AI agent adoption: demand, evidence and the measurement problem

On this page

Chapter 2 of The Hitchhiker's Guide to the Agentic Economy: Australia 2026

Report hub · Previous: what is the agentic economy? · In this chapter: current evidence · Australian baseline · measurement framework · job record

The measurement deficit

Commercial adoption in enterprise technology is counted in work delivered, accepted, and paid for. In the emerging market for autonomous software, almost none of the published numbers measure that sequence.

Instead, current market reporting mixes four completely different operational stages under the single headline of adoption: product availability in cloud catalogues, developer experimentation with model APIs, self-reported executive interest in surveys, and production systems executing real business processes. A vendor announcing that an agent feature is generally available is reporting product supply. A survey where corporate managers report exploring agents is measuring exploratory sentiment. Developer platform metrics tracking token burn or tool invocations record compute consumption. None of these signals establish that an organization has delegated an ongoing business process to software and accepted the financial and operational consequence.

Measuring genuine economic adoption requires tracking the full commercial path: what job was delegated, what authority was granted, what external actions were executed, whether the output satisfied business acceptance criteria, what human review was required to fix mistakes, and whether the customer returned to use the service again.

Key Takeaways

• Published studies use different denominators for agent adoption in 2026; each supports a different claim.
• Product availability, reported use and accepted economic work are different market events.
• Australian evidence points to early, piecemeal use and mixed realised returns among the medium and large firms consulted by the RBA.
• Transaction, token and tool-call counts can show activity without proving independent demand, successful delivery or profit.
• The strongest operating measure is the accepted outcome per unit of cost and human intervention.

The denominator problem

Headline adoption statistics routinely conflate distinct operational boundaries. Survey research tracks what corporate leaders intend to buy. Cloud telemetry records event volumes inside a single vendor's infrastructure. Task benchmarks measure success rates across synthetic programming puzzles. Payment networks track authorization requests, while software marketplaces report published catalog entries.

These signals describe different layers of an emerging market. They cannot be summed into a single percentage. The word adoption is currently used to describe seven distinct operational states:

1. An employee using an interactive conversational assistant to summarize text.
2. A team incorporating an AI writing or coding copilot into an existing manual workflow.
3. A deterministic workflow script triggering an LLM API call for text transformation.
4. An agent dynamically selecting tools and parameters under synchronous human approval.
5. An unattended production system executing multi-step tasks across internal software.
6. An autonomous agent initiating an external purchasing transaction or contract.
7. A supplier receiving recurring, commercially accepted orders from external software buyers.

These states do not form a uniform maturity curve. A business may run high-volume assistant usage across its workforce while prohibiting any autonomous external actions. Another firm may operate a mission-critical, unattended data pipeline without using conversational assistants at all. Similarly, a merchant platform processing agent-initiated payments does not establish that underlying suppliers are generating sustainable commercial margins.

Evaluating any market claim requires answering two structural questions:

• What specific event was counted? A user prompt, an active seat, an API call, an automated action, a payment authorization, a delivered file, or a formally accepted business result?
• What was the denominator? Survey respondents, cloud customer accounts, total API invocations, unique corporate buyers, or completed production jobs?

Without both parameters explicitly defined, percentage adoption figures obscure more than they explain.

What the current evidence establishes

Empirical market evidence is currently concentrated in four distinct categories:

Provider documentation establishes product capability

Model providers and hyperscalers document extensive agentic infrastructure: function-calling interfaces, multi-agent frameworks, computer-use capabilities, managed sandboxes, and stateful execution runtimes. Protocol developers publish open specifications and reference implementations for tool discovery and task coordination. Payment processors document machine checkout flows, shared credentials, and token settlement mechanisms.

This documentation proves that specific technical interfaces exist and can be called under designated terms. It does not establish how many enterprise buyers have deployed them, how reliably they operate under messy production conditions, or whether the resulting workflows generate positive economic returns.

Independent evaluations measure bounded technical capability

Evaluation organizations test models on structured, synthetic problem sets. Research on task-completion time horizons conducted by METR measures the difficulty of software engineering tasks that frontier models can resolve with defined probability (METR, "Task-Completion Time Horizons of Frontier AI Models").

These evaluations provide clear visibility into task definitions, test boundaries, and failure rates. However, technical performance on an isolated programming test does not demonstrate whether an agent can navigate ambiguous business goals, operate across enterprise legacy systems, or operate safely within unstated corporate policies.

Platform telemetry records behavior within closed environments

Platform telemetry provides visibility into customer usage patterns inside specific vendor ecosystems. Enterprise reporting from OpenAI describes an ongoing shift from information retrieval toward task execution across its business customer base (OpenAI, "From assistance to execution"). Research by Anthropic analyzes real-world deployment telemetry to study human oversight patterns and effective agent autonomy (Anthropic, "Measuring AI agent autonomy in practice").

These datasets provide detailed views of session duration and tool utilization across large user cohorts. However, they reflect the activity of self-selected customer bases within specific software tools. They rarely capture the downstream human review, manual data cleaning, or cross-system reconciliation required to convert platform output into an accepted commercial deliverable.

Surveys capture executive sentiment and planned investment

Enterprise surveys track corporate expectations and self-reported experimentation. A 2026 McKinsey study reported that 40 per cent of respondents at large organizations claimed their companies were scaling AI agents, up from 27 per cent in the prior year, while smaller organizations reported an adoption rate of 22 per cent. Crucially, the same study revealed that only 37 per cent of all respondents saw any positive EBIT impact from AI investments, and just 6 per cent qualified as high performers under McKinsey's criteria (McKinsey, "The State of AI in 2026: On the road to ROI").

The value of this data lies in the gap between the headline scaling figure (40 per cent) and the realized financial return (37 per cent seeing any EBIT contribution). It reflects widespread capital allocation and active experimentation, alongside significant friction in capturing net economic value.

The Australian enterprise baseline

The Reserve Bank of Australia surveyed 105 medium and large firms between June and August 2025 during its liaison program on technology and artificial intelligence investment. The RBA found that AI deployment remained largely piecemeal, practical adoption of autonomous agents was low, and realized financial returns were mixed, even though many participating executives anticipated substantial long-term productivity gains (RBA, "Technology investment and AI: what are firms telling us?").

This finding provides the most rigorous empirical baseline currently available for the Australian enterprise sector. Its boundaries are clear: it reflects medium and large corporate respondents that agreed to participate in central-bank liaison interviews. It does not measure small businesses, professional partnerships, or specialized service firms where most Australian service transactions occur.

The RBA evidence reveals a consistent structural condition: global tool availability is running well ahead of operational enterprise adoption and measurable commercial payback.

Australian market adoption is also constrained and enabled by domestic institutional infrastructure:

• Financial and payment rails: The New Payments Platform (NPP), PayTo mandates, and the Peppol eInvoicing framework provide the domestic rails needed to bind digital payments to verified accounting records.
• Regulatory governance: Privacy guidance from the Office of the Australian Information Commissioner (OAIC), governance standards from the Australian Securities and Investments Commission (ASIC), and security frameworks from the Australian Signals Directorate (ASD) define the legal compliance and cyber-security boundaries within which Australian firms must operate.

These institutional systems govern risk and liability; they do not by themselves generate commercial demand.

Market-measurement framework

To evaluate claims consistently, this report structures market evidence across three distinct dimensions:

Dimension	Values used in this report	Why it matters
Observed event	Specified, announced, documented, available, observed, invoked, paid, delivered, accepted, reconciled, repeated	Identifies how far through the operational lifecycle the transaction progressed
Source origin	Provider, customer, Agentic Economy test, third-party researcher, official body	Identifies who generated and controls the underlying data
Study design	Product telemetry, reproducible test, survey, interview, case record, audit, accounting record, census	Identifies the sample population, methodology, and inherent blind spots

A protocol can be fully drafted without a single production deployment. A cloud vendor can expose a service in documentation without disclosing active enterprise customer counts. A company can operate a proprietary workflow that delivers accepted work without publishing public case studies. Evaluating market adoption requires separating what is specified from what is commercially settled.

Each market claim in this report tracks six core fields: claim ID · event · source origin · study design · date/version · limitation.

Distinguishing technical activity from economic demand

Digital systems generate extensive machine-readable event logs. Because these events are easily quantified, they are frequently misinterpreted as direct measures of economic demand.

Tool invocations

A tool invocation records an attempt by an agent to query a database, call an API, or execute code. Complex tasks frequently trigger multiple recursive tool calls, failed executions, and retries. An elevated invocation count often reflects workflow inefficiency, brittle error-handling, or poorly bounded context rather than increased customer demand.

Token consumption and compute hours

Token usage measures raw infrastructure utilization. It tracks operational cost, not commercial value. High token burn can indicate a runaway reasoning loop or excessive context padding as easily as it indicates a valuable business deliverable.

Directory listings and endpoints

A catalog entry in an agent registry or an open API endpoint confirms discoverability. It does not prove that an autonomous agent can authenticate, pass corporate security gates, authorize payment, receive usable data, or recover cleanly from unexpected downtime.

On-chain and transaction records

A payment event records an asset transfer. On its own, a transaction hash does not prove whether the payment was a commercial purchase, an automated synthetic test, a vendor subsidy, an internal service loop, or a retry following a failed attempt. Independent directories tracking machine-payment protocols like x402 provide valuable visibility by actively probing live endpoints, but their maintainers explicitly note high concentration and coverage limitations (x402 List methodology).

Strategic partnership announcements

Vendor announcements of ecosystem partnerships indicate distribution intent. They do not prove that partner engineering teams have shipped production-grade integrations, or that joint customers are using the integration for consequential business workloads.

Operational metrics must be matched to the exact stage of the lifecycle they record.

The job-level measurement record

Evaluating the operational performance of an agent system does not require a nationwide dataset. It requires linking existing operational, human, and financial data for a single workflow.

An accepted job is a deliverable that an authorized business owner or automated verification rule has evaluated and confirmed complete against explicit quality and scope criteria.

Field	Operational data to capture	Why it changes the business decision
Objective	Scope, client identifier, and explicit acceptance test	Separates raw activity from verified completion
Agent and principal	Running software workload and sponsoring organization	Identifies legal and corporate accountability
Permitted choice	Specific actions and tools the system could select	Distinguishes autonomous agency from fixed automation
Authority	Data access boundaries, budget ceiling, time limits, and approval rules	Defines the maximum potential financial and operational exposure
Job and events	Stable job ID, timestamps, latency, retries, and error states	Tracks execution efficiency and failure overhead
Delivery	Output file, state change, or database update produced	Verifies that the promised asset was transmitted
Acceptance	Formal sign-off record, inspecting reviewer, and decision timestamp	Confirms that the deliverable satisfied business requirements
Human effort	Review duration, manual corrections, and exception handling	Exposes hidden labor costs required to make the output usable
Economics	Gross revenue or utility minus model, tool, rail, and labor costs	Determines net contribution margin and economic viability
Repeat use	Subsequent customer orders or renewed workflow authority	Validates durable market demand and client satisfaction

This structure establishes four core performance ratios:

accepted jobs / eligible jobs
first-pass accepted jobs / delivered jobs
cost / accepted job
human minutes / accepted job

Retries, sub-agent delegations, and error loops remain internal events within a single job ID. Partial deliveries and manual rework represent operational friction that degrades gross margin. A system achieving a 95 per cent completion rate may remain commercially unviable if the remaining 5 per cent of failures cause severe data leaks, unauthorized payments, or regulatory breaches.

The minimum operational dataset

Tracking twenty consecutive production jobs—or four continuous weeks of operation—provides an initial empirical baseline for a bounded workflow. It reveals labor overhead, rework rates, and edge-case failures that synthetic benchmarks and single-day demonstrations conceal:

1. Eligible population: Total business tasks or customer requests eligible for delegation.
2. Active utilization: Unique users or production workflows that invoked the agent within the window.
3. Delegated scope: The specific decisions and actions the software was permitted to execute.
4. Verified completion: Deliverables formally inspected and accepted, not raw generated responses.
5. Operational reliability: Rates of unhandled exceptions, retries, duplicate actions, and human escalations.
6. Human review overhead: Labor minutes expended on validation, prompt corrections, and exception handling.
7. Unit economics: Fully loaded cost (compute, APIs, payment fees, human review) per accepted deliverable.
8. Retention and repeat rate: Recurring usage patterns across stable customer or workflow cohorts.
9. Risk and failure events: Security near-misses, policy violations, and data integrity incidents categorized by severity.
10. Pre-agent baseline: Identical metrics captured from the legacy human or automated process being replaced.

Evaluating these fields produces defensible operating records for individual business cases. It avoids reliance on generic industry averages that mask catastrophic tail risks. An agent workflow that successfully completes 19 out of 20 tasks at high speed is an unacceptable operational failure if the single failed job transfers funds to an unverified counterparty or exposes protected customer records.

The investment and decision record

Individual job logs should feed into a structured governance record before an enterprise scales deployment:

Decision field	Baseline assumption	Observed production record
Operational baseline	Historical job volume, processing delays, error frequency, and fully loaded labor cost	Like-for-like performance data across identical task classes
Capital investment	Upfront engineering, prompt design, security hardening, data pipeline prep, and change management	Actual capital deployed, including ongoing workflow maintenance
Unit operating cost	Anticipated token consumption, tool API fees, payment interchange, and human spot-checking	Realized cost per delivered and accepted deliverable
Commercial return	Expected cycle-time reduction, expanded capacity, labor savings, or net revenue growth	Realized business impact verified by the sponsoring business unit
Loss ceiling	Maximum acceptable financial loss, data breach exposure, or customer churn impact	Empirical record of near-misses, policy violations, and error containment
Governance rule	Explicit performance thresholds governing expansion, mandatory redesign, or immediate termination	Formal review record, decision rationales, and scheduled audit cadences

Twenty production jobs do not eliminate uncertainty. They make trade-offs, operational failure modes, and unit economics transparent. For high-consequence corporate operations, validation may require larger sample cohorts, adversarial red-teaming, or maintaining mandatory human-in-the-loop controls permanently.

Empirical evidence gaps

The market lacks reliable cross-industry datasets that separate conversational assistance from autonomous execution, accepted deliverables, and fully loaded operating costs. Similarly, there is negligible longitudinal data proving that enterprise AI scaling reliably correlates with sustainable EBIT expansion or productivity growth. Until those empirical gaps are filled by audited disclosures, market participants must distinguish between visible technical capability and unmeasured commercial adoption.

Four distinct market conditions are routinely collapsed into the single term adoption:

• Definitional scope: What an organization or survey classifies as an agent.
• Published supply: What a software vendor documents and sells in its catalog.
• Observed deployment: What a customer configures and runs in a sandbox or internal pilot.
• Economic settlement: What an autonomous workflow delivers, verifies, reconciles, and earns repeat demand for.

Evidence from an earlier stage in that sequence cannot prove the arrival of a later one.

The trajectory of agent measurement

Market measurement is shifting away from intermediate activity metrics toward completed business work. Token counts, tool invocations, directory listings, and transaction volumes remain valuable infrastructure signals, but institutional buyers will increasingly evaluate agent providers on verified acceptance rates, labor overhead, failure remediation, and gross margin contribution.

This shift will reveal significant commercial divergence across software sectors. Workflows operating in structured, low-consequence digital domains will demonstrate rapid adoption and compelling unit margins. Tasks involving ambiguous goals, external regulatory compliance, or material financial liability will remain bounded by human review and strict governance gates.

Chapter 3 maps the whole market: the four connected systems that support delegated software work, the five operational flows that move between them, and the commercial boundaries that separate suppliers.

Sources and method

• Primary sources: Australian enterprise evidence, independent evaluation research, provider telemetry disclosures, and empirical survey data checked as of 13 September 2026. Data points retain their original populations, sample frames, and study designs.
• Analytical frameworks: The three-dimension measurement model, the claim-record schema, the job-level performance ledger, and the investment decision card are original analytical tools developed to evaluate agentic commerce.
• Evidence boundaries: The Reserve Bank of Australia liaison reflects 105 medium and large corporate respondents from June to August 2025. McKinsey survey data reflects self-reported responses under proprietary survey definitions. Neither represents a complete census of Australian business operations.
• Editorial record: Published by Agentic Economy, Perth, Western Australia. Updates and corrections are logged via the editorial record.

Back to the report hub · Next: The agentic economy market map.