The Enterprise AI Trust Gap Is Now a Deployment Problem

Adoption keeps rising while trust falls, and three October reports show how vendors, benchmarkers and buyers are being pushed to verify AI outcomes rather than assume them.

When Usage and Sentiment Move in Opposite Directions

Enterprise and consumer AI adoption is not slowing down. MIT Tech Review's October 5 analysis, drawing on Pew Research Center data, reports that half of US adults now use a chatbot—more than double the 2023 share—and one in four use one daily. Sensor Tower figures cited in that piece put ChatGPT at a billion monthly users in May and Google DeepMind's Gemini at 950 million in July. More than a third of adults across all 38 OECD countries reported using generative AI tools in the prior three months.

The sentiment data runs the other way. Pew found more US adults expect a negative personal and societal impact from AI than a positive one, with pessimism strongest among younger people. Stanford research cited in the piece found more than half of people worldwide say AI products make them nervous. A May Gallup poll found 71% of US adults would oppose a new AI data center locally—against 53% who would oppose a nuclear power plant. All 50 US states have passed or proposed AI laws, producing more than 2,100 bills, a tenfold increase in three years, according to the MIT Tech Review analysis.

The author, Will Douglas Heaven, frames his central claim as interpretation rather than measured finding: the dislike is aimed not at the technology but at companies pushing it into as many parts of life as possible. That distinction is worth holding carefully. The same coverage notes the correlation—the Global North, where adoption is highest, skews negative, while the Global South is more optimistic—without establishing causation.

What is established is more mundane and more useful: adoption data and sentiment data must be read together. Neither substitutes for the other.

A Benchmark That Grades Databases, Not Sentences

The strongest technical evidence for why trust is fraying arrived two days earlier. Microsoft and Hugging Face released ThinkingBox, an agent sandbox, and ThinkingBox-Bench, which grades AI agents on the terminal state and side effects left in a backend database rather than on the sentences they generate, according to the joint blog post covered here.

The results are unflattering. Across a common-set ablation covering 121,680 valid trials across 12 LLM models, 79,853 attempts failed the executable checks. Of those failures, 67.24% still terminated cleanly, invoked a state-changing tool and reported no final tool error. In other words, roughly two-thirds of failed work looked like finished work.

The failure diagnostics are dominated by tool handling at 79.9%, ahead of wrong state updates at 10.3%, incomplete user resolutions at 7.0% and no state-changing action at 2.9%. The post describes these as unweighted averages of per-model shares and observable labels, not unique causal explanations.

The consistency data cuts against headline rankings. Claude Opus 5.5 leads the task-weighted pass@1 table at 67.16%; Kimi-K3 is the strongest open-weights model at 57.37%. But Kimi-K3 solves 93.89% of the benchmark at least once while succeeding on all 20 attempts for just 68 of 507 tasks. Claude Opus 5 solves fewer tasks at least once, 79.09%, but completes 47.53% of the benchmark on every attempt.

Costs diverge in the same way. GPT-5.6 Sol has the lowest cost per single success at $0.127, yet sits at $9.76 per dependable task. GPT-5.4 is cheapest per dependable task at $6.80, though only 128 tasks clear the full-20 bar. The post states these are modelled comparative indices at undiscounted list rates, not invoices.

The Gap Between Tool Access and Verified Outcomes

If the benchmark explains why trust is eroding, the SMB software market shows the commercial response. Salesforce's October 5 guide frames small business AI as early-stage but already funded: three out of four small businesses are investing in AI, while 88% of AI-engaged SMBs remain in the exploring phase, per the vendor's SMB guide covered here.

Salesforce attributes that gap to complexity rather than willingness to spend, and positions no-code tools as the fix. The article also cites a research finding that 91% of SMBs with AI report it boosts revenue, with gains concentrated among businesses that implemented AI in a connected, systematic way rather than through point solutions. That figure is correlational in the article's own telling; the direction of causation is not established.

The guide draws a distinction that matters operationally. A no-code AI tool automates a specific task—drafting an email, summarising a record, routing a lead. An AI agent is autonomous: it takes a sequence of actions, makes decisions and completes multistep workflows without the user initiating each step. Those two categories require different review, error-handling and escalation rules.

Salesforce also describes its Einstein Trust Layer as protecting customer data with enterprise-grade security even when that data powers AI features. That is a vendor description of its own architecture, and the coverage says so explicitly. Procurement teams are advised to verify data location and processing terms against their own requirements rather than treat the description as third-party validation.

What Bayer's Ohio Timeline Says About Capital Patience

Not every AI-era trust problem is computational. Bayer's plan to invest 2.2 billion US dollars in a new pharmaceutical manufacturing site in New Albany, Ohio, announced October 3, sets a longer clock. The company expects the first drug substance module to become operational in 2031, with a second drug product module planned for 2034, as reported here.

Bayer describes a flexible, modular campus combining drug substance and drug product manufacturing, initially supporting oncology, cardiovascular and renal care. The company estimates around 600 high-value jobs once operating and roughly 1,500 construction jobs during the build. Both dates are company projections about future milestones, not confirmed completions.

The comparison is not that pharma and software share a trust problem. It is that both now ask stakeholders to accept stated intent ahead of demonstrated outcome. Bayer frames the site as strengthening resilience across its global Product Supply network—a claim the coverage flags as being about intended rather than demonstrated outcomes. The source does not disclose production volumes, capital phasing by year, expected returns or how the new capacity relates to current utilisation at existing sites.

For procurement and supply-chain teams, the practical signal is timing. A US drug substance module targeted for 2031 will not relieve current supply constraints, so sourcing strategies should not assume this site is available within existing planning cycles.

The Verification Question Buyers Should Ask Now

Four announcements, four different verification problems. When a user trusts an AI feature, a model, a carrier exception remains open behind a ticket marked solved: 67.24% of ThinkingBox failures terminated cleanly with no final tool error. When an SMB buys a no-code platform, security assurance arrives as vendor architecture description. When a state approves a data centre, 71% of neighbours may object. When Bayer books 2.2 billion dollars, revenue-bearing capacity sits five and eight years out.

The common thread is not that any of these claims is false. It is that each is stated intent, and the evidence to confirm it arrives later—sometimes much later.

That produces a narrow, practical checklist. Check terminal state before committing rather than trusting a model's summary of it. Classify tool and system errors so retries target recoverable ones. Narrow the tool surface to what the workflow needs. Treat a repeat rate, not a single success, as the design input. Verify vendor security claims in contract terms. Read adoption figures and sentiment figures together.

The ThinkingBox post is explicit that the lift from its own recommended remedies has not been measured on the benchmark, and that no third-party verification or independent replication is claimed. Deployment decisions should treat it as an evaluation environment to reproduce rather than a settled ranking. The evidence to watch is whether documented reliability, not capability, becomes the metric that buyers demand—and whether vendors can produce it.

Related reporting: Bayer Plans 2.2 Billion Dollar Ohio Manufacturing Site · Thinkingbox Benchmark Grades AI Agents on Database State · AI Use Climbs as Public Sentiment Sours, MIT Tech Review Says · Salesforce Maps No-code AI Tools for Small Business · NVIDIA Details Agent-built Simulation Projects on Omniverse Libraries · Balyasny Deploys Gemini Models Across 200 Investment Teams

Evidence note: This analysis connects four previously published Business 2.0 news posts. It does not represent additional independent source-page research. Reported company claims, forecasts and planned outcomes remain attributed and qualified.