaining uncertainty, rather than the superficial category of the tool. A read-only connector can still expose confidential information. A write operation inside a sandbox may be entirely reversible.

September gives buyers more options for persistent execution and managed runtimes. It also raises the importance of testing recovery. Can a workflow resume without repeating a financial transaction? Can its owner explain which business rule governed a disputed action? Can the same acceptance test run after a model, connector or prompt update? These are proposed buyer tests, not capabilities inferred from a launch announcement. S10AWSNew AgentCore Runtime generally available2026-09-18 · official release notes S14Google CloudGemini Enterprise Agent Platform release notes2026-09 · official product documentation S30OpenAIIntroducing the Agents API2026-09-10 · vendor product announcement

Figure 2 asks how far the evidence has progressed. Its stages describe what the reviewed source supports. They are not a universal ladder: a customer-measured result can still have weaker causal credibility than a controlled research study, and general availability does not imply scaled use.

The investment portfolio should admit three outcomes: expand, redesign or retire. Teams need a credible route to the third. Otherwise every pilot becomes an indefinite request for more integration work, and the organization never learns which constraints actually bind.

08 Operating model: the business owner owns acceptance

Central IT should maintain shared identities, runtime patterns, observability and procurement rules. A business owner should remain accountable for the definition of a satisfactory result. This allocation avoids a common ambiguity: IT reports successful executions while operations reports unresolved work.

Fund shared infrastructure as a reusable capability, but assess workflow expansion against its own economics. The allocated cost of a platform and the marginal cost of one more workflow answer different questions. Report both where the distinction changes a decision. Do not make a pilot appear profitable by excluding a dedicated integration team, or make a sound marginal expansion appear unattractive by charging it the entire historical platform build.

The practical artifact is a short operating agreement for each production workflow. It names the owner, accepted result, allowed actions, budget, evidence retained, change policy and exit route. Its value comes from resolving decisions, not from the number of pages in a governance pack.

09 Microsoft: integration increases both convenience and dependence

Microsoft's September direction connects a familiar work surface with persistent execution, organizational context and consumption billing. That combination can reduce the effort needed to start a workflow. It can also place identity, business context, execution policy and the invoice within one vendor's operating model. The decision is where that integration earns its cost and where the buyer needs an independent control or exit.

Availability has to be read capability by capability

The 25 September launch groups Home, Code and Autopilot under one Copilot experience. Code enters Frontier rollout; Autopilot expands a private preview. Managed Runtime is preview. These distinctions matter when a business sponsor assumes that a launch presentation describes what every licensed employee can use next week. Fabric IQ in Chat and Cowork has a different status from the preview for IQ Sharing. S01MicrosoftIntroducing the new Copilot with Home, Code, and Autopilot2026-09-25 · vendor announcement S06MicrosoftNew Microsoft data innovations unlock what only your business knows2026-09-28 · vendor announcement

Cowork itself is not a September invention. Its release history places general availability and consumption billing earlier in the year. September's changes include working on existing Office files and additional app-skill capabilities. Preserving version history is useful, but the buyer still needs to establish who approved a consequential change and whether the workflow leaves an adequate business record. S05Microsoft LearnWhat is new in Copilot Cowork2026-09 · official product documentation

The Microsoft 365 release summary also spans late August and September. It should not be converted into a count of September launches, nor into a promise of simultaneous tenant rollout. Tenant, geography, license and release channel remain part of every deployment check. S04Microsoft LearnMicrosoft 365 Copilot release notes2026-09-23 · official product documentation

Commercial architecture now belongs in the workflow review

The broad commercial direction combines a user subscription with usage-priced advanced work. The narrower CSP notice is more actionable: from 2 November, new Copilot Business purchases through CSP, including standalone and bundled purchases, default to pay-as-you-go billing. It is not a statement that every existing enterprise license changes on that date. S01MicrosoftIntroducing the new Copilot with Home, Code, and Autopilot2026-09-25 · vendor announcement S03Microsoft LearnSeptember 2026 Partner Center announcements2026-09-16 · official product documentation

The buyer should ask for a worked bill from the intended workflow, including failed attempts, retries and review. An expenditure cap is a budget control; it does not establish value. A low unit price is also insufficient if the workflow generates many attempts for each accepted result. In the scenario later in this edition, review effort can consume the apparent saving even with unchanged software expenditure.

Pricing comparisons in the retrieved Microsoft community material could not be checked against the full page. They are excluded from the numerical case. Contract rates, discounts and credit conversion rules should come from the organization's actual commercial terms, not a generalized benchmark. S02MicrosoftEvolution of the Copilot pricing model2026-09-25 · vendor pricing explanation

GitHub makes policy and migration immediate concerns

GitHub's managed permissions cover shell, file and network operations across supported agent surfaces. The release says users cannot weaken enterprise policy through local settings or saved approvals. That is a strong reason to test the mechanism against the organization's own forbidden actions. It is not a substitute for that test. S07GitHubEnterprise-managed permissions for GitHub Copilot agent operations2026-09-09 · official release notes

The feature-default change effective 22 October and the selected-model retirements scheduled for 19 October require different responses. The first is a policy inventory and explicit-choice exercise. The second is an evaluation and migration exercise. A replacement model may preserve a product label while changing behavior on internal tasks, tool use or refusal boundaries. Both should have named owners before the effective date. S08GitHubDefault enablement of Copilot features for Copilot Business and Enterprise2026-09-24 · official release notes S09GitHubUpcoming deprecation of selected GitHub Copilot models in mid-October2026-09-18 · official release notes

Fabric, Power Platform and Dynamics concentrate business context

Fabric IQ brings semantic context into Copilot. Existing permissions are a useful starting point, but inherited access can reproduce old entitlement errors. A technically authorized query may also combine definitions that the business has never reconciled. Give a domain owner responsibility for the meaning of the answer, including which metric definition prevailed. IQ Sharing adds a further boundary between organizations and remains a preview in the reviewed announcement. S06MicrosoftNew Microsoft data innovations unlock what only your business knows2026-09-28 · vendor announcement

The announced Dynamics and Power Platform grounding path belongs on the test agenda with its stated future rollout, not in a September inventory of proven outcomes. A successful demonstration should become a bounded process comparison, including exception handling and integration effort. S01MicrosoftIntroducing the new Copilot with Home, Code, and Autopilot2026-09-25 · vendor announcement

Identity, security, partners and exit

Entra, Purview, Defender and Agent 365 can fit an integrated governance design. This edition does not establish a new measured September outcome for each component. The buyer should test effective permissions and retained evidence across the actual chain rather than assume that adjacent product names imply complete coverage. A partner's implementation can be necessary and valuable; its dependency should appear in the cost model and operating agreement.

The exit question is concrete: can the organization export the workflow specification, acceptance set, business definitions, relevant logs and cost records in a usable form? Ask an internal team to replay a sample outside the original interface. A nominally open connector is weak evidence of portability if business rules or approval history remain inaccessible.

Microsoft's strongest proposition is reduced coordination cost across familiar work. Its principal buyer risk is that convenience makes the dependencies difficult to see. Both should be measured within a particular workflow, rather than settled by a platform-wide endorsement or rejection.

10 Five bounded tests for October

These are proposed experimental designs. Their thresholds are decision rules chosen for a pilot, not measured effects from the sources. Agree them before observing the result and amend them only with a recorded reason.

For randomized tests, check whether the planned sample can detect a decision-relevant effect. The suggested cohort sizes are operational starting points, not statistical power calculations. Where random assignment is impossible, record selection differences and avoid a causal label. Pre-register how reopened cases, partial completions and missing review time enter the denominator.

11 Announcement reality check

No item is marked abandoned or delayed without affirmative evidence. Failure to find a new release is not evidence that a project has stopped. These labels concern the reviewed evidence, not undisclosed customer deployments.

12 Hyperscaler competition: the control point differs

This comparison is deliberately not a weighted vendor ranking. The relevant weights depend on installed systems, data sensitivity, latency, internal engineering capability and the cost of exit. A single score would disguise those differences.

Figure 3, the runtime evidence exhibit, asks what the AWS benchmark actually establishes. The source reports ranges across artifact sizes. Showing a band preserves that structure and avoids turning the most favorable pair of endpoints into a universal speedup.

For a workflow dominated by human review, a faster cold start may matter little. For a frequently invoked interactive tool, it may improve responsiveness. Measure the latency budget by stage before purchasing capacity on the strength of a component benchmark.

13 Enterprise software: the interface can move while the rules stay

Salesforce's AIforce proposition makes business data, actions and governance accessible from more interfaces. SAP positions process knowledge, enterprise architecture and adoption tooling around autonomous work. Databricks emphasizes governed business knowledge. LSEG and Snowflake extend licensed data distribution. The common direction is that the visible interface becomes less reliable as a guide to where durable control resides. S15SalesforceSalesforce unveils AIforce2026-09-15 · vendor announcement S19SAPThe autonomous enterprise: business transformation management solutions2026-09-22 · vendor announcement S27DatabricksUnity Catalog Pages: a governed home for business knowledge2026-09-22 · vendor announcement S29Snowflake / LSEGLSEG and Snowflake expand collaboration2026-09-30 · joint partnership announcement

This is an analytical map of potential power, not a claim that every vendor already controls every layer. The strategic question for buyers is which dependencies are worth accepting because they reduce coordination costs, and which dependencies make future bargaining or recovery unnecessarily difficult.

The strongest negotiation may concern exports and decision rights rather than the model menu. A buyer that can reproduce accepted work, explain its permissions and move its evidence has more credible alternatives than one that can only switch the model behind the same opaque workflow.

14 Security: contain actions and preserve the evidence

Anthropic's September investigation concerns four incidents in third-party cybersecurity evaluations with reduced safeguards and misconfigured internet access. Three were already disclosed in July; the additional incident occurred in January, was identified in August and was disclosed in September. This chronology matters. The article is new evidence, not four new September production breaches. Its scan populations cannot be used as a denominator for an enterprise incident rate. S23AnthropicAlignment assessment: cybersecurity incidents2026-09-09 · first-party incident investigation

The lesson for CIOs is a containment question. Can an agent's effective authority exceed the intended scope when it encounters untrusted material, follows an indirect instruction or calls another tool? Model refusals are one defense. Runtime permissions, network boundaries, credential scope and approval requirements constrain what happens when that defense is insufficient. September's control releases make those mechanisms more accessible, but the security property still has to be tested in the buyer's environment. S07GitHubEnterprise-managed permissions for GitHub Copilot agent operations2026-09-09 · official release notes S14Google CloudGemini Enterprise Agent Platform release notes2026-09 · official product documentation

A practical exercise starts with a harmless substitute for a consequential asset. Attempt unauthorized data egress, a forbidden filesystem change, a credential misuse and a repeated external action after a simulated timeout. Record the denied action, the identity, the policy version and the owner who receives the alert. Verify that an operator can stop the workflow without losing the evidence needed to explain it.

Supply-chain review should include connectors and their scopes, not only model providers. A model upgrade, connector update or new app action can change the effective attack surface while the workflow's business label stays the same. Pair change approval with replay of critical permitted and forbidden cases. For high-consequence actions, keep an accountable human approval until the organization can justify a narrower gate with evidence.

15 Data architecture: context is an engineered input

WorkWorlds tests enterprise-like tasks inside a synthetic environment with different employee seats and agent configurations. In the September v2 results, curated context improves access to relevant evidence more than it improves task completion. That gap is useful: finding information and acting correctly on it are distinct problems. S26WorkWorlds authors / arXivWorkWorlds: evaluating agents in realistic enterprise environments, v22026-09-23 · research preprint

Figure 4 asks how much curation changes each outcome. It plots the paper's standardized differences and their confidence intervals. The accompanying table supplies each condition's level. Neither v1 results nor raw appendix ratios are mixed into these estimates.

Fabric IQ, Unity Catalog Pages and SAP's transformation-management direction address related problems through different product surfaces. Their presence does not remove the need for semantic ownership. The difficult work often concerns a disputed definition, an obsolete instruction or a record whose status depends on another system. S06MicrosoftNew Microsoft data innovations unlock what only your business knows2026-09-28 · vendor announcement S19SAPThe autonomous enterprise: business transformation management solutions2026-09-22 · vendor announcement S27DatabricksUnity Catalog Pages: a governed home for business knowledge2026-09-22 · vendor announcement

Infrastructure planning should follow the workflow's latency and recovery requirements. Separate model time, retrieval, tool execution, queueing and human review. Persistent execution may reduce repeated setup, but it also retains state that must be governed. Record retention, deletion, credential expiry and the behavior of interrupted runs alongside throughput and price.

16 Economics dashboard: accepted work is the denominator

The stable metric carried forward from August is cost per verified outcome (CPVO). Here, verification means an outcome passes a defined acceptance rubric; it does not imply an external audit. The broad numerator includes model and platform charges, tools, integration and support, human review, rework and an explicit treatment of risk. Transition expenditure and recurring expenditure should be visible separately.

Figure 5 asks when an apparently inexpensive workflow ceases to save money. It is an illustrative scenario, not a vendor quote or an observed enterprise result. Assume 1,000 attempted cases, $1,400 in platform and usage charges, $1,200 in rework, and review labor at $60 per hour. At 85% acceptance and four review minutes per attempt, CPVO is $6,600 / 850 = $7.76. The manual comparator is assumed to cost $12 for an accepted outcome.

At the same acceptance rate, the scenario reaches $12 at 7.6 review minutes. At eight minutes it costs $12.47. Lower acceptance increases the cost of every accepted result even when the software bill stays flat. The formula is (2,600 + 1,000 × review minutes) / (1,000 × acceptance rate) under the stated $60 hourly rate.

The interactive version lets readers vary acceptance and review time and export the assumptions. Its purpose is to expose a break-even boundary. It cannot estimate an organization's return without actual prices, workload mix, review observations and a credible baseline.

17 Productivity evidence check: better output can coexist with unequal learning

The September working paper by Autor and co-authors studies AI assistance in patent drafting through a pre-registered, three-month field experiment involving 133 lawyers at 11 US firms. Blinded assessments show assisted quality gains of 0.34 standard deviations at day 10 and 0.38 at day 90. These are standardized quality effects, not percentages of time saved or profit gained. S25Autor et al., MIT/NBERDoes AI assistance enhance or erode expertise?2026-09 · research working paper

Figure 6 asks whether assistance transfers into independent capability. The study also reports a 0.32-standard-deviation effect on unaided redlining overall. Senior lawyers show a 0.45 effect; juniors do not show an average gain and their outcomes become more dispersed. The figure does not put an invented numerical estimate on the junior result.

The result weakens a simple claim that using AI necessarily erodes expertise. It also weakens the equally simple claim that better assisted output means everyone is learning. Prior expertise and the design of work appear relevant to what the worker retains. Attrition, the narrower unaided assessment sample and possible contamination require caution. The study concerns professional patent tasks, not every knowledge-work occupation.

Across the edition, evidence sits at several levels. WorkWorlds tests a technical mechanism in a synthetic organization. The patent experiment estimates effects in a real professional task. Smarsh reports customer workflow results. None alone establishes firm-wide productivity or economy-wide output. A stronger firm-level claim would require the cost of complementary work, displaced tasks, demand responses and sustained performance over time. S17SmarshSmarsh scales Agentforce after strong customer support results2026-09-03 · customer primary case S25Autor et al., MIT/NBERDoes AI assistance enhance or erode expertise?2026-09 · research working paper S26WorkWorlds authors / arXivWorkWorlds: evaluating agents in realistic enterprise environments, v22026-09-23 · research preprint

There is no fixed productivity-paradox verdict here. September supports a more specific belief: AI can improve selected work, while the organization must still build the conditions that convert the gain into reliable operations and useful human capability. The next test is whether those gains survive complete cost accounting and expansion to ordinary cases, not just the cases chosen for a demonstration.

18 Regulation and sovereignty: the relevant boundary is operational

The Cyber Resilience Act's reporting obligations began applying on 11 September. The Commission describes early warning within 24 hours and notification within 72 hours for the covered categories, with subsequent reporting deadlines that vary with the incident type. This is an immediate issue for manufacturers within scope, not an undifferentiated obligation imposed on every CIO or every SaaS service. Product security and counsel should establish scope and reporting responsibility together. Main remaining obligations arrive later, in December 2027. S24European CommissionCyber Resilience Act reporting obligationsundated · official regulatory guidance

Figure 7 separates September disclosures from dates on which action becomes necessary. A notice of an October or November change belongs in the September radar because it changes preparation, but it must not appear as a September effective event.

For the affected manufacturer, a reporting exercise should establish who starts the clock, what evidence is available, who decides which reporting category applies and who can submit the notification when a primary owner is absent. Agent logs and supply-chain records may support that process; their existence alone does not establish compliance.

ServiceNow's German partner-cloud announcement and the Mistral–Cloudera partnership supply different sovereignty options. The first describes a live hosting/service arrangement in Germany; the second describes deployment flexibility across customer environments. Neither should be stretched into a claim that every model, support channel, encryption key, backup or disaster-recovery dependency lies under the buyer's desired jurisdiction. S21ServiceNowT-Systems goes live as Sovereign Partner Cloud provider in Germany2026-09-07/08 · vendor partner announcement S28Mistral AIMistral AI and Cloudera partnership2026-09-10 · vendor partnership announcement

A procurement review should map those dependencies, contractual rights and recovery routes. Local data location can be valuable. Operational control also depends on who can change the software, access production, revoke credentials, restore service and export evidence. No new September AI Act timetable is asserted in this edition because the reviewed sources do not establish one.

19 Skills and organizational change

The patent experiment makes seniority an explicit design variable. Training should therefore distinguish operating a tool, judging its output and retaining the underlying professional skill. An employee may become faster at supervising a familiar task while becoming less able to diagnose an unfamiliar failure. A single adoption or training-completion measure will miss the difference. S25Autor et al., MIT/NBERDoes AI assistance enhance or erode expertise?2026-09 · research working paper

For junior staff, protect opportunities to form a judgment before seeing the model's answer. Ask them to explain a decision, identify missing evidence and repair a flawed result. For experienced staff, test whether assistance expands useful capacity without removing the difficult cases through which the next generation learns. These are proposed training responses to heterogeneous evidence, not interventions already validated by the study.

Managers also need a different performance conversation. Measure quality, accepted throughput, review effort and learning. Avoid incentives that reward agent activity alone or encourage employees to hide the repair work that makes a tool appear successful. A pilot should make failures easy to report without treating them as a personal adoption failure.

In Germany and other settings with formal employee representation, implementation planning should include the relevant local consultation and work-design process. This edition does not offer a new legal interpretation of those obligations. Its operational point is narrower: monitoring and task redesign affect people, and their effects belong in the project design before deployment, not only in an adoption campaign afterward.

20 Enterprise adoption cases: useful evidence with unequal limits

The evidence-quality scale here is the profile's case scale: 1 for an unsupported anecdote, 2 for a named deployment, 3 for a defined metric with a plausible comparison, 4 for scaled transparent outcomes and 5 for unusually rigorous causal or independent evidence. It differs from both signal strength and product maturity.

Cases selected by vendors or participating customers can overrepresent successful deployments. The absence of comparable failed projects prevents a representative success-rate estimate. These cases therefore inform mechanisms and pilot design. They do not yield an industry adoption rate or a forecast that the next implementation will succeed.

21 Transactions and alliances: access to the workflow matters

The material September items in the reviewed set are alliances and distribution arrangements. Salesforce and Siemens connect industrial product knowledge with customer processes. Mistral and Cloudera address deployment choices. LSEG and Snowflake extend the distribution of licensed content. Each can reinforce a different control point: workflow context, execution environment or rights to proprietary information. S16SalesforceSiemens and Salesforce deepen AI partnership2026-09-15 · vendor customer case S28Mistral AIMistral AI and Cloudera partnership2026-09-10 · vendor partnership announcement S29Snowflake / LSEGLSEG and Snowflake expand collaboration2026-09-30 · joint partnership announcement

Those mechanisms are more useful to a CIO than a transaction count. A partnership may reduce integration work or enlarge the set of usable data. It may also add dependencies, overlapping support responsibilities and restrictions on reuse. Ask which party owns the unresolved case when the combined workflow fails, and which contract governs the evidence needed to investigate it.

SAP's Prior Labs acquisition is relevant background to the TabPFN product direction, but its completion belongs to July. It is not counted as September M&A. No verified, sufficiently material September completed acquisition is established by this edition's reviewed set. That is a coverage limit, not a claim that the market had no transactions. S20SAPTabPFN-3.5-Plus now available in SAP AI Core2026-09-15 · official product announcement

22 Decision agenda

23 Questions for the leadership team

• Board and CEO: Which business result would make the current AI portfolio worth its full recurring cost, and which projects would we stop if the result did not appear?
• CFO: Can we reconcile paid usage and human repair work to accepted outcomes without counting the same benefit twice?
• CISO: Which agent can create an irreversible external consequence, and what evidence shows its authority cannot expand through an indirect tool path?
• Data leader: Who settles a disputed definition when two authorized sources disagree, and how quickly does the correction reach deployed workflows?
• HR and line managers: Are junior employees developing independent judgment, or only becoming more proficient at accepting suggestions?
• Procurement: Which defaults, renewals and usage terms change without a new implementation project, and who reviews them?
• Vendors and partners: Can you show the complete trace, actual bill and rejected cases behind the claimed outcome, and can we export them?
• Business units: What counts as completion, who signs it off, and which exception would cause you to withdraw delegated authority?

24 Contrarian conclusion: owning acceptance may matter more than owning the interface

The visible contest is for the place where employees ask an agent to work. September's evidence suggests another source of durable value: the rules and context that determine whether the work is acceptable. Application vendors can retain influence even as their interface recedes. Buyers can retain bargaining power when their acceptance standards, business definitions and operating evidence remain intelligible outside one vendor's stack.

This is an inference, not an established market outcome. The practical prediction is testable: in a matched enterprise workflow, investment in owned context, a clear acceptance rubric and usable execution controls should improve quality-adjusted economics enough to justify its maintenance cost. If a model or interface upgrade reliably produces the same improvement without those complements, the claim should be weakened.

The buyer's scarce capability may be the ability to say precisely what good work means, recognize it, and assign responsibility when it fails. September made more execution available. It did less to relieve organizations of that obligation.

25 What would change the conclusion in October?

The October watch list is a set of observable tests, not a promise that each result will arrive. No post-cutoff evidence has been used to resolve them.

26 Sources, methods and limitations

The research log contains 45 screening records, including one explicit coverage gap; 44 concern concrete developments or research candidates. The source ledger contains 32 external records and one authored scenario record. Two restricted-retrieval records are retained for auditability, not promoted into numerical evidence. Material claims are mapped in claims.csv and radar.json; illustrative calculations are identified as scenarios and have no external-source claim.

The evidence window is the calendar month of September in Europe/Berlin. A September publication may describe an earlier incident, quarter or experiment. Product documentation is mutable; this edition freezes the dated September entries retrieved on 1 October. Future effective dates are recorded separately. Exact days are not invented where a paper or release history supplies only a month. The ServiceNow page's 7/8 September date discrepancy is retained in the ledger.

Primary documentation establishes what a vendor or institution states. It does not automatically establish adoption, safety, causality or customer value. Customer measurements, synthetic experiments, causal field studies and official financial disclosures receive different interpretations. No population adoption rate, aggregate CIO index, general vendor ranking or economy-wide productivity effect is inferred from this evidence set.

Continuity uses the complete August package, including its machine-readable scores and frozen forecast. The live August archive could not be retrieved during research, so package-to-live consistency remains unverified. Only that complete previous edition was available for detailed visual comparison; the design history is not represented as a six-month audit.

Source coverage includes Microsoft, AWS, Google, enterprise applications, data platforms, named customer cases, research, incident analysis, regulation and European/German deployment options. A transparent cross-industry September CIO survey and a sufficiently verified September completed acquisition were not established. These gaps limit population and transaction-market conclusions; they do not justify filling the article with estimates.

The seven figures carry exact data, source IDs and assumptions in the package. Research estimates preserve their published definitions and versions. Raw WorkWorlds ratios are reproduced separately from standardized estimates. The patent figure does not fabricate confidence intervals from p-values. The runtime chart preserves reported ranges without constructing unmatched speedup ratios. The cost model is fully reproducible and explicitly illustrative.

1. S01

   MicrosoftIntroducing the new Copilot with Home, Code, and Autopilot

   2026-09-25vendor announcementClaims C01, C02

   Home, Code and Autopilot have different release states; launch does not establish tenant access or benefits.

2. S02

   MicrosoftEvolution of the Copilot pricing model

   2026-09-25vendor pricing explanationLimited retrieval

   September 28 update; cached search text available, body not retrievable. No price comparison or rate used.

3. S03

   Microsoft LearnSeptember 2026 Partner Center announcements

   2026-09-16official product documentationEffective 2026-11-02Claims C01, C06

   November 2 default billing applies to NEW Copilot Business purchases through CSP. Preset limit amount not verified.

4. S04

   Microsoft LearnMicrosoft 365 Copilot release notes

   2026-09-23official product documentation

   Release summary spans August 26–September 22; gradual deployment means publication date is not universal tenant availability.

5. S05

   Microsoft LearnWhat is new in Copilot Cowork

   2026-09official product documentation

   September changes separated from June GA and August browser/event task entries. Month precision only.

6. S06

   MicrosoftNew Microsoft data innovations unlock what only your business knows

   2026-09-28vendor announcementClaims C03

   Fabric IQ GA in Chat/Cowork; IQ Sharing preview. Existing permissions do not guarantee correct entitlements.

7. S07

   GitHubEnterprise-managed permissions for GitHub Copilot agent operations

   2026-09-09official release notesClaims C04

   GA policy enforcement for shell, file and network operations; no independent containment test in this release.

8. S08

   GitHubDefault enablement of Copilot features for Copilot Business and Enterprise

   2026-09-24official release notesEffective 2026-10-22Claims C05

   Effective October 22 for eligible unconfigured GA feature policies. Explicit settings preserved; previews remain opt-in.

9. S09

   GitHubUpcoming deprecation of selected GitHub Copilot models in mid-October

   2026-09-18official release notesEffective 2026-10-19Claims C07

   Retirement date October 19. Version substitutions need buyer-side regression tests.

10. S10

    AWSNew AgentCore Runtime generally available

    2026-09-18official release notesClaims C08

    V2 runtime available in five named regions. Vendor cold-start P75 ranges are not end-to-end process productivity.

11. S11

    AWSThe new AgentCore Runtime: elastic, optimized, and consistently fast starts

    2026-09-18vendor technical benchmarkClaims C08

    5,000 cold invocations per agent across five sizes; lightweight echo workload. V1/V2 range endpoints not paired for ratios.

12. S12

    AWS DocumentationAmazon Bedrock AgentCore release notes

    2026-09official product documentation

    September shells/evaluators/consent changes; Registry GA belongs to August and is continuity only.

13. S13

    Google CloudGemini Enterprise release notes

    2026-09official product documentationClaims C09, C10

    Dated entries September 1–28. Resource IAM, ownership, write actions, billing and Code Assist have different scopes.

14. S14

    Google CloudGemini Enterprise Agent Platform release notes

    2026-09official product documentationClaims C11

    September 9 Computer Use and Shell sandboxes GA. Perimeter controls support rather than prove safe workflows.

15. S15

    SalesforceSalesforce unveils AIforce

    2026-09-15vendor announcementClaims C12

    Salesforce in Claude beta; activated agents are activity. No independent benefit estimate.

16. S16

    SalesforceSiemens and Salesforce deepen AI partnership

    2026-09-15vendor customer caseClaims C13

    100% inbound engagement across 132 countries for 18,000 sellers. New Teamcenter integration and onboarding promises separated from current deployment.

17. S17

    SmarshSmarsh scales Agentforce after strong customer support results

    2026-09-03customer primary caseClaims C14, C15

    Q2 observations published September; 405 Archie interactions. Employee use and reported time savings not randomized or audited.

18. S18

    Adecco GroupAdecco Group and Salesforce expand Agentforce to more than 40 countries

    2026-09-15customer primary caseClaims C16

    27,000 employees is rollout population, not active users. 2.5m candidate interactions accumulated since April 2025.

19. S19

    SAPThe autonomous enterprise: business transformation management solutions

    2026-09-22vendor announcementClaims C17

    Company Memory preview; LeanIX/Signavio/WalkMe/Cloud ALM positioning has no causal outcome evaluation here.

20. S20

    SAPTabPFN-3.5-Plus now available in SAP AI Core

    2026-09-15official product announcementClaims C18

    Structured-data prediction option. Prior Labs acquisition completed July, not a September transaction. Benchmark superiority not independently checked.

21. S21

    ServiceNowT-Systems goes live as Sovereign Partner Cloud provider in Germany

    2026-09-07/08vendor partner announcementClaims C19

    Page header September 7, body September 8. German hosting/service does not establish sovereignty of every model, support or recovery dependency.

22. S22

    OracleOracle Q1 fiscal 2027 earnings release

    2026-09-10official earnings releaseClaims C20

    Quarterly company results, not September-only activity; RPO is future contracted obligation, not recognized revenue or buyer ROI.

23. S23

    AnthropicAlignment assessment: cybersecurity incidents

    2026-09-09first-party incident investigationClaims C21

    Four incidents in reduced-safeguard third-party evaluation settings. January incident identified August, disclosed September; three already disclosed July. METR investigation planned, not independent verification completed.

24. S24

    European CommissionCyber Resilience Act reporting obligations

    undatedofficial regulatory guidanceEffective 2026-09-11Claims C22

    September 11 reporting obligations for covered manufacturers. Scope and reporting routes require product-specific review; other main obligations December 2027.

25. S25

    Autor et al., MIT/NBERDoes AI assistance enhance or erode expertise?

    2026-09research working paperClaims C23, C24

    Pre-registered 3-month field experiment, 133 patent lawyers at 11 US firms. Attrition, task and seniority limits; no independent replication claimed.

26. S26

    WorkWorlds authors / arXivWorkWorlds: evaluating agents in realistic enterprise environments, v2

    2026-09-23research preprintClaims C25, C26

    v2 September 23 supersedes v1 September 20. Synthetic PharmaCo, 192 runs, 8 tasks, 6 seats. Standardized estimates and raw ratios kept separate.

27. S27

    DatabricksUnity Catalog Pages: a governed home for business knowledge

    2026-09-22vendor announcementClaims C27

    Governed business definitions and sources. No measured retrieval or enterprise performance lift in announcement.

28. S28

    Mistral AIMistral AI and Cloudera partnership

    2026-09-10vendor partnership announcementClaims C28

    Deployment flexibility planned across customer environments; not a customer outcome or proof every integration is available.

29. S29

    Snowflake / LSEGLSEG and Snowflake expand collaboration

    2026-09-30joint partnership announcementClaims C29

    Five-year agreement and data distribution. No disclosed financial commitment amount or causal productivity result.

30. S30

    OpenAIIntroducing the Agents API

    2026-09-10vendor product announcementClaims C30

    Managed persistent execution using Codex harness. Customer evaluation testimonial lacks independent protocol; not used as headline benchmark.

31. S31

    CEPRCEPR Discussion Paper 21944

    2026-09-15research abstractLimited retrieval

    Market-implied productivity expectations and modelled GDP channels are not realized enterprise productivity. Screened but not used for quantitative conclusions.

32. S32

    Imai, Inan, Alikhani / arXivFrom Task Success to Productive Success

    2026-09research abstractLimited retrieval

    Cost comparison discovered; full methods not verified. Excluded headline cost ratio.

33. S33

    Michael Schymura / editorial modelIllustrative CPVO scenario assumptions

    2026-10-01authored scenarioClaims C31, C32

    Packaged data/radar-data.json and code/reproduce.py define all assumptions. Creation date after cutoff is production metadata, not post-cutoff external evidence.