VentureBeat Intelligence surveyed people who build, run or use AI at organizations with 100 or more employees in September 2026. We asked where AI agents get their business context, how often agents give wrong answers because of bad context, and what happens next.
Executive summary
AI agents answering business questions depend on business context, meaning the definitions behind company data, such as what counts as revenue or which customer table is current. When a definition is missing or inconsistent, an agent can give a confident but wrong answer. We call that a context failure.
Most enterprises running AI agents on company data have had a context failure. In the past six months, 69% traced at least one wrong answer to bad business context, and 41% traced more than one.
Many of those wrong answers did real damage. In 56% of the most significant cases, the wrong answer reached a customer, partner or supplier, or a financial, regulatory or board report, or a decision built on it had to be reversed. Across enterprises running AI agents, 39% had a failure with one of those consequences, which we call a high-impact failure.
A governed semantic layer gives agents and dashboards one agreed set of business definitions. Yet enterprises running, piloting or building one report repeated context failures at twice the rate of enterprises without one, 51% against 24%. The survey cannot say whether a semantic layer helps teams find more of their failures, or whether repeated failures pushed them to build one. Only 23% of enterprises with a semantic layer in production make it their agents' primary source of business context.
Retrieval is the technology that feeds company documents and data to AI agents. Use of OpenAI's or Google's retrieval rose to 80% of enterprises running production retrieval, from 66% in August, while use of standalone retrieval tools fell to 40% from 60%. OpenAI or Google runs the primary retrieval platform at 64% of those enterprises. Asked about the next 12 months, 21% expect to consolidate onto one model provider's context tools, and 72% expect to keep standalone tools for some or all of their work.
Enterprises rate their retrieval platforms 4.3 out of 5, and 63% of those running production retrieval plan to adopt a new, additional or replacement platform within 12 months. The most satisfied are the most likely to add or switch: 68% of those rating their platform 5 out of 5, against 46% of those rating it 3 or 4.
Finding 1. 41% of enterprises running AI agents traced more than one wrong answer to bad business context in six months
Another 28% traced one, so 69% have had at least one context failure.
An AI agent answering a business question depends on company data and on the definitions behind that data. When a definition is missing, or two systems disagree, the agent can return a fluent answer built on the wrong figure. Because the agent states the answer with confidence, the person reading it may have no reason to check the figure.
We asked respondents whether, in the past six months, their AI agents had produced confident but wrong answers traced to missing or inconsistent business context. We named wrong metric definitions and stale or misinterpreted data as examples, and asked respondents to leave out errors by the model itself.
Finding 1 — Confident but wrong answers traced to business context, past six months
Base: 109 respondents whose organizations have AI agents working on company data, the group the question applies to. One answer each. Shares are rounded, so they sum to 101%. Leaving out respondents who do not track root cause, 72% traced at least one context failure.
A majority of enterprises running agents, 69%, traced at least one context failure, and 41% traced more than one. Among in-house respondents alone, 63% traced at least one.
Finding 2. 56% of enterprises' most serious context failures reached customers or official reports, or forced a decision to be reversed
Only 39% were caught inside the organization before they did any harm.
A wrong answer caught by the team that uses it can be fixed before anyone acts on it. A wrong answer that reaches a customer, a regulator or a board, or that a decision is built on, may already have been acted on by the time anyone notices.
We asked respondents who reported a context failure to think of the most significant instance in the past six months and say what happened as a result.
Finding 2 — What happened in the most significant case
Base: 75 respondents who reported at least one context failure, the group the question applies to. One answer each. Reached a customer, partner or supplier, or a financial, regulatory or board report, or a decision had to be reversed: 56%.
Of respondents with a context failure, 39% said their most significant case was caught internally with no external impact. In 25% of cases, a customer, partner or supplier received incorrect information, and in 12% an incorrect figure reached a financial, regulatory or board report.
We call a context failure with one of those consequences a high-impact failure. Across all enterprises running AI agents, 39% had one.
Finding 2 — Most significant case was a high-impact failure
Bases: 34 consultants and 41 in-house respondents who reported at least one context failure.
Consultants were far more likely than in-house respondents to report a high-impact failure, 74% against 41%. Because of that difference, the overall figure leans on consultants: among in-house respondents alone, 26% of enterprises running AI agents had a high-impact failure. We did not ask respondents what explains the difference.
Finding 3. Enterprises running or building a semantic layer report repeated context failures at twice the rate of those without one
51% against 24%. The survey cannot say whether a semantic layer helps teams find more failures, or whether repeated failures led them to build one.
A governed semantic layer, or a shared set of business definitions, holds one agreed definition for each business term and metric. Agents and dashboards that query company data then use the same definitions. With a semantic layer in place, a team has an agreed definition to check an agent's answer against, though an agreed definition can still be wrong.
We asked respondents whether their organization uses a governed semantic or context layer to give agents and business intelligence tools a shared understanding of company data.
Finding 3 — Uses a governed semantic or context layer
Base: 113 qualified respondents, since the question applies to every organization. One answer each.
Finding 3 — Traced more than one context failure
Bases: 70, 38, 31 and 39 respondents whose organizations have AI agents working on company data. Each figure is a share of its own group. One respondent who did not know about a semantic layer is left out. The gap between the first two figures holds when respondents who do not track root cause are left out, and among in-house respondents alone.
Enterprises running, piloting or building a semantic layer were twice as likely as enterprises only evaluating one or with no plans to trace repeated context failures, 51% against 24%. Counting any context failure, once or more, the two groups were about even, at 71% and 66%.
We cannot separate the explanations. Teams with a semantic layer have an agreed definition to compare an answer against, so they may find more of their failures. Or enterprises that kept hitting context failures may be the ones that went on to build a semantic layer.
Finding 4. Only 23% of enterprises with a semantic layer in production make it their agents' primary source of business context
Retrieval over documents or a vector index is the most common primary source, named by 41% of enterprises running agents.
Agents get business context in several ways. Retrieval pulls passages from documents or a vector index, direct queries read live systems such as databases, and long-context loading puts large inputs straight into the model. An enterprise can run a semantic layer for its dashboards while its agents draw context from somewhere else.
We asked respondents what their AI agents use today as their primary source of business context when they need to understand company data.
Finding 4 — Primary source of business context
Base: 109 respondents whose organizations have AI agents working on company data. Four who said they have none are left out. One answer each. Shares are rounded, so they sum to 101%.
Finding 4 — Named a semantic layer as primary source
Bases: 31 and 70. Each figure is a share of its own group.
Retrieval over documents or a vector index, known as retrieval-augmented generation, is the most common primary source, named by 41% of enterprises running agents. A governed semantic layer is named by only 12%.
Among enterprises with a semantic layer in production, only 23% make it their agents' primary source, and 32% use retrieval instead.
Finding 5. 80% of enterprises now use OpenAI's or Google's retrieval, up from 66% in August, while use of standalone retrieval tools fell to 40% from 60%
Yet 72% expect to keep standalone retrieval tools for some or all of their work over the next 12 months.
Model providers now bundle retrieval, memory and orchestration into their own platforms. Consolidating onto one provider means fewer tools to connect and one vendor to manage. The tradeoff is dependence, since moving agents to another model later can mean rebuilding the context stack that feeds them.
We asked respondents which retrieval systems they run in production and which one is their primary platform.
Finding 5 — Retrieval systems in production and named primary
Base: 111 respondents who run a production retrieval system. Two who said they run none are left out. Several answers allowed for systems in production, so those shares sum past 100%. One answer each for primary platform.
Use of OpenAI's or Google's retrieval in production rose to 80% of enterprises running production retrieval, from 66% in August. Use of standalone retrieval tools, meaning Pinecone, Weaviate, Milvus, Qdrant, Elasticsearch or OpenSearch, and Postgres-based vector search, fell to 40% from 60%. Both changes are statistically significant.
OpenAI retrieval is the primary platform at 38% of enterprises running production retrieval, and Google Vertex AI Search at 26%. Half of enterprises running production retrieval run more than one retrieval system.
We then told respondents that model providers are bundling retrieval, memory and orchestration into their platforms, citing OpenAI, Anthropic and Google. We asked for the most likely direction for their organization's retrieval and context stack over the next 12 months.
Finding 5 — Most likely direction over the next 12 months
Base: 111 respondents who run a production retrieval system, the group with a retrieval stack to direct. One answer each. Shares are rounded, so they sum to 101%.
Most enterprises running production retrieval, 72%, expect to keep standalone tools for some or all workloads.
Finding 6. The most satisfied retrieval customers are the most likely to add or switch platforms
68% of enterprises rating their retrieval platform 5 out of 5 plan to adopt another platform within 12 months, against 46% of those rating it 3 or 4.
An enterprise satisfied with its main retrieval platform may still add a second one, or replace the first as its needs change.
We asked respondents to rate their primary retrieval platform from 1 to 5, and whether they plan to adopt a new, additional or replacement retrieval platform in the next 12 months.
Finding 6 — Rating of primary retrieval platform, average out of 5
Bases: 94 for overall satisfaction, 87 for value for money, 91 for ease of implementation.
Finding 6 — Plan to adopt a new, additional or replacement platform
Base: 111 respondents who run a production retrieval system, the group with a current platform to add to or replace. One answer each.
Of enterprises running production retrieval, 63% plan to adopt a new, additional or replacement platform within 12 months. Among respondents who rated overall satisfaction 5 out of 5, 68% plan to adopt one, against 46% of respondents who rated it 3 or 4.
Finding 7. Only 13% of enterprises running production retrieval expect vector-only retrieval to dominate their systems by the end of 2027
Hybrid retrieval is expected by 32%, and tool-first or long-context retrieval by 25%.
Vector search was the first standard way to give AI models access to company documents. The alternatives include hybrid retrieval, querying live systems directly and loading material into a long context window. The architecture an enterprise standardizes on shapes which retrieval tools it needs.
We asked which retrieval architecture respondents expect to dominate their production retrieval systems by the end of 2027.
Finding 7 — Expected dominant architecture by the end of 2027
Base: 111 respondents who run a production retrieval system, the group the question applies to. One answer each. Shares are rounded, so they sum to 101%.
Only 13% of enterprises running production retrieval expect vector-only retrieval to dominate. Hybrid retrieval is expected by 32%, and tool-first or long-context retrieval by 25%.
We asked respondents who expect multiple architectures which one they would choose if they had to standardize on one by the end of 2027.
Finding 7 — Would standardize on
Base: 20 respondents who expect multiple architectures by use case, the only group asked. One answer each.
Among respondents expecting several architectures, 90% would standardize on tool-first or long-context retrieval, from four options that did not include hybrid.
Enterprises choosing retrieval infrastructure weigh how easily they can load their data, how accurately the system finds the right passage, and how much work it takes to run. We asked which factor most influenced their choice.
Finding 7 — Factor that most influenced the choice
Base: 111 respondents who run a production retrieval system, the group the question applies to. One answer each.
Ease of data ingestion, retrieval accuracy and operational simplicity each drew 21% to 22% of enterprises running production retrieval.
What changed since the August Pulse
We asked the questions below in the same words in August and September. Six changes are statistically significant.
What changed: statistically significant
What changed: statistically significant
Statistically significant means the difference passes the test described in the methodology. Bases: August 122 and September 107 for the first figure; August 117 and September 109 for the rest. Standalone retrieval tools are Pinecone, Weaviate, Milvus, Qdrant, Elasticsearch or OpenSearch, and Postgres-based vector search.
More September respondents use OpenAI's or Google's retrieval in production, and more named a mix that varies by use case as their agents' primary source of business context. Fewer use any standalone retrieval tool, fewer run Elasticsearch or Weaviate in production, and fewer named ease of data ingestion as the factor that most influenced their choice.
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, so we cannot tell whether the figure moved. Bases, August then September: semantic layer figures 130 and 111; primary source figures 122 and 107; retrieval figures 117 and 109.
The September figures in both groups leave out the two flagged respondents described in the methodology, so a few differ by a point from the findings above. Both months count only the respondents a question applies to, so some August figures differ from those in the August report.
What we did not compare
We did not compare the context failure question, because the August version offered an answer for respondents without agents on company data. We did not compare the expected dominant architecture, because the August version asked about the end of 2026.
We did not compare the expected move to consolidate onto one model provider or plans to adopt a new platform. We have only the published August figures for those two questions, and they include respondents with no production retrieval system. We did not compare primary retrieval platforms. In August, 34% of respondents named a primary platform they had not said they run, so the August answers are not clean enough to compare.
The questions on what happened after a context failure and on standardizing on one architecture are new in September.
The bottom line: Most enterprises running AI agents have had wrong answers from bad business data, and at 39% those answers reached customers or official reports or forced a reversed decision
Among enterprises running AI agents, 69% traced a confident but wrong answer to bad business context in the past six months. In 56% of the most significant cases, the wrong answer reached a customer, partner or supplier, or a financial, regulatory or board report, or a decision built on it was reversed.
Enterprises running, piloting or building a semantic layer report repeated failures at twice the rate of those only evaluating one or with no plans. Only 23% of enterprises with a semantic layer in production make it their agents' primary source of business context.
In the next Pulse, VentureBeat Intelligence will measure whether more enterprises make their semantic layer their agents' primary source of business context. We will also track whether the 39% share of enterprises with a high-impact failure rises or falls.
Respondent profile
| Respondent type | Respondents |
|---|---|
| Works on AI for the organization that employs them | 61% (69) |
| Consultant who works on AI for client organizations | 39% (44) |
Base: 113 qualified respondents. One answer each.
| Organization size (main client, for consultants) | Respondents |
|---|---|
| 100–250 employees | 35% (40) |
| 251–1,000 employees | 32% (36) |
| 1,001–5,000 employees | 21% (24) |
| 5,001–10,000 employees | 8% (9) |
| 10,001 or more employees | 4% (4) |
Base: 113 qualified respondents. One answer each.
| Job level | Respondents |
|---|---|
| Manager | 28% (32) |
| VP or director | 26% (29) |
| C-suite | 23% (26) |
| Individual contributor | 15% (17) |
| Partner or principal | 7% (8) |
| Other | 1% (1) |
Base: 113 qualified respondents. One answer each.
| Role in AI purchasing | Respondents |
|---|---|
| Final decision maker | 60% (68) |
| Recommender or influencer | 29% (33) |
| No involvement | 6% (7) |
| User | 4% (5) |
Base: 113 qualified respondents. One answer each. Consultants answered for their main client's AI purchases.
| Industry (main client, for consultants) | Respondents |
|---|---|
| Technology or software | 30% (34) |
| Financial services, banking or insurance | 13% (15) |
| Healthcare or life sciences | 13% (15) |
| Retail or e-commerce | 11% (12) |
| Manufacturing | 10% (11) |
| Professional services or consulting | 6% (7) |
| Education | 5% (6) |
| Telecommunications or media | 4% (4) |
| Transportation or logistics | 4% (4) |
| Government or public sector | 2% (2) |
| Energy or utilities | 1% (1) |
| Other industries | 2% (2) |
Base: 113 qualified respondents. One answer each.
Methodology
VentureBeat Intelligence fielded this VB Pulse survey in September 2026 and received 170 responses, of which 113 qualified.
Respondents who work on AI for their own organization qualified if they chose a size of 100 to 250 employees or larger for it. Consultants qualified if they chose that size or larger for their main client. We went by the size each respondent chose, so an organization of exactly 100 employees qualified only when the respondent chose the 100-to-250 option.
We excluded analysts, academics and journalists, staff at technology vendors serving their own customers, and respondents who described their role only in their own words. We also excluded consultants who do not work directly on AI systems.
The survey asked consultants to think of the client organization where they spent the most time on AI work in the past six months. Later questions asked about "your organization," so a consultant's answers may describe the client or the consultant's own firm. Of the consultants, 84% have day-to-day access to that client's AI systems.
Each percentage is based on the respondents a question applies to, named in the note under each figure.
Two respondents said their organization is only evaluating a semantic layer, yet named a semantic layer as their agents' primary source of business context. We flag them rather than remove them. Sentences use the full base, and the table below shows the main figures both ways.
| Figure | Full base | Leaving out the two flagged |
|---|---|---|
| Context failure in the past six months | 69% (75 of 109) | 68% (73 of 107) |
| More than one context failure | 41% (45 of 109) | 41% (44 of 107) |
| High-impact failure | 39% (42 of 109) | 38% (41 of 107) |
| Repeat failures, only evaluating a semantic layer or no plans | 24% (9 of 38) | 22% (8 of 36) |
| Expect to consolidate onto one model provider | 21% (23 of 111) | 20% (22 of 109) |
Every finding holds on both bases.
The question on retrieval systems in production allowed several answers, and its figures count people, not selections.
Some groups in this report have fewer than 40 respondents: semantic layer in production, 31; piloting or building one, 39; only evaluating or no plans, 38; consultants with a context failure, 34; multiple architectures, 20. The two checks in Finding 3 use two more groups: only evaluating or no plans, leaving out respondents who do not track root cause, 34; and in-house respondents only evaluating or with no plans, 23. 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.
A difference counts 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. Unless a sentence says otherwise, every difference between groups or between August and September that this report describes passed its test. Differences that fail are described as statistically insignificant or left out of the text.
A change from August counts only when the test confirms it twice: on all September respondents and on September's in-house respondents alone, both leaving out the flagged pair.
In September we changed the screening questions, adding a separate path for consultants and screening out respondents who do not work directly on AI. August's respondents were screened differently, so some movement between the two months may reflect who answered rather than how enterprises changed. To guard against this, every change from August reported here also passed the test among September's in-house respondents alone, who were screened the same way in both months. That includes the rise in use of OpenAI's or Google's retrieval and the fall in use of standalone retrieval tools.
VentureBeat Intelligence surveys a self-selected group of VentureBeat readers and panel respondents. The results describe those respondents, not every enterprise, and apply to the wider market only as far as this mix of respondents allows.