The short answer
In a year of public posts by US enterprise IT and business decision makers, 0.9% of mentions fell under the view that AI vendors are reliable in production. The highest reported shares were for the view that a vendor is only as trustworthy as the lab behind it (13.8%) and the view that vendors are down more often than they can accept (11.3%).
The data types mentioned most often as kept out of third-party AI tools were internal databases an agent could reach (29.7%), customer personal data (28.6%), and source code (21.9%).
Views of the AI vendors they work with
Watchtower grouped 285,988 mentions under the question “What do you think of the AI vendors you work with?” Sentiment is the overall tone of the mentions grouped under each label, from −100% to +100%.
Views of AI vendors, share of mentions
Share of mentions and sentiment for each view
| View | Share of mentions | Sentiment |
|---|---|---|
| Only as trustworthy as the lab behind them | 13.8% | −72.1% |
| Down more often than we can accept | 11.3% | −92.1% |
| Learning our business from our data | 7.0% | +9.2% |
| A trusted long-term partner | 5.9% | +30.7% |
| Worth what we pay | 5.3% | −62.0% |
| Transparent about how models work | 4.5% | +20.5% |
| Too much power over our roadmap | 3.9% | −51.5% |
| Careless with customer data | 3.4% | −100.0% |
| Partnered with our rivals | 2.9% | −100.0% |
| Reliable in production | 0.9% | 0.0% |
| Harder to integrate than promised | 0.4% | −100.0% |
| Unaccountable when things go wrong | 0.2% | −100.0% |
| Fair on pricing | No share reported | −9.1% |
Data kept out of third-party AI tools
Watchtower grouped 152,083 mentions under the question “What data do you keep out of third-party AI tools?”
Data kept out of third-party AI tools, share of mentions
Share of mentions and sentiment for each type of data
| Type of data | Share of mentions | Sentiment |
|---|---|---|
| Internal databases an agent could reach | 29.7% | +36.7% |
| Customer personal data | 28.6% | −81.2% |
| Source code | 21.9% | −72.2% |
| Health information | 5.4% | −5.8% |
| Unreleased product plans | 4.3% | −100.0% |
| Live production systems | 3.9% | −100.0% |
| Regulated records under audit | 2.8% | 0.0% |
| Anything covered by client confidentiality | 2.3% | +100.0% |
| Financial statements | 0.7% | −100.0% |
| Sales pipeline and customer lists | 0.3% | −100.0% |
Three kinds of data account for about four in five mentions.
How the study was run
Intercept Watchtower studies public posts on X, Reddit, Quora, Bluesky, and other public platforms from about 20 million business and IT professionals. The sample is drawn at random using coupling from the past, a sampling method from statistical physics published in peer-reviewed research. AI reviews each account’s full posting history to screen out bots and vendor accounts.
This study covers enterprise IT and business decision makers in the United States, all industries, from August 18, 2025 to August 18, 2026. Each share is the percentage of the question’s mentions grouped under that label. Sentiment is the overall tone of those mentions, from −100% (wholly negative) to +100% (wholly positive). Results are reported in aggregate at a 95% confidence level, with a margin of error of up to 3%.
The data reflects what people say in public. It does not measure beliefs they keep to themselves.
How to cite this study
- Intercept Watchtower, AI vendor trust study, US enterprise IT and business decision makers, August 18, 2025 to August 18, 2026. interceptgroup.com/insights/ai-vendor-trust-study/