Read the sample before the findings
The full export contains 919 rows, including invited contacts. There are 314 eligible non-invited applications. After invited-only rows are excluded and responses are deduplicated by email for each question, individual item bases range from 71 to 101, which is 23 to 32 percent of the 314 eligible applications.
That item completion range is a central limitation, not a footnote. The 314 eligible applications are not used as a universal denominator. Every percentage in this report carries its own question base.
Definitions used in this report
| Term | Meaning in this report | What it does not prove |
|---|---|---|
| Production agent | An AI agent the respondent counted as running in production in the instrument | The survey did not independently verify uptime, autonomy, value or control quality |
| Embedded AI | AI whose removal would change the organization's cost structure or hiring plan | It is not simply another name for deployed software |
| Full real-time cost visibility | The respondent selected “full real-time visibility” for AI operating costs | It is self-reported and was not tested against a required cost taxonomy or audited ledger |
| Applicable access response | A respondent who selected an access method rather than “not applicable yet” | It does not represent all 314 eligible applications |
What changed since the Preview
The September Preview relied on small technology-seat cuts from other instruments, with bases of 21 to 32. Edition 1 uses a dedicated production instrument, and each published headline clears Open Future Forum's 40-response publication floor.
The instrument directly measures agent counts, production constraints, operating-cost visibility and agent access to enterprise data. It therefore replaces proxy conclusions with an operator evidence book while retaining the limits of a selective, self-reported sample.
Preview inference versus current test
The distinction matters because a proxy can identify a question without answering it. The September Preview asked whether technical leadership was becoming a production operating role. The current instrument tests that proposition directly through several separate questions.
| September Preview question | Current October test | Result now available | What remains untested |
|---|---|---|---|
| Is AI reaching production? | Four-stage maturity and production-agent count | 54 of 91 report deployed or embedded AI; 60 of 75 report at least one production agent | Independent verification of production status, reliability or business value |
| Is AI becoming a fleet-management problem? | Number of production agents | 22 of 75 report more than twenty agents | Whether respondents count agents consistently or whether every reported agent is material |
| Can leaders see operating economics? | Real-time AI cost visibility | 34 of 77 report full visibility; among 60 matched agent operators, 30 report full visibility | The cost taxonomy, accuracy and workflow-level reconciliation behind the answer |
| What is blocking production? | Multi-select bottleneck question | Data access and quality leads at 39 percent (29 of 75) | Severity, duration and whether a named bottleneck caused delay or failure |
| How do agents reach data? | Enterprise-data access method | 24 of 49 applicable responses use per-agent credentials and 18 shared service accounts | Permission scope, logging quality and whether any configuration meets a specific control standard |
October therefore replaces several plausible narratives with measured distributions. It does not convert self-reported answers into audits. The correct step is to use each result to request a more specific production record.
Who answered
| Respondent seat | Count | Share |
|---|---|---|
| CEO or founder | 36 | 36% |
| Data or AI leader | 23 | 23% |
| CTO, CIO or engineering | 16 | 16% |
| Other | 10 | 10% |
| Investor | 8 | 8% |
| CMO or marketing | 4 | 4% |
| CISO or security | 2 | 2% |
| Board director | 2 | 2% |
Base 101.
The sample is operator-led but not exclusively technical. CEO and founder respondents form the largest group, followed by data or AI leaders and technology executives.
Production context and agent fleets
Maturity is context here, not this report's primary question: 54 of 91 respondents report deployed or embedded AI. The AI Transformation Report owns the four-stage chart, complete table and operating-model interpretation. This source report begins its detailed evidence with the production-agent fleet.
| Production-agent count | Count | Share |
|---|---|---|
| None | 9 | 12% |
| 1 to 5 | 27 | 36% |
| 6 to 20 | 11 | 15% |
| More than 20 | 22 | 29% |
| Not sure | 6 | 8% |
Base 75.
Sixty of 75 respondents report at least one production agent, including 22 who report more than twenty. The modal answer is one to five, but the upper tail makes a fleet-level operating view necessary for many respondents.
For an operating model that covers lifecycle states, controlled releases, service tiers, fleet-level containment and retirement, read How to Operate More Than 20 Production AI Agents as a Fleet.
Operating-cost visibility
| Real-time cost visibility | Count | Share |
|---|---|---|
| Full real-time visibility | 34 | 44% |
| Partial visibility | 30 | 39% |
| No real-time visibility | 13 | 17% |
Base 77.
The unpaired headline compares 80 percent reporting at least one production agent (base 75) with 44 percent reporting full real-time cost visibility (base 77). The 36-point difference is descriptive because the question bases differ.
The matched analysis is more useful. Among the same 60 respondents who report at least one production agent and answered the cost question, 30 report full visibility, 23 partial visibility and seven none. In other words, half of the matched agent operators select full visibility and half do not.
| Agent fleet | Matched base | Full visibility | Partial visibility | No visibility |
|---|---|---|---|---|
| None | 9 | 22% | 44% | 33% |
| 1 to 5 | 27 | 37% | 44% | 19% |
| 6 to 20 | 11 | 91% | 9% | 0% |
| More than 20 | 22 | 45% | 45% | 9% |
| Not sure | 6 | 33% | 33% | 33% |
Matched respondents answering both questions; base 75. Small row bases are directional.
Ten of the eleven respondents reporting six to twenty agents select full visibility. The more-than-twenty group splits evenly between full and partial visibility, with two reporting none. These rows are too small to establish a general relationship between fleet size and cost control.
The maturity cut provides additional context: full visibility is reported by 20 of 33 deployed respondents, compared with four of 18 explorers. This cross-sectional association does not establish causation or movement between stages. The complete exhibit and maturity interpretation appear in the AI Transformation Report.
Where production gets stuck
| Production bottleneck | Count | Share |
|---|---|---|
| Data access and quality | 29 | 39% |
| Compute | 22 | 29% |
| Integration with existing systems | 21 | 28% |
| Inference cost | 19 | 25% |
| Governance and approval | 16 | 21% |
| Agent identity and permissions | 11 | 15% |
| Talent | 11 | 15% |
Base 75; any mention, so shares can sum above 100 percent.
Data access and quality leads, while compute, integration and inference cost form a second group. The distribution shows that the production problem is not reducible to model choice: data, infrastructure, economics, governance and identity all appear in the answers.
How agents reach enterprise data
| Access method, all respondents | Count | Share |
|---|---|---|
| Per-agent credentials | 24 | 34% |
| Shared service accounts | 18 | 25% |
| Delegated user identity | 4 | 6% |
| Credential-free or brokered access | 3 | 4% |
| Not applicable yet | 22 | 31% |
Base 71.
| Access method, applicable responses | Count | Share |
|---|---|---|
| Per-agent credentials | 24 | 49% |
| Shared service accounts | 18 | 37% |
| Delegated user identity | 4 | 8% |
| Credential-free or brokered access | 3 | 6% |
Base 49.
The applicable cut answers a narrower question: among respondents already using one of the listed access methods, how are agents authenticated? Shared service accounts account for 18 of 49 answers. The all-respondent table retains “not applicable yet” so the report does not imply that every respondent operates agents with enterprise-data access.
Buying authority and planned spend
| Signoff role | Count | Share |
|---|---|---|
| CEO | 48 | 56% |
| CIO or CTO | 24 | 28% |
| Business unit | 13 | 15% |
| No single owner | 11 | 13% |
| Finance | 10 | 12% |
Base 86; any mention, so shares can sum above 100 percent.
| Planned spend over twelve months | Count | Share |
|---|---|---|
| Under $100K | 31 | 41% |
| $100K to $1M | 17 | 22% |
| $1M to $10M | 9 | 12% |
| $10M to $100M | 2 | 3% |
| Over $100M | 5 | 7% |
| Not sure | 12 | 16% |
Base 76.
The CEO appears in 48 of 86 sign-off answers. Planned spend is widely distributed: 33 of 76 respondents plan at least $100,000, including 16 planning at least $1 million. The bands are too broad, and the sample too selective, to support a credible average or market-size extrapolation.
Production control register
This register is the practical output of the report. It is a management template, not a survey result. Complete one row for every material production agent and reconcile it with finance, security and the business owner.
| Register field | Minimum entry | Control question |
|---|---|---|
| Agent and version | Stable identifier, model and deployment version | Can a reviewer identify exactly what ran? |
| Business workflow | Workflow served and intended outcome | Is the agent tied to a real operating process? |
| Accountable owners | Business owner and technical owner | Who approves changes and accepts results? |
| Identity method | Per-agent, shared, delegated or brokered credential | Are actions attributable and access revocable? |
| Data scope | Systems and data classes the agent can reach | Is access limited to what the workflow requires? |
| Cost record | Cost center, unit of work and visibility status | Can operating cost be reconciled without a manual audit? |
| Service evidence | Quality, reliability and exception measure | Does performance meet the production standard? |
| Human control | Review, override and escalation path | Who intervenes when the agent is wrong or uncertain? |
| Last review | Date, reviewer and unresolved exception | Is the record current enough for a decision? |
The CEO, CFO and CISO production handoff
Production evidence becomes useful when the business, finance and security records describe the same agent and workflow. The handoff below is an operating recommendation derived from the findings, not a surveyed allocation of executive duties.
| Executive owner | Evidence received | Decision that owner makes | Evidence returned to the shared record |
|---|---|---|---|
| CEO or business executive | Workflow purpose, accountable operator, materiality and expected outcome | Whether the workflow belongs in production and which exceptions the organization will accept | Named business owner, approved outcome, priority and stop or escalation condition |
| CFO | Agent identifier, volume, model and infrastructure costs, human-review cost and budget source | Whether cost is visible enough to scale and whether the economic case meets the approved threshold | Cost center, unit-of-work measure, current run rate, value method and unresolved allocation gap |
| CISO | Agent identity, data scope, permissions, tools, logging and human override | Whether access is bounded, attributable and revocable; whether an exception requires treatment or acceptance | Approved access method, exception owner, remediation date and incident path |
| AI or technology leader | All three decisions plus service and change evidence | Whether the release remains within technical, economic and control boundaries | Current version, reliability result, change log and reconciled production-register entry |
The most important handoff rule is a shared identifier. If finance tracks a platform invoice, security tracks a service account and the business tracks a workflow name without a common key, the company may have three accurate records that cannot be reconciled. The register should bridge those records before a cost challenge or incident forces a manual reconstruction.
What AI leaders want next
Forty-three nonblank, non-invited open-text answers were reviewed qualitatively. Recurring topics include enterprise adoption, agents, infrastructure, deployment and governance. Respondents also ask how to scale agents reliably and connect them to enterprise systems. This is a thematic reading, not a coded distribution, so no percentage shares are assigned.
Tested against the record
Open Future Forum's results can be set beside broader research without treating unlike samples as one benchmark.
| External source | Useful context | Open Future Forum evidence | Limit on the comparison |
|---|---|---|---|
| McKinsey, State of AI 2026 | Wider evidence on agent scaling, operating cost and uneven financial impact | Fleet-size, bottleneck, spend and matched cost-visibility distributions from one operator instrument | Different samples and question wording; percentages should not be pooled |
| IBM, Cost of a Data Breach 2026 | Context on identity-based access, human attribution and agent auditability | 18 of 49 applicable responses report shared service accounts, alongside 24 reporting per-agent credentials | The survey records the selected method, not control quality or breach outcomes |
| NIST AI Risk Management Framework | A general structure for governing AI risk across the lifecycle | A production register that attaches ownership, access, cost and service evidence to the individual agent | NIST is a framework, not a validation of any respondent's implementation |
Open Future Forum adds a first-party operator view: data access is the leading named bottleneck, and only 30 of the 60 matched respondents operating at least one production agent report full real-time cost visibility. External sources provide context only; their publishers are not affiliated with this report and do not endorse it.
What this means for the AI leader
For management, the useful operating unit is the individual production agent attached to a workflow, identity, data scope, owner, cost record and service measure. A fleet becomes governable when those fields can be reconciled in one register and reviewed as operating evidence, not reconstructed after an incident or budget challenge.
Questions this report answers
How many respondents run production agents?
Sixty of 75 report at least one production agent, including 22 who report more than twenty.
Can agent operators see what AI costs to run?
Among the 60 matched respondents with at least one production agent, 30 report full real-time cost visibility, 23 partial visibility and seven none.
What is the leading production bottleneck?
Data access and quality is named by 39 percent, ahead of compute at 29 percent, integration at 28 percent and inference cost at 25 percent (base 75, any mention).
How do agents access enterprise data?
Among 49 applicable responses, 24 use per-agent credentials, 18 shared service accounts, four delegated user identity and three credential-free or brokered access.
How much do respondents plan to spend?
Thirty-three of 76 plan at least $100,000 over the next twelve months, including 16 planning at least $1 million. The bands should not be converted into a market-size estimate.
Does reporting a production agent prove that it is autonomous or reliable?
No. “Production agent” is the respondent's answer to the instrument. The survey did not independently test autonomy, uptime, quality, materiality or the control environment around the system.
What is the minimum record for a cross-functional production review?
Use one stable agent identifier linked to the workflow, business and technical owners, access method, data scope, cost center, unit of work, service evidence, human-control path and last review. The record should reconcile CEO or business, CFO and CISO decisions.
Key citable facts
- Open Future Forum's October 2026 AI Leaders instrument finds 60 of 75 respondents reporting at least one production agent.
- Among the 60 matched respondents running at least one production agent, 30 report full real-time cost visibility, 23 partial visibility and seven none.
- Data access and quality is the leading production bottleneck at 39 percent (base 75, any mention).
- Shared service accounts represent 37 percent of applicable agent-access answers (18 of 49).
- Thirty-three of 76 respondents plan at least $100,000 in AI spend over the next twelve months.
Methodology and honesty notes
This edition uses first-party questions embedded in the Enterprise AI at Microsoft application flow. Its 919-row export contains 314 non-invited applications. Invited-only records are excluded. Respondents are deduplicated by email for each question, with the latest answer retained. Individual item bases range from 71 to 101, or 23 to 32 percent of eligible non-invited applications, because questions were added or completed at different points in the flow. The 314 eligible applications are not used as a universal base.
Multi-select questions use the any-mention convention. No headline is published below 40 responses. The access-method applicable cut excludes respondents who answered “not applicable yet.” Matched analyses include only respondents who answered both relevant questions; their row bases are shown, and small rows are directional.
This is a selective sample drawn from Open Future Forum's network, not a probability sample of all enterprises. Production status, fleet size and cost visibility are self-reported and were not independently audited. Planned-spend bands are reported as bands; no midpoint estimate or market-size extrapolation is made. The open-text section is a qualitative review of 43 answers, not a coded distribution. No causal or longitudinal claim is made from cross-sectional answers.
Related reading
- AI Transformation Report: deployed versus embedded AI and the transition gates
- CISO AI Leverage Report: identity, access and security funding
- Executive AI Statistics: full source tables and sample accounting
- How Many AI Agents Are in Production in 2026?
- Do Companies Have Real-Time Visibility into AI Operating Costs?
- What Prevents AI from Reaching Production in 2026?
- How Do AI Agents Access Enterprise Data in 2026?
Citation and editions
Suggested citation: Newlands, M. (2026). AI Leaders AI Leverage Report, Edition 1. Open Future Forum, October 2026. This edition supersedes the September 2026 Preview.
Work These Questions with AI Leaders Peers
The AI Leaders Forum meets through small, off-the-record gatherings. Membership is by application and referral.