Of enterprises that replaced the model within a production AI application in the past year, 59% saw performance drop and reworked the application, or went back to the original model.
VentureBeat Intelligence surveyed people at organizations with 100 or more employees, and consultants serving such organizations, in September 2026. We asked which agent orchestration platforms enterprises run, what they plan to adopt next, and what happened when they changed the AI model in a production system.
Executive summary
Enterprises are swapping the AI models inside their production apps, and for many of them the swap backfires. Of the enterprises surveyed, 62% replaced the model in a production AI system in the past 12 months while keeping the surrounding application in place. Among enterprises that swapped, 59% saw performance drop and had to rework the application, or went back to the original model.
For enterprises that swapped a model, the swap was seldom a configuration change. Only 19% of enterprises that swapped a model said the swap took almost no engineering effort. Half of enterprises that swapped spent a few weeks or more on the swap, or stopped before finishing.
Consultants see far more model switching than in-house teams do. Of consultants, 81% said their main client had switched models in the past year, against 51% of in-house respondents at their own companies. So the overall 62% leans heavily on what consultants see at their clients.
Few enterprises tie their agent platform to one AI model. Only 7% named how well a platform works with one model as their top reason for choosing it, though that is up from 2% in August.
OpenAI's platform is used by 43% of enterprises that run agent orchestration, more than Anthropic's at 22% or Google's at 21%. Use of Google's platform fell to 21% of enterprises running agent orchestration, from 40% in August. Most of that drop came from enterprises that used Google's platform alongside a different main platform. Google's share as the primary platform was 18% in September, against 24% in August.
Most enterprises expect to keep at least some control of their AI agents outside the model providers' own platforms. By the end of 2027, 72% of enterprises planning to run agents at scale expect to control them through their own tools, an outside platform, or a mix of their own tools and the provider's. Only 28% expect to rely on the provider's platform alone.
Finding 1. Switching AI models in a live app usually meant extra work: Only 19% of enterprises called it almost effortless
Half of the enterprises that swapped a model spent a few weeks or more on the swap, or stopped before finishing.
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 while keeping the application lets an enterprise move to a cheaper or stronger model. 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: 109 qualified respondents, since the question applies to every organization. One answer each.
Of the enterprises surveyed, 62% swapped a model. Among enterprises that swapped, 41% said the new model performed as well or better, 38% saw performance drop and reworked the application, and 21% went back to the original model.
We then asked how much engineering effort the swap took, beyond changing the model itself. We report the answers of enterprises that swapped a model.
Finding 1 — Engineering effort beyond changing the model
Base: 68 enterprises that swapped a model, the group the question applies to. Every respondent saw the question; answers from enterprises that had not swapped are left out. One answer each.
Among enterprises that swapped a model, 29% spent a few weeks, 12% more than a month, and 9% stopped before finishing. Only 19% said the swap took almost no effort.
Finding 1 — Replaced the model, in-house respondents vs. consultants
Bases: 67 in-house respondents and 42 consultants answering about their main client. One answer each.
Consultants reported swaps more often than in-house respondents, at 81% against 51%. Among enterprises that swapped, in-house respondents and consultants reported a performance drop or a return to the original model at rates that are statistically insignificant.
Finding 2. OpenAI's agent platform is used by 43% of enterprises that run agent orchestration, ahead of Anthropic's and Google's
Of enterprises running agent orchestration, 73% use a single platform.
An agent orchestration platform coordinates a company's AI agents: It hands out tasks, calls tools and keeps agents within their permissions. OpenAI, Anthropic and Google each sell an orchestration platform alongside their own models. Open-source frameworks such as LangChain work with models from many providers.
We asked which agent orchestration platforms the organization uses today, then which one is primary.
Finding 2 — Orchestration platforms used and named primary
Base: 107 enterprises that 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.
OpenAI's platform is used by 43% of enterprises running agent orchestration, more than Anthropic's at 22% or Google's at 21%. OpenAI's platform is primary at 25%, and Google's and Anthropic's at 18% and 16%.
Among enterprises using Google's platform, 83% name it their primary platform. The share is 71% among Anthropic's users and 59% among OpenAI's users, gaps that are statistically insignificant.
Finding 2 — Orchestration platforms in use
Base: 107 enterprises that use a listed orchestration platform.
Finding 2 — Rating of primary orchestration platform, average out of 5
Base: 90 enterprises that use a listed platform and rated their primary platform, or 89 for value for money; the rest skipped the ratings.
Finding 3. 63% of enterprises plan to adopt a new, additional or replacement orchestration platform within 12 months, most of them within six months
Of enterprises considering Anthropic's orchestration platform, 60% do not use it today.
Adopting a new orchestration platform means moving or rebuilding agent workflows, connecting the platform to company data and tools, and setting up its permissions and monitoring. A platform an enterprise does not use yet means new integrations and new controls to set up.
We asked whether the organization plans to adopt a new, additional or replacement orchestration platform in the next 12 months, and which platforms the enterprises planning one are considering.
Finding 3 — Plan to adopt a new, additional or replacement platform
Base: 109 qualified respondents, since the question applies to every organization. One answer each.
Finding 3 — Platform under consideration
Base: 61 enterprises planning to adopt a platform within 12 months that named a platform; six more are evaluating without a shortlist, and two named none. Several answers allowed. Platforms named by fewer than 4 enterprises are not shown.
Of enterprises planning a platform that named one, 33% are considering Anthropic's, 30% OpenAI's and 25% Google's, gaps that are statistically insignificant. Among enterprises planning a platform, 78% expect to adopt it within six months.
Finding 4. Only 7% of enterprises name how well a platform works with one AI model as their top reason for choosing it
Flexibility across models and tools was named by 23%, more than three times as many.
An enterprise can pick an orchestration platform to stay free to mix models and tools, or to get the best fit with one provider's model. A platform tied to one model can work more smoothly with it, but moving away later can mean more rework.
We asked which factor most influenced the organization's orchestration platform selection. We report the answers of enterprises that use a listed platform.
Finding 4 — Factor that most influenced platform selection
Base: 107 enterprises that use a listed platform. One answer each. Percentages are rounded, so they sum to 98%.
Alignment with a state-of-the-art base model is named by 7% of agent orchestration platform users, up from 2% in August, a statistically significant change. Agent orchestration platform users naming performance as the main factor rose from 1% to 7%, also a statistically significant change.
Finding 5. 72% of enterprises expect to keep at least some control of their AI agents outside the model providers' own platforms by 2027
Among enterprises planning to split control between a model provider and their own tools, only 22% have written down where the line falls.
The agent control plane decides which agent runs, which tools and data each agent can reach, and what gets logged. A provider-managed control plane runs inside a model provider's own agent service. A hybrid control plane combines the 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: 94 enterprises that expect to deploy autonomous agents at scale, the group the question applies to. Another 15 qualified respondents answered that they do not expect to deploy autonomous agents at scale and are left out. One answer each.
A hybrid approach is chosen by 34% and a provider-managed service by 28%, a gap that is statistically insignificant. Taken together, 72% chose a hybrid, external or custom in-house control plane.
We asked enterprises expecting a hybrid control plane how work is divided between the model provider's built-in controls and their own control layer.
Finding 5 — How a hybrid control plane divides work
Base: 32 enterprises expecting a hybrid control plane, the only group asked. One answer each.
Only 22% of enterprises expecting a hybrid control plane split work by workload and have documented where the boundary sits.
Finding 5 — Top risk if orchestration control lives inside a model provider's platform
Base: 109 qualified respondents. One answer each.
Asked about the top risk if orchestration control lives inside a model provider's platform, 28% chose limited visibility and observability and 22% security and permissioning limitations, a gap that is statistically insignificant.
Finding 6. 38% of enterprises say most of their AI agents now carry out multi-step tasks, not single requests
Of enterprises with agents deployed, 27% say a quarter or fewer of their agents are multi-step workflows.
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 and more model calls to pay for.
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: 104 enterprises with agents deployed, the group the question applies to. Five with no agents deployed are left out. One answer each.
Of enterprises with agents deployed, 24% say 51% to 75% of their agents are multi-step workflows and 13% say 76% to 100%. At the other end, 13% say none are.
Finding 7. 21% of enterprises running AI agents have nothing that automatically stops agents' token spending
31% send their agents' most frequent, routine tasks to cheaper AI models to hold down costs.
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. We report the answers of enterprises that use a listed platform.
Finding 7 — How token spending is controlled
Base: 107 enterprises that use a listed platform. One answer each.
Of agent orchestration platform users, 11% monitor spending only after the fact and 10% have no specific controls, together 21% of agent orchestration platform users. Another 12% do not know how spending is controlled.
What changed since the August Pulse
We asked the questions below in the same words in August and September. The first table shows the measures with statistically significant changes.
What changed: statistically significant
What changed: statistically significant
Statistically significant means the difference passes the test described in the methodology. Bases, August then September: agent orchestration platform users 162 and 107.
Fewer September agent orchestration platform users run Google's platform. More agent orchestration platform users named alignment with a state-of-the-art base model, or performance, as their top reason for choosing their platform.
What did not change: statistically insignificant
What did not change: statistically insignificant
Statistically insignificant means the difference does not pass the test described in the methodology. Bases, August then September: agent orchestration platform users 162 and 107; other rows 166 and 109. Planner rows use all enterprises planning a new platform, including those evaluating without a shortlist, to match August: 100 and 69. Finding 3 leaves out enterprises planning a new platform without a shortlist, so its shares differ.
The bottom line: Most enterprises have switched AI models in a live app, and 59% of those had to rework it or switch back
Of the enterprises surveyed, 62% switched the AI model in a live app in the past year. Among those enterprises, 59% had to rework the app or switch back, and half spent a few weeks or more on the switch or stopped before finishing.
Flexibility across models and tools is named by 23% of enterprises running agent orchestration as the main reason they chose their platform. We did not test whether enterprises that chose a platform for flexibility swap models more easily.
In the next Pulse, VentureBeat Intelligence will measure whether more enterprises that swap a model report the new model performing as well or better, and whether use of Google's platform recovers.
Respondent profile
| Respondent type | Share of 109 |
|---|---|
| Works on AI for the organization that employs them | 61% (67) |
| Consultant who works on AI for client organizations | 39% (42) |
Base: 109 qualified respondents. One answer each.
| Organization size (main client, for consultants) | Share of 109 |
|---|---|
| 100 to 250 employees | 38% (41) |
| 251 to 1,000 employees | 24% (26) |
| 1,001 to 5,000 employees | 16% (17) |
| 5,001 to 10,000 employees | 10% (11) |
| More than 10,000 employees | 13% (14) |
Base: 109 qualified respondents. One answer each.
| Role in AI purchasing | Share of 109 |
|---|---|
| Final decision maker | 33% (36) |
| Recommender or influencer | 41% (45) |
| User | 20% (22) |
| No involvement | 6% (6) |
Base: 109 qualified respondents. One answer each. Consultants answered for their main client's AI purchases.
| Job level | Share of 109 |
|---|---|
| Manager | 32% (35) |
| Director | 18% (20) |
| Individual contributor | 18% (20) |
| C-suite | 13% (14) |
| Partner or principal | 9% (10) |
| VP | 6% (7) |
| Other (small business owner, customer service, none) | 3% (3) |
Base: 109 qualified respondents. One answer each.
| Industry (main client, for consultants) | Share of 109 |
|---|---|
| Retail and e-commerce | 17% (19) |
| Healthcare and life sciences | 17% (18) |
| Technology and software | 14% (15) |
| Manufacturing | 11% (12) |
| Telecommunications and media | 8% (9) |
| Transportation and logistics | 6% (7) |
| Professional services and consulting | 6% (7) |
| Education | 5% (5) |
| Financial services, banking and insurance | 5% (5) |
| Government and public sector | 5% (5) |
| Energy and utilities | 4% (4) |
| Other (construction, wholesale, search quality rating) | 3% (3) |
Base: 109 qualified respondents. One answer each.
Methodology
Who answered. VentureBeat Intelligence fielded this VB Pulse survey in September 2026 and received 144 responses, of which 109 qualified. Respondents qualified if they work at an organization with 100 or more employees, or consult for one as their main client with direct access to its AI systems.
Who did not qualify. We excluded analysts, academics and journalists, staff at technology vendors serving their own customers, and consultants who do not work directly on AI systems. We also excluded consultants without direct access to their main client's AI systems, and organizations or clients under 100 employees.
Consultants. Consultants answered about the client organization where they spent the most time on AI work. Of the consultants, 69% have day-to-day access to that client's AI systems.
Who each figure covers. Each percentage is based on the respondents a question applies to, named in the note under each table. Questions on platforms, ratings, selection factors and token spending use enterprises that use a listed orchestration platform. The control-plane question uses a narrower base, explained in the note under its table.
Contradictory answers. We ran four checks on every qualified respondent. The first two looked for a primary platform the respondent did not say they use, and for platforms under consideration alongside the answer "no plans to change". The other two looked for a plan to adopt alongside the answer "not considering a change", and for listed platforms alongside the answer "not currently using agent orchestration". None of these checks found a contradiction, so no respondent is flagged and no figure needs a second version. We did not treat two screener patterns as contradictions: consultants who work entirely on their own firm's systems yet answered for a client, and in-house respondents with no direct AI role. Four agent orchestration platform users have no agents deployed yet, which we read as building rather than contradiction.
Two respondents who do not use agent orchestration also answered the selection-factor and token-spending questions, and one of them rated a platform. We excluded their answers from all platform figures.
Several answers. The questions on platforms in use and platforms under consideration allowed several answers. Their figures count people, not selections.
Some groups in this report have fewer than 40 respondents: in-house respondents who swapped a model, 34; consultants who swapped a model, 34; enterprises expecting a hybrid control plane, 32. Further groups: enterprises considering Anthropic's platform, 20; users of Anthropic's platform, 24; of Google's, 23; of Amazon Bedrock, 9.
Percentages for groups this small are less precise; for a group of 38, a result could differ by about 15 percentage points either way from the true figure. We report a difference between groups only when a statistical test confirms it is unlikely to be due to chance. The test used for each comparison is listed below. Differences that fail the test are shown and marked statistically insignificant.
Tests. A difference counts only when its test gives p below 0.05. Two percentages from separate groups use the two-proportion z-test when both groups have 40 or more respondents, as in the comparison of consultants and in-house respondents on swapping a model and the comparisons with August. Fisher's exact test is used when either group has fewer than 40, as in the comparisons among users of different platforms and of outcomes among consultants and in-house respondents who swapped. 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 also uses an exact binomial test.
Because the September screening questions differed from August's, a change from August counts only when it passes the test twice: on September's respondents and on September's in-house respondents alone.
Respondents chose to take this survey, and they come from VentureBeat's readers and a research panel. Because the sample is self-selected, the results describe these respondents and may not hold for all enterprises.
Published by VentureBeat Intelligence.