What changed since Edition 2
Edition 2 relied on proxy signals because a direct maturity question was not available. Edition 3 uses the first four-stage maturity measure from the Enterprise AI at Microsoft application flow. It establishes an October baseline; it does not establish a trend.
The new instrument changes what this report can say and what it still cannot say. It can distinguish exploration, pilots, production deployment and the stricter embedded condition using one direct question. It cannot show how any individual organization moved between those states, how long a transition took or whether one management practice caused a more mature answer.
What has not changed is the analytical standard. Production is not treated as proof of value, and a cross-functional observation is not treated as if every respondent answered every question. The sections below use four markers of transformation, but each marker retains its own instrument, respondent population and denominator.
A four-marker model of AI transformation
Transformation is broader than technical release. For an executive review, four separate markers make the claim testable: work, economics, workforce or value, and governance or control. The model below organizes existing October evidence. It does not combine the measures into an index or score.
| Transformation marker | October evidence | Base and source boundary | What management still has to prove |
|---|---|---|---|
| Work | 42 report deployed AI and 12 report embedded AI; together, 54 of 91 report deployed or embedded | Base 91, AI Leaders maturity question | Which workflow changed, who owns it and whether the new process is reliable enough to remain in use |
| Economics | 34 report full real-time cost visibility, 30 partial visibility and 13 none | Base 77, AI Leaders cost question | Whether workflow-level cost can be reconciled to an outcome, budget owner and unit of work |
| Workforce and value | 101 name doing the work of more people, 99 creating content faster and 94 knowing the customer better | Base 218, marketing impact question; any mention | Whether the selected benefit has a baseline, attribution method and decision consequence |
| Governance and control | Among 49 applicable access responses, 24 use per-agent credentials and 18 shared service accounts | Base 49 applicable responses from the AI Leaders access question; the all-response base is 71 | Whether agent actions are attributable, permissions are bounded and access can be revoked without disrupting the workflow |
These markers should not be averaged. The workforce and value answers come from a marketing-oriented instrument, while the maturity, cost and access answers come from the AI Leaders instrument. They illuminate different parts of an operating model; they do not describe one matched cohort.
Marker one: work changes
The maturity baseline begins with the operating workflow. Exploring and piloting can produce learning, prototypes and local productivity, but the production and embedded categories ask whether the system is part of real work. Even then, the two upper stages are not interchangeable. A production system can remain an additional layer around the old process. An embedded system changes the assumptions used to allocate capacity, cost or hiring.
The 12 embedded answers therefore carry a more demanding meaning than “widely used.” A defensible embedded claim should be accompanied by a before-and-after workflow, a named owner and an explicit statement of which operating decision now depends on AI. Without those records, management may be describing adoption rather than transformation.
Marker two: economics become visible
Transformation changes the cost of delivering work, not merely the software budget. Full real-time visibility is reported by 34 of 77 respondents, while 43 report partial or no visibility. That result does not show that those 43 have no cost management. It shows that they did not select the strongest visibility answer offered by the instrument.
The useful economic question is not simply “What did we spend on AI?” It is “What does this workflow cost at a defined unit of work, and what value or capacity decision is management making because of it?” Model fees, inference, data movement, infrastructure, observability, human review and exception handling may sit in different cost centers. Transformation evidence has to reconnect them.
Marker three: workforce or value changes
The marketing impact instrument offers one view of how value is being described. Workforce leverage, content speed and customer knowledge all appear frequently, and no single answer dominates. Because the question allows multiple selections, the findings should not be turned into exclusive value archetypes.
For transformation purposes, each value claim needs a different record. “Doing the work of more people” should connect to capacity, role design or hiring assumptions. “Creating content faster” needs both elapsed-time and quality measures. “Knowing the customer better” needs an observable decision or customer result. A selected impact is a signal about where to investigate, not an audited benefit.
Marker four: governance follows the system into production
Identity is a practical test of whether controls have moved with the workflow. Among respondents using one of the listed access methods, 24 report per-agent credentials and 18 report shared service accounts. The latter can make attribution, least-privilege design and revocation harder, although the survey did not inspect how any account was configured.
Governance becomes operational when a reviewer can identify the agent, its version, the human and business owners, the systems it can reach, the authority it can exercise and the path for stopping or correcting it. A policy that cannot be reconciled to this system-level record is not yet a production control.
The four-stage maturity baseline
| Stage | Count | Share |
|---|---|---|
| Exploring | 19 | 21% |
| Piloting | 18 | 20% |
| Deployed in production | 42 | 46% |
| Embedded | 12 | 13% |
Base 91.
Fifty-four respondents report AI deployed or embedded. The distinction between those categories is material: deployment means a system is live, while the embedded answer is reserved for organizations where removing AI would change the cost structure or hiring plan.
The 33-point difference between deployed and embedded is therefore a difference in reported maturity states at one point in time. It should not be described as a failure rate, a transition rate or the share of deployments that will never become embedded.
Why deployment is not the same as transformation
A live model can automate a task without changing how the company allocates work, budgets capacity or plans headcount. Under this instrument, transformation becomes visible only when AI is tied to operating decisions that would have to be reversed if the system disappeared.
That makes the embedded threshold intentionally demanding. It asks for structural dependence, not enthusiasm, experimentation or the existence of a production endpoint. The October result shows a large production population and a much smaller group reporting that level of operating-model change.
Maturity and cost visibility: a matched view
The maturity and cost questions can be joined for 75 respondents who answered both. This matched table is more informative than comparing two unpaired headline percentages, but its smaller rows remain directional.
| Maturity stage | Matched base | Full visibility | Partial visibility | No visibility |
|---|---|---|---|---|
| Exploring | 18 | 22% | 33% | 44% |
| Piloting | 14 | 36% | 43% | 21% |
| Deployed in production | 33 | 61% | 33% | 6% |
| Embedded | 10 | 50% | 50% | 0% |
Matched respondents answering both questions; base 75. Small row bases are directional.
Twenty of 33 deployed respondents report full visibility, compared with four of 18 exploring respondents. None of the ten embedded respondents selects no visibility, although half select only partial visibility. The embedded row is too small to establish a stable estimate, and the entire table is cross-sectional. It does not show that cost visibility caused a higher maturity stage, that maturity created visibility or that any respondent moved from one row to another.
The operating interpretation is narrower: maturity and cost evidence belong in the same review. If a production or embedded claim cannot be connected to a cost record, management should label the economic evidence incomplete rather than downgrade or upgrade the technical stage by assumption.
Four-stage transition-gates framework
The following is a management framework derived from the survey's four stages. The gates are recommendations, not additional survey findings.
| Current stage | Evidence available now | Gate before the next stage | Decision record to retain |
|---|---|---|---|
| Exploring | A defined problem and a plausible AI use case | Name the business owner, baseline the current workflow and document the risk screen | Use-case brief, baseline measure and stop/go owner |
| Piloting | A bounded test with users and an observable result | Set an exit criterion for quality, reliability, economics and control before production approval | Pilot scorecard, exceptions and production decision |
| Deployed in production | A live system serving a real workflow | Connect the system to a named workflow owner, service expectations, cost record and workforce plan | Production register, operating review and change log |
| Embedded | Removing AI would change the cost structure or hiring plan | Re-test the dependence periodically and retain a viable contingency | Budget or capacity decision, workforce impact and continuity plan |
The framework prevents a common category error: treating technical release as proof of enterprise transformation. An organization should not move its own status to embedded merely because usage is high; it should be able to point to the operating decisions that now depend on AI.
A practical transformation review
Use these prompts in a monthly operating review. They are designed to produce evidence rather than a maturity label alone.
| Review question | Evidence to request | What the answer establishes |
|---|---|---|
| Which workflow changed? | Before-and-after process map and accountable owner | Whether AI altered work rather than adding another tool |
| Which operating measure moved? | Agreed baseline and current result | Whether the change is observable |
| Which budget or capacity decision changed? | Approved budget, capacity or hiring record | Whether the system affects the operating model |
| What would happen if AI were removed for thirty days? | Documented continuity test or management assessment | Whether dependence is operationally understood |
| Who can reverse or redesign the change? | Named executive and technical owners | Whether the transformation remains governable |
To convert the 30-day removal question into an executable continuity exercise with thresholds, rollback and remediation, use How to Run a 30-Day AI Continuity Drill.
Failure modes that make transformation look deeper than it is
Counting tools instead of changed work
License counts, prototypes and agent inventories show activity. They do not show that a workflow changed or that the old process was retired. Ask for the specific handoff, decision or task that now operates differently, then confirm that the responsible business owner recognizes the change.
Declaring value without preserving a baseline
A faster process can still cost more, produce more exceptions or shift work to reviewers. The baseline should include time, quality, operating cost and exception handling in the same measurement frame. If management cannot reconstruct the pre-AI workflow, a later value claim will be difficult to audit.
Treating partial cost visibility as a complete unit-economics record
Partial visibility may be sufficient for a pilot but inadequate for a material production workflow. The review should identify which costs are visible, which remain allocated or estimated, and whether the missing amount could change the decision to scale.
Moving responsibility to a central AI team
The central team can provide platforms, standards and expertise. It cannot replace the business owner who accepts the workflow result, the CFO who reconciles economics, the CISO who governs identity and access, or the people leader who manages role change. A transformation program without those distributed responsibilities is vulnerable to orphaned systems.
Using an embedded label without a contingency
If removing AI would change cost or hiring, the organization has created operational dependence. That dependence should have a continuity plan: what happens during provider failure, unacceptable quality, a security incident or a regulatory restriction? Embedded status raises the need for resilience; it does not remove it.
Interpreting the maturity result
Three readings are supported by the data:
- Production is common in this selective sample: 42 of 91 respondents choose deployed in production.
- Structural dependence is much less common: 12 of 91 choose embedded.
- The data does not show how quickly organizations move between stages, because the direct maturity measure is new and cross-sectional.
The report does not combine finance, security or production-agent measures into a synthetic transformation score. Those questions have different bases and, in some cases, different respondent populations. Readers seeking the underlying production-control evidence should use the AI Leaders report; readers seeking the complete count and base book should use Executive AI Statistics.
Tested against the record
McKinsey's 2026 State of AI reports broader enterprise scaling and identifies a small high-performer group. Open Future Forum uses a different and narrower maturity measure: AI counts as embedded only when removing it would change costs or hiring. Under that definition, 13 percent of respondents are embedded (base 91).
| External source | Useful context | What Open Future Forum adds | Comparison limit |
|---|---|---|---|
| McKinsey, State of AI 2026 | Broader evidence on enterprise scaling, operating practices and high-performing organizations | A direct four-stage maturity item plus matched maturity-and-cost visibility evidence | The studies use different samples and wording; their percentages should not be pooled or treated as a trend |
| NIST AI Risk Management Framework | A general structure for governing and managing AI risk | A practical transition-gate record connecting workflow, cost, identity, ownership and contingency evidence | NIST is a framework, not a benchmark or validation of this survey |
| IBM, Cost of a Data Breach 2026 | Context on identity, access and auditability for agentic systems | A reported distribution of enterprise-data access methods among applicable AI Leaders respondents | The IBM study and this report answer different questions and do not establish causality |
These sources are used as context only. Their publishers are not affiliated with this report and do not endorse it.
What this means for the executive team
Treat “in production” as a release milestone, not the final transformation claim. The executive team should be able to name the workflow, owner, operating measure and budget or workforce decision changed by AI. If those links are absent, the organization may have a functioning deployment without an embedded operating model.
For boards
Ask management to reconcile the material-system inventory to workflow ownership, cost visibility, access method and the decisions that depend on the system. The board does not need every prototype. It does need pre-agreed escalation rules for material failures, cost exceptions, unauthorized access, provider concentration and workforce consequences.
For talent and people leaders
Translate an embedded claim into role-level evidence. Identify which tasks changed, which skills are newly required, which review work increased and which hiring assumption moved. “Doing more with AI” is not a workforce plan until it connects to capacity, accountability and employee transition.
For investors and transaction teams
Separate product availability from operating dependence. A production count can support a scale narrative, but an embedded claim should be tested against cost records, hiring or capacity decisions, workflow outcomes and contingency planning. The absence of one record does not invalidate the system; it identifies the next diligence question.
Questions this report answers
How much enterprise AI is deployed or embedded?
Fifty-nine percent report AI deployed in production or embedded in the operating model (54 of 91).
What share meets the stricter embedded test?
Thirteen percent report that removing AI would change the organization's cost structure or hiring plan (12 of 91).
Is the 33-point difference a transition rate?
No. It is the cross-sectional difference between the deployed share of 46 percent and embedded share of 13 percent. The instrument does not follow organizations through the stages over time.
What evidence should support an embedded claim?
Management should be able to identify the workflow change and the associated budget, capacity or hiring decision, then explain what would have to change if the AI system were removed.
Can the four transformation markers be combined into one score?
No. They come from separate questions and, in the workforce or value case, a separate respondent population. They are a review structure, not a matched index.
Does the matched cost table show that visibility causes maturity?
No. It shows how 75 respondents who answered both questions are distributed at one point in time. The pattern is descriptive and does not establish direction, transition or causality.
Why does embedded status require a continuity question?
The embedded definition says removal would change cost structure or hiring. That level of dependence makes provider failure, control failure or model withdrawal an operating-model issue, so management should retain a viable contingency.
Key citable facts
- Open Future Forum's October 2026 maturity instrument finds 46 percent deployed in production and 13 percent embedded (base 91).
- Deployed and embedded respondents together account for 59 percent of the maturity sample (54 of 91).
- The deployed category exceeds the embedded category by 33 percentage points; the difference is cross-sectional and is not a transition rate.
Methodology and honesty notes
The maturity question comes from the Enterprise AI at Microsoft application flow. Invited-only records are excluded, and responses are deduplicated by email for the question, with the latest answer retained. The maturity base is 91.
The maturity-by-cost table contains 75 respondents who answered both questions. Its stage rows have bases from 10 to 33 and are descriptive; the smaller rows are directional. The access-method marker uses 49 respondents who selected one of the listed methods and excludes 22 who answered “not applicable yet.” The workforce and value marker comes from a separate marketing instrument with a base of 218 and uses the any-mention convention. Those measures are kept separate and are not combined into a transformation score.
The direct four-stage measure is new and has no prior-period like-for-like comparison. The 13 percent figure is a baseline, not a trend. The sample is selective and drawn from Open Future Forum's network; it is not a probability sample of all enterprises. No causal claim is made from these cross-sectional answers.
Related reading
- AI Leaders AI Leverage Report: production-agent fleets, costs, bottlenecks, data access and planned spend
- Executive AI Statistics: complete source tables, bases and methodology
- Executive AI Leverage Report: cross-functional executive synthesis
- What Does “Embedded AI” Mean in 2026?
- What Prevents AI from Reaching Production in 2026?
Citation and editions
Suggested citation: Newlands, M. (2026). AI Transformation Report, Edition 3. Open Future Forum, October 2026. This edition supersedes Edition 2, September 2026.
Work These Questions with Executive Peers
Forum Select meets through small, off-the-record gatherings. Membership is by application and referral.