VentureBeat Intelligence surveyed 106 people in October 2026, and 100 qualified: people who build, run or use AI at organizations with 100 or more employees, and consultants serving such organizations. We asked which agent orchestration platforms their organizations run, what happened when they changed the AI model in a production system, and whether an AI agent or automated workflow had failed in a way that affected a customer, an employee or a business process.
Executive summary
Enterprises can change the AI model inside a production app to cut costs or get better results. The app around the model, meaning its prompts, tools and tests, often has to be reworked to match.
Of respondents, 47% say their organization replaced the model in a production AI system in the past 12 months while keeping the surrounding application. Of those, 69% spent a few weeks or more on the swap, and 36% spent more than a month, up from 12% in September. Only 13% said the swap was close to a configuration change. And 44% either saw performance drop and reworked the application or went back to the original model; in September that figure was 58%, and the change is not statistically significant.
Of respondents, 36% say an AI agent or automated workflow failed in the past 12 months in a way that affected a customer, an employee or a business process. Another 22% do not know whether one did. In 56% of the most significant failures, the failure reached customers or a financial, regulatory or board report.
Of respondents, 65% run AI agents on two or more orchestration platforms. OpenAI's platform is used by 61%, Anthropic's by 47% and Microsoft's by 40%. We report platform use alongside September's figures as two separate survey readings, and the section on changes since September explains why.
Of respondents, 61% say their organization adopted a new, additional or replacement orchestration platform in the past 12 months, and 45% plan to adopt one in the next 12 months, down from 64% in September. Of those planning one who named a candidate, 54% are considering Anthropic's platform and 28% OpenAI's.
Of respondents who expect to deploy AI agents at scale, 77% expect to keep some control outside the model providers' own platforms by the end of 2027. Of those expecting a hybrid of provider controls and their own, 21% have documented which workloads each side governs.
Of respondents with agents deployed, 16% say more than half of their agents are multi-step workflows, down from 35% in September.
Of respondents who answered, 42% cannot point to anything that automatically stops their agents' token spending: 28% do not know how spending is controlled, 9% have no specific controls and 5% monitor only after the fact. Another 33% use limits built into their orchestration platform.
Finding 1. Swapping the AI model in a production app took weeks of engineering for 69% of respondents who did it
Of respondents whose organizations swapped the model, 36% needed more than a month, up from 12% in September. Only 13% said the swap was close to a configuration change.
An enterprise AI system wraps a model in an application: an agent loop, tools, prompts and evaluations tuned to how that model behaves. Swapping the model lets an enterprise move to a cheaper or stronger one. When the new model behaves differently, the prompts, tools and tests built around the old model may need rework.
We asked whether the organization had replaced the underlying model in a production AI system in the past 12 months while keeping the surrounding application in place.
Finding 1 — Replaced the model in a production AI system, past 12 months
Base: 100 qualified respondents, since the question applies to every organization. One answer each.
Of respondents, 47% say their organization swapped a model. Among those respondents, 56% said the new model performed as well or better. Another 33% saw performance drop and reworked the application, and 11% went back to the original model.
We then asked respondents whose organizations swapped a model how much engineering effort the swap took, beyond changing the model itself.
Finding 1 — Engineering effort beyond changing the model
Base: 45 respondents whose organizations swapped a model, the only group asked. One answer each.
Among respondents whose organizations swapped a model, 36% spent more than a month on the swap, up from 12% in September. Only 13% said the swap took almost no effort.
Finding 2. 36% of respondents say an AI agent failed and affected their business in the past year
Another 22% do not know whether one did. In 56% of the most significant failures, the failure reached customers or a financial, regulatory or board report.
An AI agent that acts on its own can give a customer the wrong answer, put an incorrect figure in a report or stall a business process. Finding such a failure depends on someone, or some system, checking the agent's output.
We asked whether, in the past 12 months, an AI agent or automated AI workflow in the organization had failed in a way that affected a customer, an employee or a business process.
Finding 2 — AI agent or workflow failure, past 12 months
Base: 100 qualified respondents, since the question applies to every organization. One answer each.
We asked respondents who reported a failure what happened as a result of the most significant instance.
Finding 2 — What happened in the most significant instance
Base: 34 respondents who reported a failure, the only group asked. One answer each.
In 56% of the most significant failures, the failure reached customers or a financial, regulatory or board report. In 38%, the failure was caught internally.
Finding 3. 65% of respondents run AI agents on two or more orchestration platforms, so no single vendor's controls cover all their agents
OpenAI's platform is used by 61% of respondents, Anthropic's by 47% and Microsoft's by 40%.
An agent orchestration platform coordinates a company's AI agents: The platform hands out tasks, calls tools and keeps agents within their permissions. OpenAI, Anthropic, Google and Microsoft each sell one alongside their own models.
We asked which agent orchestration platforms the organization uses today, then which one is primary.
Finding 3 — Platform
Base: 100 qualified respondents, all of whom use a listed orchestration platform. Several answers allowed for use, so that column sums past 100%; one answer for primary. Platforms used by under 5% are not shown.
More respondents selected OpenAI's platform than Anthropic's, both for use (61% against 47%) and as the primary platform (39% against 25%), but neither gap is statistically significant.
Finding 3 — Orchestration platforms in use
Base: 100 qualified respondents.
Respondents rated their primary orchestration platform 3.89 out of 5 for overall satisfaction, 3.72 for ease of implementation and 3.71 for value for money.
Base: 100 qualified respondents for each rating.
Finding 4. Anthropic leads the shortlist for respondents' next orchestration platform: 54% of those planning a new one are considering it, against 28% for OpenAI
Of the 39 respondents planning a new orchestration platform within 12 months who named a candidate, 54% are considering Anthropic's and 28% OpenAI's. Overall, 61% of respondents say their organization adopted a platform in the past year, and 45% plan to adopt one in the next 12 months, down from 64% in September.
Adopting an orchestration platform means moving or rebuilding agent workflows, connecting the platform to company data and tools, and setting up its permissions and monitoring.
We asked whether the organization adopted a new, additional or replacement agent orchestration platform in the past 12 months.
Finding 4 — Adopted a new, additional or replacement platform, past 12 months
Base: 100 qualified respondents. One answer each.
We also asked whether the organization plans to adopt a new, additional or replacement platform in the next 12 months. Of respondents, 45% plan to adopt one: 12% within three months, 20% within three to six months and 14% within six to 12 months. Another 55% reported no plans to adopt one in that period.
Base: 100 qualified respondents. One answer each.
We asked respondents planning a new platform which platforms they are actively considering.
Finding 4 — Platform under consideration
Base: 39 respondents planning a platform within 12 months who named one; three more are evaluating without a shortlist, and one named none. Several answers allowed. Platforms named by fewer than 8 respondents are not shown.
Among respondents planning a new platform who named any candidate, 54% are considering Anthropic's platform and 28% OpenAI's, a statistically significant gap. Of the respondents considering Anthropic's platform, 48% do not use it today. Custom in-house orchestration is named by 41%.
Finding 5. 77% of respondents planning AI agents at scale expect to keep some control outside OpenAI's, Anthropic's and other model providers' platforms by 2027
Within that 77%, 45% expect a hybrid of provider controls and their own, 24% a custom in-house control plane and 8% an independent orchestration platform. Of the 34 who expect a hybrid, 21% have documented which workloads each side governs.
The agent control plane decides which agent runs, which tools and data each agent can reach, and what gets logged. A hybrid control plane combines a model provider's controls with the enterprise's own layer, so the enterprise has to decide which side governs what.
We asked where the organization expects the primary control plane for enterprise agents to live by the end of 2027.
Finding 5 — Expected primary control plane, end of 2027
Base: 75 respondents who expect to deploy autonomous agents at scale, the group the question applies to. Another 20 qualified respondents do not expect to and are left out. One answer each.
Taken together, 77% of these respondents chose a hybrid, custom in-house or external control plane. The remaining 23% expect a provider-managed agent service to be their primary control plane.
We asked respondents expecting a hybrid control plane how work is divided between the model provider's built-in controls and their own control layer. The question is in the present tense, and one answer refers to how the boundary is drawn today, so we read the answers as describing current arrangements at organizations that expect a hybrid by 2027.
Finding 5 — How a hybrid control plane divides work
Base: 34 respondents expecting a hybrid control plane, the only group asked. One answer each.
Finding 5 — Top risk if orchestration control lives inside a model provider's platform
Base: 100 qualified respondents. One answer each.
Asked about the top risk if orchestration control lives inside a model provider's platform, 42% of respondents chose security and permissioning limitations, more than chose any other risk.
Finding 6. Only 16% of respondents say most of their deployed AI agents run multi-step workflows, down from 35% in September
For 69% of respondents with agents deployed, multi-step workflows make up half or fewer of those agents. The rest either report more than half (16%) or do not know (16%).
A multi-step orchestrated workflow has an agent carry out a sequence of steps, calling tools and other agents along the way. A single-prompt assistant answers one request at a time. Multi-step workflows can take on longer business processes, but they add more places for an error to spread.
We asked what share of the organization's currently deployed AI agents are multi-step orchestrated workflows, rather than single-prompt assistants or chat interfaces.
Finding 6 — Share of deployed agents that are multi-step orchestrated workflows
Base: 90 respondents with agents deployed, the group the question applies to. Five with no agents deployed are left out. One answer each.
Of respondents with agents deployed, 16% say more than half of their agents are multi-step workflows, down from 35% in September. Another 16% do not know the share.
Finding 7. 42% of respondents cannot point to anything that automatically stops their AI agents' token spending
Of respondents who answered, 28% do not know how token spending is controlled, 9% have no specific controls and 5% monitor only after the fact. Another 33% use limits built into their orchestration platform.
An agent that loops, retries or calls other agents can run up model charges quickly, because every step uses tokens. A real-time control stops an agent when it reaches a limit. Monitoring after the fact shows what an agent spent, but only once the charges have been incurred.
We asked how the organization controls token spending across the agent orchestration layer.
Finding 7 — How token spending is controlled
Base: 81 qualified respondents who answered; 14 left the question blank. One answer each.
In October, 33% of respondents named controls built into their orchestration platform, up from 20% in September, a Directional change. The share answering "don't know" rose from 10% to 28%, a statistically significant change.
What changed since the September Pulse
September's figures cover only respondents who would pass October's screen, so they differ slightly from the published September report. We asked the questions below in the same words in September and October.
We checked whether changes in the respondent mix could account for the results. In October, 88% of qualified respondents came through a research panel, and we did not record where September's respondents came from. We cannot tell how much of any change between the two surveys reflects who answered, and how much reflects changes in what organizations do.
Findings: statistically significant changes
Findings: statistically significant changes
A Finding meets the 95% significance threshold overall and among in-house respondents alone. Bases, September then October: respondents whose organizations swapped a model, 65 and 45; all qualified respondents, 100 and 95; respondents expecting agents at scale, 89 and 75; respondents with agents deployed, 96 and 90; token spending, 99 and 81.
Directional changes
Directional changes
Directional means the difference meets the 90% threshold, or meets the 95% threshold overall but not among in-house respondents alone. Bases as in the table above; the risk question uses all qualified respondents, 100 and 95.
Platform use, two readings
Platform use, two readings
Platform use is reported as two survey readings until a second month or a quarterly comparison confirms a change. Bases: 99 and 95 respondents using a listed platform. Margins of error assume a random sample; this sample is self-selected, so they understate the uncertainty.
Not statistically significant
Not statistically significant
Not statistically significant means the change does not meet the threshold described in the methodology. Bases as in the tables above; ratings 83 and 95, value for money 82 and 95.
What we did not compare
The question on replacing the model offered a "Don't know" answer in October, which 14% chose, and no September respondent gave that answer, so we do not compare the overall share that swapped.
The question on platforms being considered was left unanswered in October by respondents with no plans to change, although the questionnaire showed it to everyone, so its base differs from September's.
The questions on adopting a platform in the past 12 months and on agent failures are new in October.
The bottom line: Swapping the AI model in a production app took weeks of engineering for 69% of respondents who did it
Of respondents whose organizations swapped the model in a production AI app, 69% needed a few weeks or more of engineering work, and 36% needed more than a month, up from 12% in September.
Of respondents, 36% report an AI agent failure that affected a customer, an employee or a business process in the past year, and 22% do not know whether one happened.
In the next Pulse, VentureBeat Intelligence will measure whether the share of respondents whose organizations spent more than a month on a model swap holds at October's level, and whether platform use holds at October's readings.
Respondent profile
| Respondent type | Respondents |
|---|---|
| Works on AI for the organization that employs them | 72% (68) |
| Consultant who works on AI for client organizations | 28% (27) |
Base: 100 qualified respondents. One answer each.
| Organization size (main client, for consultants) | Respondents |
|---|---|
| 100–250 employees | 23% (22) |
| 251–500 employees | 17% (16) |
| 501–1,000 employees | 11% (10) |
| 1,001–5,000 employees | 14% (13) |
| 5,001–10,000 employees | 5% (5) |
| More than 10,000 employees | 31% (29) |
Base: 100 qualified respondents. One answer each.
| Role in AI purchasing | Respondents |
|---|---|
| Final decision maker | 14% (13) |
| Recommender or influencer | 56% (53) |
| User | 28% (27) |
| No involvement | 2% (2) |
Base: 100 qualified respondents. One answer each. Consultants answered for their main client's AI purchases.
| Job level | Respondents |
|---|---|
| Individual contributor | 39% (37) |
| Manager | 22% (21) |
| Director | 14% (13) |
| VP | 8% (8) |
| Partner or principal | 7% (7) |
| C-suite | 5% (5) |
| Other (founder, office manager, business owner, customer service) | 4% (4) |
Base: 100 qualified respondents. One answer each.
| Industry (main client, for consultants) | Respondents |
|---|---|
| Technology and software | 22% (21) |
| Financial services, banking and insurance | 21% (20) |
| Education | 18% (17) |
| Professional services and consulting | 11% (10) |
| Manufacturing | 8% (8) |
| Retail and e-commerce | 7% (7) |
| Healthcare and life sciences | 5% (5) |
| Transportation and logistics | 3% (3) |
| Energy and utilities | 2% (2) |
| Other (nonprofit) | 2% (2) |
Base: 100 qualified respondents. One answer each.
Methodology
Who answered. VentureBeat Intelligence fielded this VB Pulse survey in October 2026. The survey received 106 responses, of which 100 qualified. Respondents qualified if they work on AI at an organization with 100 or more employees, or consult for one as their main client with day-to-day or partial access to its AI systems.
Who did not qualify. We excluded staff at technology vendors serving their own customers, and one consultant without access to the main client's AI systems. No respondent was excluded for organization size, for a research or journalism role, or for not working directly on AI systems.
Organizations. The 100 qualified respondents give 73 distinct organization names; for consultants, this is the firm they work for. Eighteen organizations are represented by more than one respondent, and every response counts in the figures. Keeping one respondent per organization gives 36% for a model swap that took more than a month and 32% for agent failures, against 36% and 36% with every response counted.
Consultants. Consultants answered about the client organization where they spent the most time on AI work. Consultants are 28% of qualified respondents, and every figure in this report includes them. Of the consultants, 56% have day-to-day access to that client's AI systems.
Where respondents came from. Of qualified respondents, 88% came through a research panel and 11% through a VentureBeat email; one response has no recorded source. We did not record where September's respondents came from.
Who each figure covers. Each percentage is based on the respondents a question applies to, named in the note under each table.
Several answers. The questions on platforms in use and platforms under consideration allowed several answers. Their figures count people, not selections.
Contradictory answers. We checked every qualified respondent for a primary platform they did not say they use, and for platforms under consideration alongside the answer "not considering a change"; neither check found a contradiction. We also checked for a plan to adopt a platform alongside "not considering a change." One respondent gave that pair of answers and is flagged rather than removed. Leaving that respondent out moves no figure in this report by more than one percentage point.
Data checks. Fourteen respondents left the token-spending question blank, although the questionnaire showed it to everyone.
Small groups. Some groups in this report have fewer than 40 respondents: respondents who reported an agent failure, 34; respondents expecting a hybrid control plane, 34; respondents planning a new platform who named any candidate, 39. Respondents considering a given platform number 8 to 21. Percentages for groups this small are less precise; for a group of 34, a result could differ by about 17 percentage points either way from the true figure. We test every comparison between groups with the tests listed below, and we say when a gap is not statistically significant.
Tests. We describe a difference as statistically significant only when its test gives p below 0.05. Two percentages from separate groups use the two-proportion z-test, or Fisher's exact test when either group has fewer than 40 respondents. Two answers from the same respondents use the exact McNemar test, or an exact binomial test when only one answer was allowed. A claim that most of a group gave an answer uses an exact binomial test. Ratings use the Mann-Whitney test. The tests treat every response as independent, although 18 organizations are represented by more than one respondent, so the tests may somewhat overstate precision.
Changes from September. Each change is labelled a Finding, Directional or not statistically significant. A Finding must meet the 95% significance threshold overall and among in-house respondents alone. Platform use is reported as two readings, each with a conventional margin of error.
Survey changes. The size question split the 251–1,000 band in October. From October, respondents whose only platform answer was "not currently using agent orchestration" skip the primary-platform and rating questions; no October respondent gave that answer.
Respondents chose to take this survey, and they come from a research panel and VentureBeat's readers. Because the sample is self-selected, the results describe these respondents and may not hold for all enterprises.
Published by VentureBeat Intelligence. © 2026 VentureBeat. All rights reserved.