Audits / Autonomy Economics Evidence Audit for September 2026
Monthly evidence audit
Autonomy Economics Evidence Audit for September 2026
What the register recorded between 1 September 2026 and 30 September 2026.
Derived from the register and published without editorial approval on 6 September 2026.
31
observations added
19
strengthened
0
weakened
What weakened
Listed first, deliberately. These figures are derived from the register and not written, so this section cannot be softened before it publishes, which is the whole reason the audit is generated instead of drafted.
Nothing weakened or was qualified in this period.
Counter-signals
No new contradictory signals were recorded in this period.
Boundary conditions observed
Salesforce Q2 results show strong durability during commercial and architectural redesign (anticipated boundary)
Bank of England records rapid gains in agentic task capability and governance constraints (anticipated boundary)
UK consumers retain purchase control when using AI shopping agents (anticipated boundary)
Salesforce reports strong growth while agentic architecture expands (anticipated boundary)
Most law firms report unchanged pricing despite clients expecting AI-driven commercial change (anticipated boundary)
Futurum survey finds buyers still prefer per-seat pricing for separately billed generative AI (anticipated boundary)
Visa survey finds only 23% trust GenAI to make payments on their behalf (anticipated boundary)
Microsoft prices Agent 365 at $15 per user despite the product governing autonomous agents (anticipated boundary)
Too early to judge
No new observations were classified as too early in this period.
What strengthened
Salesforce explicitly prices Help Agent on successful outcomes (strength 5)
Wall Street banks push Big Law to pass AI productivity into lower fees (strength 5)
FT reports consulting clients cutting external work and pressing fees as AI expands internal capability (strength 5)
Salesforce and Anthropic announce Claudeforce (strength 5)
Salesforce expands Headless 360 so agents can invoke enterprise capabilities directly (strength 5)
The Economist reports that China's AI drive threatens the world's largest workforce (strength 5)
EY survey finds C-suites pivoting from AI adoption to unlocking value as token costs trigger fiscal scrutiny (strength 5)
Worldline, ING and Mastercard complete live end-to-end European agentic payment (strength 5)
OpenAI launches Deployment Company with major capital partners (strength 5)
Anthropic and private-equity consortium form enterprise AI services company (strength 5)
Blackstone links realised AI productivity gains to early repricing and a shift from seats to outcomes (strength 4)
Atlassian prices customer-service AI at $1 per successful resolution (strength 4)
FCA finds enterprise value depends increasingly on the AI harness around the model (strength 4)
Atlassian introduces metered Rovo credits for AI and Teamwork Graph usage (strength 4)
FSB elevates autonomous-AI control and resilience into G20 financial-system discussion (strength 4)
Salesforce documentation states that Help Agent pricing is strictly outcome-based. Unresolved sessions are not billed, while successful resolutions are billable outcomes.
Law 2Law 3Commercial alignment
Prediction testedAs autonomous systems perform work, value capture can migrate from access and activity units toward successful outcomes.
AssessmentSupportiveDirectly observable
Why it mattersThe outcome itself becomes the billing unit, while a verification rule determines whether payment is triggered.
The FT reported that Morgan Stanley, Citigroup and Goldman Sachs are pressing major law firms to reflect AI-enabled efficiencies in fees, including through competitive bidding, fixed fees and requests for evidence of AI-driven savings.
HBPLaw 2Law 3Profit ParadoxRevenue Durability
Prediction testedCommercial-unit pressure should appear once buyers can observe that less human progression is required, even if demand for high-value legal judgement remains.
AssessmentSupportiveDirectly observable
Why it mattersThis is a direct buyer response to productivity change that is already happening, with no reliance on forecasts about future legal pricing.
The FT reported that companies are bringing more technology work in-house, reducing reliance on large consulting teams and demanding lower, fixed or performance-linked fees as AI reduces the human effort required for implementation and analysis.
HBPCSDLaw 2Law 3Profit ParadoxRevenue Durability
Prediction testedConsulting can remain in demand while billable progression becomes less defensible and value moves toward senior judgement, leadership advice and measurable outcomes.
AssessmentSupportiveDirectly observable
Why it mattersThis is buyer behaviour, which carries more weight than a vendor prediction. Clients are connecting AI-enabled productivity directly to lower consultant dependence and pricing pressure.
Salesforce and Anthropic announced Claudeforce, making Salesforce data, workflows, business logic, actions and governance accessible directly in Claude, initially through a plugin with 37 prebuilt sales skills.
HBILaw 1Law 3
Prediction testedExternal agents can become the user-facing interaction layer while the incumbent protects economic relevance by becoming the governed execution and context layer underneath.
AssessmentSupportiveDirectly observable
Why it mattersThe incumbent is voluntarily allowing a third-party agent to become the interface, giving up the attempt to preserve UI scarcity.
Salesforce expanded Headless 360 across its platform, exposing reusable enterprise capabilities through MCP, developer tools and headless experiences so authorised AI agents can discover and take action without relying on conventional application interfaces.
HBILaw 1Commercial alignment
Prediction testedAs agents become an interaction layer, interface dependence weakens and durable value can migrate toward underlying systems of record, workflow and governed action.
AssessmentSupportiveDirectly observable
Why it mattersThe vendor is explicitly designing for software to remain valuable even when humans do not navigate its interface directly.
ByteDance released Seedance 2.0 AI model in February 2026. Microdramas now take 'days to produce instead of weeks' and are 'up to 90% cheaper (the bill is mostly tokens, instead of humans).' Li Dazhi's daily salary 'fell by half' and 'gigs dried up.' He and colleagues switched to unrelated work (insurance sales, vendor stalls, food delivery, rideshare). The microdrama industry had 700,000 direct j
HBPCSDProfit Paradox
Prediction testedTests HBP and CSD, asking whether humans carrying work through stages (scriptwriting, shooting, editing and post-production) can be replaced by a single autonomous system, and whether the scarcity of human video production skill is declining. The text shows both. Production time collapsed, cost per unit fell 90%, and workers were displaced. Worker transition to lower-wage work implies that video production expertise no longer commands the same wage premium.
AssessmentSupportiveDirectly observable
Why it mattersThis demonstrates HBP, where human progression through stages of work is now replaced by autonomous capability. Video production that required weeks of human labour and human skill progression now requires days and is tokenised. It shows CSD (the scarcity of video production and editing expertise, which supported job creation and wage income, has declined). The economic consequence is measurable. Worker income halved, work unit cost fell 90%, and workers were displaced into lower-skill service work. This is a shift in the structure of how work is compensated and organized.
98% of senior leaders using AI tools that require tokens say token usage costs caused their organisation to reconsider their approach. 82% are concerned about AI token usage and costs. Only 64% actively monitor token usage and have clear budgetary guardrails. 76% say off-the-shelf software solutions no longer meet their needs. 91% view in-house AI-built software as critical. 87% have fully deploye
Law 2Profit ParadoxRevenue DurabilityCapital-Structured Autonomy
Prediction testedEnterprises actively building in-house autonomous systems in response to consumption cost structures incompatible with traditional SaaS licensing. Per-seat pricing model anticipated to lose relevance. Spending projections missing targets suggest commercial structures are not sustaining projected revenue.
AssessmentSupportiveDirectly observable
Why it mattersToken consumption costs are driving enterprises away from vendor commercial software toward in-house autonomous systems, and consumption-based cost drivers are making traditional per-seat licensing incompatible with autonomous activity structures. Value capture is restructuring around cost governance of autonomous throughput, and less around access.
The firms announced Europe’s first live end-to-end agentic payment transaction in production, with an AI agent initiating an authenticated payment using existing payment infrastructure.
HBILaw 1Law 3
Prediction testedAgents can mediate economic interaction and execute transactions without requiring the user to navigate each conventional interface step.
AssessmentSupportiveDirectly observable
Why it mattersThe HBI mechanism is observable in production as well as in product concepts.
OpenAI launched the OpenAI Deployment Company to embed forward-deployed engineers in organisations, agreed to acquire Tomoro, and described a model for scaling deployment across the economy with capital and consulting partners.
Law 1Capital-Structured Autonomy
Prediction testedValue can move above individual operating companies toward actors that coordinate autonomous deployment, learning and workflow redesign across wider networks.
AssessmentSupportiveDirectly observable
Why it mattersThe organisation is explicitly designed around repeatable deployment capability, and model access is only one part of it.
Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced a new AI services company for mid-sized enterprises.
Law 1Capital-Structured Autonomy
Prediction testedCapital owners can organise deployment, engineering capability, standards and learning across networks of firms, instead of each operating company adopting independently.
AssessmentSupportiveDirectly observable
Why it mattersThe structure closely matches the paper’s predicted ecosystem-level orchestration model.
In a September 2026 investor presentation, Blackstone President and COO Jon Gray described substantial realised AI productivity gains across portfolio companies, then highlighted multiple compression in professional services, software and information-services businesses even where underlying operations remained sound, and said management teams that move the model from seats to outcomes should be better positioned.
Prediction testedWhether autonomous capability can generate substantial operating gains while markets simultaneously question the durability of legacy commercial structures, and whether the proposed response shifts value capture away from seat-based access toward outcomes.
AssessmentSupportiveStrong
Why it mattersIt combines three parts of the proposed mechanism in one post-publication primary-source presentation, namely realised productivity gains, repricing before broad operating deterioration, and an explicit recommendation to move from seats to outcomes.
Atlassian states that Customer Service Management AI agent resolutions will be charged at $1 per successful resolution from 3 December 2026, only when the AI fully resolves a customer request without handing it to a human.
Law 2Law 3Commercial alignment
Prediction testedAutonomous software can move value capture from user access toward successful outcomes, with verification determining when the economic event has occurred.
AssessmentSupportiveStrong
Why it mattersA second major SaaS vendor is independently adopting successful-resolution pricing, making the pattern less idiosyncratic.
Demand for law firm hours rose 4.2% through the first half of 2026 against a normal rate of about 1.5%, while two-thirds of firms reported daily associate use of AI and clients pressed to move away from the billable hour.
HBPLaw 2Revenue DurabilityCommercial alignment
Prediction testedTests whether realised changes in human progression are visible alongside pressure on effort-based pricing, while recognising that aggregate hours are not the same measure as effort per unit of work.
AssessmentAmbiguousStrong
Why it mattersIt is the clearest counterweight on the register to the legal pricing-pressure records. Billed hours rising at nearly three times the normal rate, while AI use is near-universal inside the same firms, is not what a simple reading of the mechanism predicts. It forces the sequence to be stated more carefully. Buyer pressure can run ahead of any measurable change in activity, driven by expectation and by the visible profitability of the supplier instead of an observed reduction in effort.
The FCA published a multi-firm review focused on frontier AI, cyber resilience, harness engineering, governance and vulnerability management, highlighting that firms must structure controls and operating environments around frontier models.
CSDLaw 1Law 3
Prediction testedAs model capability diffuses, value and operational scarcity should move toward the systems that organise, constrain, validate and absorb that capability.
AssessmentSupportiveStrong
Why it mattersA regulator is examining the harness and control environment as a distinct operational layer, and does not treat the model alone as the relevant system.
The FT reported that PE-backed software companies are using shorter amend-and-extend transactions, paying higher yields and accepting stronger creditor protections as lenders assess AI-related threats to long-term software business-model durability alongside leverage and maturity risks.
CSDHBIRevenue Durability
Prediction testedMarkets and creditors can reprice the durability of exposed commercial structures before full operating-model or revenue deterioration is visible.
AssessmentAmbiguousStrong
Why it mattersIt suggests a transmission channel from autonomy exposure into cost of capital and creditor protections.
Atlassian documents Rovo credits as a meter for AI usage and enriched Teamwork Graph context, including calls made through the Rovo MCP server and CLI.
HBICSDLaw 1Law 2
Prediction testedAs agents, in place of users, consume software capability, the seat becomes a weaker standalone measure of value and new activity and context units emerge.
AssessmentSupportiveStrong
Why it mattersAtlassian is monetising agent and third-party access to underlying context even when work takes place outside the conventional product interface.
The FSB Chair warned G20 finance ministers and central bank governors that increasingly autonomous frontier AI may materially change the speed, scale and economics of cyber risk and stressed resilience and responsible deployment.
Law 3
Prediction testedThe layer that constrains and governs autonomous throughput should become more important as the scale and consequences of autonomous action increase.
AssessmentSupportiveStrong
Why it mattersGovernance is moving from enterprise policy into systemic-risk architecture.
Vercel's September 2026 AI Gateway Production Index shows open-weight models processing 56% of gateway token volume in August while accounting for 14% of estimated spend. Average price per token fell 23.2% during the month, while Anthropic retained 64% of gateway spend. The same dataset therefore shows both rapid diffusion of lower-cost capability and persistent concentration of premium spending.
CSD
Prediction testedTests whether broad availability of cheaper and open-weight capability erodes scarcity rents, or whether premium frontier capability continues to sustain concentrated spending.
AssessmentAmbiguousStrong
Why it mattersThe same production dataset contains evidence in both directions. It is therefore useful for defining where CSD may weaken first and where premium capability can remain scarce enough to command materially higher spending.
Salesforce reported $10.8 billion in subscription and support revenue, up 12% year on year, nearly $3.9 billion in Agentforce and Data 360 ARR, 7 billion Agentic Work Units delivered to date, and reorganised revenue disclosure around Agentforce Apps and Data 360, Headless Platform and Other.
Prediction testedFirms that realign commercial structure with autonomous activity can remain economically durable while interface dependence and conventional activity units change.
AssessmentBoundary conditionStrong
Why it mattersIt is a live counterexample to a simplistic 'agentic AI destroys SaaS' thesis and therefore a useful positive test of the redesign proposition.
DeepSeek released an open-source agent harness in developer preview under the MIT licence, built around an everything-is-a-plugin architecture.
CSDLaw 1
Prediction testedAs autonomous capability diffuses through open systems, scarcity should move away from raw access to the capability and toward how it is integrated, governed and applied.
AssessmentSupportiveStrong
Why it mattersOpen agent infrastructure broadens access to execution capability as well as to model inference.
Deloitte’s survey of 121 senior legal leaders found 61% in AI deployment phases, 61% experimenting with or piloting agentic AI, 78% wanting external-provider AI to reduce costs, and 85% expecting AI to change law-firm pricing.
HBPCSDLaw 2Law 3Revenue Durability
Prediction testedWhen structured legal production requires less human progression, billable-hour durability should come under pressure while judgement and accountability remain more defensible.
AssessmentSupportiveStrong
Why it mattersThe evidence connects adoption to expected change in the commercial model as well as to productivity.
The July Financial Stability Report said frontier models can sustain longer, more complex multi-step tasks with less human intervention, while also warning that agentic workflows can be costly and that autonomous payments raise questions around authorisation, traceability, liability and governance.
HBPLaw 3
Prediction testedTechnical maturity changes the timing of commercial pressure, while governance and control become more important as autonomous execution expands.
AssessmentBoundary conditionStrong
Why it mattersAn independent central bank is observing both rising autonomy and the constraints that determine when it becomes economically deployable.
KPMG and Microsoft announced global deployment of Agent 365 and Copilot, including tooling to manage, monitor and secure AI agents across KPMG and client organisations, alongside Copilot deployment to more than 276,000 professionals.
HBPLaw 1Law 3
Prediction testedAs autonomous throughput expands, value and organisational importance move toward monitoring, control, verification and governed execution.
AssessmentSupportiveStrong
Why it mattersGovernance is being operationalised as infrastructure for enterprise-scale agent deployment.
Salesforce reported record first-quarter results, more than $1 billion in Agentforce ARR, $3.4 billion in combined AI and data ARR, 3.8 billion Agentic Work Units delivered, and more than one million active Slack MCP users within six weeks of launch.
Prediction testedFirms that redesign around autonomous activity can preserve durability even while the old interaction structure weakens.
AssessmentBoundary conditionStrong
Why it mattersIt prevents the theory from being misread as 'AI must cause SaaS decline' and provides a live comparator for redesign versus non-redesign.
Thomson Reuters' 2026 Future of Professionals survey, conducted in March and April before the Autonomy Economics publication date, found a clear gap between client expectations and law-firm pricing response. Seventy-one per cent of in-house legal professionals expected outside firms to change their commercial models as AI usage increased, while only 28% of law firms reported pricing changes and 62% said pricing remained unchanged. Because the underlying survey predates publication, this record is treated as contemporaneous background and does not count as post-publication evidence.
HBPLaw 2Commercial alignment
Prediction testedTests whether AI-enabled changes in professional work translate into changes in effort-based commercial units and whether commercial structures adapt alongside activity structures.
AssessmentBoundary conditionStrong
Why it mattersIt provides direct survey evidence that client expectations and law-firm pricing structures are currently out of alignment. That gap is exactly where a boundary observation is useful. Pressure is observable while redesign has not become dominant.
The FT reported continued high redemption requests at Blackstone’s large retail private-credit fund and linked investor concern partly to exposure to leveraged software companies facing uncertainty over AI’s impact, while also identifying broader private-credit pressures.
CSDHBIRevenue Durability
Prediction testedAutonomy exposure can influence valuations and capital structures before operating revenue fully resets, including through credit and portfolio channels.
AssessmentAmbiguousModerate
Why it mattersIt may represent a second-order capital-market consequence of software durability uncertainty.
Futurum's 2H 2026 Enterprise Software Decision-Makers Survey finds a split in buyer pricing preferences. Per-seat pricing ranks last for core software at 12.6%, but first for separately billed generative AI functionality at 42.3%, ahead of consumption pricing at 36.6% and outcome pricing at 21.1%.
Law 2
Prediction testedTests whether buyers move consistently away from access-based pricing as autonomous capability expands, or whether uncertainty can preserve seat pricing in some categories.
AssessmentBoundary conditionModerate
Why it mattersThe same buyers express different commercial preferences for core software and AI. That makes the observation more useful than a simple vendor pricing example, because it shows that the commercial unit can depend on the risk characteristics of the category as well as on the underlying technical capability.
Visa's Trust Index found that only 23% of US consumers trusted generative AI to handle payment transactions on their behalf. Trust rose to 61% when Visa was named as the party securing the transaction. The survey was fielded from 26 to 28 May 2026, after publication of Autonomy Economics.
HBI
Prediction testedTests whether autonomous systems displace human interaction at economically consequential transaction points, as well as during search and recommendation.
AssessmentBoundary conditionModerate
Why it mattersThe evidence is useful because it measures consumer willingness after the framework was published and focuses on the payment decision itself. It narrows any reading of HBI that assumes rapid removal of human approval throughout the commerce journey.
Microsoft made Agent 365 generally available as a control plane for AI agents while retaining a per-user commercial unit. The standalone product is priced at $15 per user per month, with each licence covering an individual who manages or sponsors agents, or uses agents to do work on their behalf.
Why it mattersThe observation is useful because the product sits close to the autonomous control layer and still remains priced per user. It narrows any overly simple reading of Law 2 that would expect agentic products to abandon user-based pricing immediately.