Research

VB Pulse Agentic Context Layers Report

78% of respondents traced an AI agent's confident wrong answer to missing or conflicting business definitions in the past six months

October 2026

VentureBeat Intelligence surveyed people who build, run or use AI at organizations with 100 or more employees, and consultants serving such organizations, in October 2026. We asked where AI agents get their business context, whether agents gave wrong answers traced to bad business context, how organizations would find out, and which retrieval platforms they use.

Executive summary

AI agents answering business questions depend on business context: 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.

Of respondents whose organizations have AI agents working on company data, 78% traced at least one confident but wrong answer to missing or inconsistent business context in the past six months. In all, 51% traced more than one.

Asked about their most significant case, 58% of respondents with a context failure said it was caught internally before it left the organization. In the remaining 42% of those cases, a customer, partner or supplier received incorrect information, an incorrect figure reached a financial, regulatory or board report, or a decision or action had to be reversed.

Asked how they would normally find out about such an answer, 57% of respondents whose organizations have AI agents name automated checks against a trusted source. Every respondent who named automated checks also tracks whether wrong answers come from business context or the model.

Of respondents, 39% say they use a governed semantic layer in production. Among them, 28% name it as their agents' primary source of business context. Respondents running, piloting or building a semantic layer were more likely to report a context failure, 88% against 60% of those only evaluating one or with no plans.

Of respondents, 55% use OpenAI's retrieval and 46% use Google Vertex AI Search, and 70% use at least one standalone retrieval tool. Asked about the next 12 months, 77% say the most likely direction is to use standalone tools for some or all workloads.

Finding 1. AI agents' confident wrong answers from bad business definitions keep recurring for most respondents

51% traced more than one.

An AI agent answering a business question works from company data and from the definitions behind that data. When a definition is missing, or two systems disagree, the agent can return a confident answer based on a misreading of the data.

We asked whether, in the past six months, the organization's AI agents had produced confident but wrong answers traced to missing or inconsistent business context rather than to model error. The question named wrong metric definitions and stale or misinterpreted data as examples.

Finding 1 — Traced a confident but wrong answer to business context, past six months

51%
Yes, more than once
28%
Yes, once
11%
No
10%
We don't track root cause at that level

Base: 97 respondents whose organizations have AI agents working on company data. Three who said they do not yet have agents working on company data are left out. One answer each.

In all, 78% traced at least one context failure. Leaving out the 10% who do not track root cause at that level, 87% traced at least one.

Finding 2. In 42% of respondents' worst context failures, an AI agent's wrong answer reached customers, official reports or business decisions

21% say a customer, partner or supplier received incorrect information, 12% say an incorrect figure reached a financial, regulatory or board report and 9% say a decision or action had to be reversed. In 58%, the wrong answer was caught internally.

The question separates wrong answers caught internally from those that reached a customer, partner or supplier, or a financial, regulatory or board report, or led to a decision that had to be reversed.

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

58%
Caught internally before it left the organization, with no external impact
21%
A customer, partner or supplier received incorrect information
12%
An incorrect figure reached a financial, regulatory or board report
9%
A decision was made or an action taken that had to be reversed
0%
Still unresolved
0%
Nothing consequential

Base: 76 respondents who reported at least one context failure, the group the question applies to. One answer each.

We classify the three outcomes involving third parties, official reports or reversed decisions or actions as high-impact failures. Of all respondents whose organizations have AI agents working on company data, 33% reported one.

Finding 3. Every respondent that catches AI agents' wrong answers with automated checks also tracks what caused them

None of the respondents who name automated checks against a trusted source say they don't track whether wrong answers come from business context or the model, against 24% of those who don't name automated checks.

A wrong answer from an agent can be found by a system, such as an automated check or a monitoring alert, or by a person who notices it. This question was moved ahead of the context-failure question in October, so respondents answered it before being asked whether a failure had happened.

We asked how the organization would normally find out that an AI agent had given a confident but wrong answer based on bad business context.

Finding 3 — How the organization would normally find out

57%
Automated checks compare agent output against a trusted source
55%
A subject-matter expert reviews outputs on a sample basis
52%
A dashboard or monitoring alert flags it
39%
An internal user notices and reports it
19%
A customer, partner or supplier notices
3%
We would probably not find out
1%
Don't know

Base: 97 respondents whose organizations have AI agents working on company data, the group the question applies to. Several answers allowed, so shares sum past 100%.

Not counting "we would probably not find out" and "don't know," 31% of these respondents named exactly one way they would find out, and 44% named three or more.

Finding 3 — Does not track root cause at that level

0%
Name automated checks as a way they would find out (0 of 55)
24%
Do not name automated checks (10 of 42)

Bases: 55 and 42 respondents whose organizations have AI agents working on company data. Ten respondents in all said they do not track root cause at that level, meaning whether a wrong answer came from business context or from the model.

Respondents naming automated checks were also more likely to report at least one context failure, 91% against 62%. But when the 10 respondents who don't track whether errors came from business context or the model are excluded, the figures are 91% and 81%, a difference that is not statistically significant.

Finding 4. Respondents investing in semantic layers are more likely to report AI agent failures from bad business definitions

88% of respondents running, piloting or building a semantic layer traced at least one context failure, against 60% of those only evaluating one or with no plans. The survey did not measure why. Among respondents with a semantic layer in production, 28% name it as their agents' primary source of business context.

A governed semantic layer provides shared definitions of business terms and metrics that agents and business intelligence tools can use. The question asked about a governed semantic or context layer; this report calls both a semantic layer.

We asked whether the organization uses a governed semantic or context layer to give agents and business intelligence tools a shared understanding of its data.

Finding 4 — Uses a governed semantic or context layer

39%
Yes, in production
28%
Yes, piloting or building one
21%
No, but actively evaluating
6%
No, no current plans
6%
Don't know

Base: 100 qualified respondents, since the question applies to every organization. One answer each.

Finding 4 — Traced at least one context failure

88%
Running, piloting or building a semantic layer (59 of 67)
60%
Only evaluating one, or no plans (15 of 25)

Bases: 67 and 25 respondents whose organizations have AI agents working on company data. Five who did not know whether their organization uses a semantic layer are left out. The gap remains statistically significant when respondents who do not track root cause are left out.

The difference is statistically significant. The survey did not measure why.

We also asked what the organization's AI agents use today as their primary source of business context when they need to understand company data.

Finding 4 — Agents' primary source of business context today

32%
Retrieval (RAG) over documents or a vector index
25%
Direct queries to live systems (SQL, APIs, MCP)
16%
A governed semantic layer or ontology
14%
A mix that varies by use case
10%
Loading large inputs directly into the model's context window
2%
The model's general knowledge, with no enterprise context layer

Base: 97 respondents whose organizations have AI agents working on company data. Three who said they do not yet have agents working on company data are left out. One answer each.

Among respondents with a semantic layer in production, 28% name it as their agents' primary source of business context, 23% name retrieval and 15% name direct queries to live systems.

Base: 39 respondents whose organizations use a semantic layer in production.

Finding 5. 77% of respondents expect to use standalone retrieval tools even as OpenAI, Anthropic and Google bundle retrieval in

44% expect a mix of provider-native and standalone tools and 33% expect to keep best-of-breed standalone tools, while 15% expect to consolidate onto one model provider. Today, 70% use at least one standalone retrieval tool.

Retrieval systems find the company documents and data an AI model needs to answer a question. Enterprises can use a model provider's built-in retrieval, a standalone retrieval tool such as a vector database, or both.

We asked which retrieval platforms the organization uses today, and which one is its primary platform.

Finding 5 — Retrieval platform

OpenAI retrieval / file search
55%
Use today
32%
Primary platform
Google Vertex AI Search
46%
Use today
16%
Primary platform
Postgres-based vector search (pgvector)
37%
Use today
9%
Primary platform
Elasticsearch / OpenSearch
30%
Use today
5%
Primary platform
Pinecone
29%
Use today
14%
Primary platform
Qdrant
22%
Use today
11%
Primary platform
Custom in-house retrieval stack
21%
Use today
5%
Primary platform
Milvus
15%
Use today
6%
Primary platform
Weaviate
5%
Use today
1%
Primary platform

Base: 100 qualified respondents, all of whom named at least one retrieval platform in use. One respondent named only a platform outside the list, shown under primary as "other" (1%). Several answers allowed for use, so that column sums past 100%; one answer for primary.

Of these respondents, 79% use two or more retrieval platforms, and 46% use both a model provider's retrieval and a standalone tool. For the 70% figure, we counted Pinecone, Weaviate, Milvus, Qdrant, Elasticsearch or OpenSearch, and Postgres-based vector search as standalone retrieval tools.

We then told respondents that model providers are bundling retrieval, memory and orchestration into their platforms, naming 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

44%
A mix: provider-native for some workloads, standalone tools for others
33%
Keep best-of-breed standalone tools alongside the model
15%
Consolidate onto a single model provider's native context stack
4%
No clear direction yet
4%
Build and own our context layer in-house

Base: 100 qualified respondents. One answer each.

In all, 77% say the most likely direction is to use standalone tools for some or all workloads, and 15% say it is to consolidate onto one model provider.

In a separate question, we asked the 66 respondents who expect hybrid retrieval or multiple architectures to dominate by the end of 2027 which approach they expect to carry the largest share of production workloads by then.

Finding 5 — Expected to carry the largest share of production workloads by the end of 2027

30%
Tool-first: direct queries against systems of record
26%
A governed semantic layer that resolves the question before retrieval
23%
Long-context: put the material in the prompt
18%
Vector search
3%
No single approach will carry a clear majority

Base: 66 respondents who expect hybrid retrieval or multiple architectures by use case to dominate by the end of 2027, the only group asked. One answer each. The question is reworded from September and starts a new series.

Finding 6. Retrieval stacks remain in flux: 57% of respondents plan to adopt a new retrieval platform within a year

In the past 12 months, 67% adopted one: 37% added a platform alongside what they had, 18% replaced one and 12% adopted their first.

We asked whether the organization adopted a new, additional or replacement retrieval platform in the past 12 months.

Finding 6 — Adopted a new, additional or replacement retrieval platform, past 12 months

37%
Yes, we added one alongside what we already had
18%
Yes, we replaced an existing one
12%
Yes, our first one
13%
No, we considered it but did not move
13%
No, we did not consider it
7%
I do not know

Base: 100 qualified respondents. One answer each.

We also asked whether the organization plans to adopt a new, additional or replacement retrieval platform in the next 12 months. Of respondents, 19% plan to do so within three months, 28% within three to six months and 10% within six to 12 months. Another 43% have no plans to change.

Base: 100 qualified respondents. One answer each.

Respondents rated their primary retrieval platform 4.02 out of 5 for overall satisfaction, 4.00 for value for money and 3.91 for ease of implementation.

Bases: 100 for overall satisfaction, 97 for value for money, 99 for ease of implementation.

What changed since the September Pulse

September's figures cover only respondents who would pass October's screen; for this tracker that is every respondent in the published September report, so the figures are unchanged. We asked the questions below in the same words in September and October. In October, we moved the question about detecting wrong answers ahead of the context-failure questions.

We checked whether changes in the respondent mix could account for the differences. October's respondents work at larger organizations than September's. Adjusting September's answers to October's mix of company sizes moves September's figures by up to about seven points and does not close either gap labelled a Finding.

The seniority mix also changed: individual contributors were 41% of October's qualified respondents, against 15% in September. Adjusting September's answers to October's seniority mix moves September's figures for each Finding by up to about four points. The respondent profile gives both months' seniority mix.

Adjusted to October's mix of company sizes, September's figures are 27% for tracing no context failure and 24% for naming operational simplicity as the factor that most influenced their choice. Adjusted to October's seniority mix, they are 31% and 18%. Against October's 11% and 7%, both differences remain statistically significant under each adjustment, overall and among in-house respondents alone.

October's respondents came from several sources, including a research panel. Within October, panel respondents differed from the rest on some questions: 57% of them use Postgres-based vector search, against 16% of other October respondents. The data cannot establish how much of any difference between the two surveys reflects who answered, or the new question order, rather than changes in what enterprises do.

Findings: statistically significant changes

Findings: statistically significant changes

27% → 11%
Traced no context failure in the past six months
21% → 7%
Name operational simplicity as the factor that most influenced their choice of retrieval platform

A Finding meets the 95% significance threshold overall and among in-house respondents alone. Bases, September then October: respondents with AI agents working on company data 109 and 97; respondents answering the selection-factor question 111 and 99.

Directional changes

Directional changes

27% → 39%
Use a semantic layer in production
15% → 25%
Name direct queries to live systems as agents' primary source of business context
39% → 58%
Most significant context failure was caught internally
56% → 42%
Most significant context failure was high-impact
21% → 31%
Name retrieval accuracy as the factor that most influenced their choice
32% → 44%
Expect hybrid retrieval to dominate by the end of 2027
25% → 14%
Expect tool-first or long-context retrieval to dominate by the end of 2027
31% → 19%
Plan to adopt a new retrieval platform within three months
4.29 → 4.02
Average overall satisfaction with the primary platform, out of 5
4.12 → 3.91
Average ease of implementation, out of 5
4.21 → 4.00
Average value for money, out of 5

Directional means the difference meets the 90% threshold, or meets the 95% threshold overall but not among in-house respondents alone. Bases, September then October: all qualified respondents 113 and 100; respondents with AI agents 109 and 97; respondents with a context failure 75 and 76. Retrieval questions: 111 and 100 (99 for the selection factor). Ratings: 94 and 100, 91 and 99, 87 and 97.

Platform use, two readings

Retrieval platform used

40% → 70%
Any standalone retrieval tool (margin of error: September 9 points, October 9 points)
5% → 37%
Postgres-based vector search (pgvector) (margin of error: September 4 points, October 9 points)
12% → 30%
Elasticsearch / OpenSearch (margin of error: September 6 points, October 9 points)
14% → 29%
Pinecone (margin of error: September 6 points, October 9 points)
7% → 22%
Qdrant (margin of error: September 5 points, October 8 points)
58% → 55%
OpenAI retrieval / file search (margin of error: September 9 points, October 10 points)
50% → 46%
Google Vertex AI Search (margin of error: September 9 points, October 10 points)

Vendor usage is reported as two survey readings until a second month or a quarterly comparison confirms a change. Bases: 111 and 100. Margins of error are conventional calculations that assume a random sample; this sample is self-selected, so they understate the uncertainty. September's report gave 60% for August. Leaving out October's panel respondents, 65% of October respondents use a standalone tool.

Not statistically significant

Not statistically significant

69% → 78%
Traced at least one context failure in the past six months
41% → 51%
Traced more than one context failure
39% → 33%
Had a high-impact context failure, of all respondents with agents
62% → 67%
Run, pilot or build a semantic layer
12% → 16%
Name a semantic layer as agents' primary source of business context
41% → 32%
Name retrieval as agents' primary source of business context
72% → 77%
Most likely direction uses standalone tools for some or all workloads
21% → 15%
Most likely direction is to consolidate onto one model provider
63% → 57%
Plan to adopt a new retrieval platform within 12 months
5% → 0%
Said nothing consequential happened after the most significant context failure

Not statistically significant means the difference does not meet the threshold described in the methodology. Bases as in the tables above.

In September, respondents running, piloting or building a semantic layer traced more than one context failure at twice the rate of those only evaluating one or with no plans, 51% against 24%. In October, the same comparison gives 58% against 40%, a difference that is not statistically significant.

What we did not compare

The primary-platform question now lists only the platforms a respondent named in use, so its shares are not comparable with September's. The question on how an organization would find out about a wrong answer moved ahead of the context-failure question, and September respondents each gave one answer while October allowed several, so we do not compare it. The question on platforms being considered was skipped in October by respondents with no plans to change, although the questionnaire showed it to everyone, so its base differs from September's. The question on which approach will carry the largest share of workloads was reworded and starts a new series. The question on adopting a platform in the past 12 months is new.

The bottom line: 78% of respondents traced an AI agent's confident wrong answer to missing or conflicting business definitions in the past six months

Of respondents whose organizations have AI agents working on company data, 78% traced at least one confident but wrong answer to missing or inconsistent business context in the past six months. In 42% of the most significant cases, the wrong answer reached a customer, partner or supplier, or an official report, or a decision had to be reversed.

Respondents investing in semantic layers are more likely to report AI agent failures from bad business definitions than those only evaluating one or with no plans; the survey did not measure why. Of respondents with one in production, 28% name it as their agents' primary source of business context.

In the next Pulse, VentureBeat Intelligence will measure whether the share of respondents naming a semantic layer as their agents' primary source of business context changes, and whether standalone retrieval use holds at October's reading.

Respondent profile

Respondent typeRespondents
Works on AI for the organization that employs them59% (59)
Consultant who works on AI for client organizations41% (41)

Base: 100 qualified respondents. One answer each.

Organization size (main client, for consultants)Respondents
100–250 employees22% (22)
251–500 employees14% (14)
501–1,000 employees15% (15)
1,001–5,000 employees20% (20)
5,001–10,000 employees11% (11)
More than 10,000 employees18% (18)

Base: 100 qualified respondents. One answer each.

Job levelRespondents
Individual contributor41% (41)
Manager26% (26)
C-suite12% (12)
Director10% (10)
VP8% (8)
Partner or principal2% (2)

Base: 99 qualified respondents who chose a listed job level; one described their level in their own words. One answer each. In September, the shares among 113 qualified respondents were 15% individual contributors, 28% managers, 23% C-suite, 21% directors, 4% VPs and 7% partners or principals.

Methodology

Who answered. VentureBeat Intelligence fielded this VB Pulse survey in October 2026. The survey received 107 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 analysts, academics and journalists, staff at technology vendors serving their own customers, and organizations under 100 employees. We also excluded consultants without access to their main client's systems, and in-house respondents with no direct role in their organization's AI work.

Consultants. Consultants answered about the client organization where they spent the most time on AI work. Consultants are 41% of qualified respondents, and every figure in this report includes them. Of the consultants, 73% have day-to-day access to that client's AI systems.

Where respondents came from. Of qualified respondents, 51% came through a research panel, 4% through a VentureBeat email and 45% through other links. September's data does not record where 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 retrieval platforms in use and on how an organization would find out about a wrong answer allowed several answers. Their figures count people, not selections.

Flagged respondents. One respondent chose "No production RAG systems" alongside a platform when asked which platforms they are considering. We flag the answer rather than remove it; it does not change any figure in this report.

Small groups. Four groups in this report have fewer than 40 respondents. Respondents with a semantic layer in production number 39, and respondents that did not name automated checks but do track root cause number 32. Respondents only evaluating a semantic layer or with no plans number 25, and respondents who do not track root cause at that level number 10. Percentages for groups this small are less precise.

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 or any count is under five. Ratings use the Mann-Whitney test. A majority uses the exact binomial test against 50%.

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. This check repeats the test without consultants, whose share rose from 39% to 41%; it does not test other changes in who answered. 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 and changed the smallest bands: September's 1–100 and 100–250 overlapped at 100 employees, while October's are 1–99 and 100–250. The rating question was skipped in October by respondents whose only platform answer was "No production RAG systems"; no October respondent gave that answer. The question on platforms being considered was skipped by respondents with no plans to change, although the questionnaire showed it to everyone.

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.