Key Takeaways
- Automated report generation and boiler‑plate code creation receive the highest trust scores (83.5 and 82.5) because they are tedious, easy to verify, and yield measurable outcomes.
- Trust in agents grows when tasks are scoped, low‑risk, and linked to clear metrics such as merge rates or data‑quality indicators.
- Data‑centric workflows (quality monitoring, stream monitoring, profiling) score in the low‑80s, showing that structured data provides a reliable foundation for agentic decisions.
- The biggest barrier to broader agent adoption is a lack of business context—not raw model capability—making contextual briefing as essential for agents as for new human analysts.
- Complex, multi‑step infrastructure tasks (service‑mesh configuration, disaster‑recovery testing, database migration planning) sit at the bottom of the confidence index (≈38‑45) due to their reliance on live systems and deep organizational knowledge.
- Keeping humans in the loop is the leading mitigation strategy, cited by 59 % of respondents, complemented by activity monitoring and tracing decision inputs.
- Accountability for agent decisions (48 %) and the risk of hallucinations or inaccurate results (47 %) are the top concerns, reflecting a gap between what agents can execute and what teams can explain afterward.
- Worry patterns differ by role: executives focus on accountability, while individual contributors fret most about hallucinations and potential loss of expertise.
- Streamlining everyday processes is seen as the biggest opportunity (51 %), with executives eyeing scale and team leads prioritizing workflow efficiency.
- A strong majority believe that using agents for cloud reliability (96 %) or AI evaluation (92 %) will boost their career prospects, indicating that confidence—not raw capability—drives adoption.
Overview of the Report
MIT Technology Review Insights, in collaboration with Microsoft, released “Agent confidence on the technical frontier,” a study that ranks 101 AI, data, and cloud workflow tasks according to how much practitioners trust autonomous agents to perform them. The research surveyed 300 technology executives, team leaders, and individual contributors from February to March 2026, covering 12 industries and organizations ranging from startups to enterprises with over $10 billion in annual revenue. Respondents evaluated only tasks within their own domain, assigning each a score from 0 to 100 that reflects their confidence level. The resulting index provides a nuanced view of where agents are already trusted, where skepticism remains, and what factors influence adoption.
Highest Confidence Tasks: Automated Report Generation & Boilerplate Code
At the top of the confidence index sits automated generation of business reports and their distribution to stakeholders, scoring 83.5. Closely following is boiler‑plate code generation for new software features, with a score of 82.5. Both tasks share characteristics that make them ripe for delegation: they are repetitive, time‑consuming for humans, and produce outputs that are straightforward to verify. Because the results can be checked against clear criteria—report accuracy against source data or code correctness via merge‑rate metrics—teams feel comfortable handing them over to agents, yielding the highest trust scores in the survey.
How Trust is Built Through Scoped, Measurable Work
The report observes a consistent pattern: the highest‑confidence tasks are those that are narrowly scoped, low‑risk, and tied to quantifiable success metrics. For boiler‑plate code, teams monitor the merge rate into the main codebase as a single indicator of whether the generated code meets quality standards. Similarly, automated report generation can be validated by comparing the output to authoritative data sources. This reliance on objective, easily tracked measures reduces uncertainty and builds confidence, suggesting that agents earn trust not through broad competence but through demonstrable, measurable performance on well‑defined tasks.
Data Workflows as Breakthrough Domain
Data‑centric activities emerge as a strong area of agentic trust. Data quality monitoring scored 82, while real‑time data stream monitoring and automated data profiling each earned 80.5. These scores reflect the fact that data workflows often operate on structured, well‑defined schemas that provide a reliable foundation for decision‑making. When domain experts closest to the data source supply the necessary context—such as knowing which fields represent revenue or how timestamps are normalized—agents can act reliably and produce outcomes that teams trust. The structured nature of data thus serves as a catalyst for higher confidence compared with more amorphous tasks.
Business Context as Primary Blocker
Despite strong performance on technical metrics, the survey identifies a lack of business context as the chief obstacle to broader agent adoption. Even seemingly simple requests—like listing the top ten customers by revenue—require the agent to know which column holds customer identifiers, how the organization calculates revenue, and whether the fiscal or calendar year applies. Without this contextual briefing, agents may produce technically correct but business‑irrelevant results. The report notes that a new data analyst would need the same orientation, underscoring that context, not raw model capability, is the missing link that prevents agents from being trusted with more nuanced responsibilities.
Low Confidence in Complex Multi‑Step Workflows
At the opposite end of the spectrum, tasks that involve live infrastructure, long‑running coordination, and deep organizational knowledge receive the lowest scores. Service‑mesh configuration and troubleshooting garnered a mere 37.5, disaster‑recovery testing 43, and database migration planning 44.5. These activities demand an understanding of inter‑system dependencies, real‑time state, and often‑undocumented operational procedures—knowledge that agents are only beginning to acquire. The uncertainty introduced by dynamic environments and the potential for irreversible mistakes drive down confidence, highlighting a current limitation of agentic systems in handling highly complex, context‑heavy workflows.
Mitigation Strategies: Humans‑in‑the‑Loop and Monitoring
To address their reservations, 59 % of respondents said they plan to keep humans in the loop, allowing agents to propose actions while retaining final human approval. Another 53 % intend to monitor agent activity closely and trace decision inputs, especially for high‑stakes or irreversible scenarios. This hybrid approach leverages the speed and consistency of agents while preserving human oversight for judgment, accountability, and risk management. The survey indicates that organizations view these practices as the primary way to bridge the confidence gap until agents can demonstrate broader reliability.
Top Concerns: Accountability and Hallucinations
When asked about their worries, respondents ranked accountability for decisions made by agents as the top concern at 48 %. Closely following is the potential for inaccurate results or hallucinations, cited by 47 %. Unpredictability of outcomes rounded out the top three concerns. These anxieties point to a shared underlying issue: although agents can execute tasks, teams often struggle to explain or justify the outcomes after the fact. Without clear audit trails or interpretable reasoning, trust erodes, especially in regulated or safety‑critical environments where accountability is paramount.
Role‑Based Differences in Worry and Organizational Response
The concern profile varies by role. Executives, who bear ultimate responsibility for organizational outcomes, focused on accountability at 54 %. Individual contributors, meanwhile, expressed the greatest worry about hallucinations (56 %) and about losing expertise in their craft, with a smaller subset fearing outright replacement. The report argues that organizations should respond by investing in junior talent rather than cutting it, using agents to augment— not replace—human skill development. By pairing agents with mentorship and up‑skilling initiatives, companies can mitigate expertise loss while still gaining efficiency gains.
Opportunities and Career Impact of Agentic AI
Looking forward, 51 % of participants identified streamlining everyday processes as the biggest opportunity agents present for regular work, followed by improvements in performance and reduction of repetitive tasks. Executives tend to value the scalability agents enable, whereas team leads prioritize optimizing their own teams’ workflows. Importantly, a strong majority believe that using agents will enhance their career prospects: 96 % of cloud‑workflow respondents said that relying on agents for system reliability will help them advance, and 92 % of AI‑workflow respondents felt the same about applying agents to evaluation and quality assurance. This sentiment underscores that confidence in agents—not just their raw capability—is the real driver of adoption, and as experience grows and business environments mature, the expectation is that confidence gaps in areas like disaster recovery will gradually close.