AI agents are pushing enterprises to rewrite security governance, put human accountability on the page, and rethink how oversight works as deployments scale. The findings come from AWS Reimagine 2026, built on confidential interviews with 154 executives at 128 organizations across 23 industries, and show where current policy falls behind the speed of agent work.<\/p>\n
Governance built around six-month IT programs is the wrong shape for projects that finish in days. When a two-week experiment waits a month for approval, some teams stop asking for permission, and policy starts pushing work underground. That is the gap the report tries to close.<\/p>\n
Security, privacy, and data leakage top the technical risk list. Proprietary data, customer personal information, and confidential business intelligence can escape through AI systems, often invisibly and through third-party tools that employees use without oversight. The report cites one leader who pointed out that chief information officers who spent years managing shadow information technology are now dealing with shadow AI at roughly ten times the scale.<\/p>\n
A separate European survey of small and medium enterprises and large enterprises, commissioned by AWS and carried out by Strand Partners, found that more than half already use AI. Only 24 percent have a documented approach to responsible AI use, and 10 percent have a data governance strategy in place. At Boston University, 40 to 50 percent of people use AI at least weekly, with some of it running in approved models and systems and some of it going rogue.<\/p>\n
Emails, chats, and meeting notes form what Reimagine 2026 calls the human layer of data, and the report sees clear value in mining it with AI. Accessing it raises privacy concerns, so the principle that emerged from the interviews is that AI looks at group-level patterns and leaves individuals out. Where salary data, human resources decisions, or personal communications were involved, organizations redacted them before processing so that the system receives the signal without the identity. Being open with employees about how their data was used was the lever that built trust.<\/p>\n
At a Houston hospital, a doctor assumed a new AI project meant leadership was reading his messages. It did not. Staff needed a year of experience with the system before they trusted that it surfaced organizational patterns and left personal content alone. That year of lived experience is the practical trust threshold the report points to.<\/p>\n
Some organizations sort AI projects by risk before they start. At Bradesco, a classification tree asks whether a use case involves personal data, whether it runs live or in batches, and whether a human must stay in the loop. Each combination of answers maps to a risk level and a set of controls.<\/p>\n
For AI agents, the report advises starting with human approval and expanding autonomy only after the agent shows it is reliable, while keeping the option to narrow it again. Treat it like probation for a new hire. Security limits should be set outside the agent, since agents can misinterpret or work around embedded rules.<\/p>\n
Regulatory differences between jurisdictions complicate things further. A chief technology officer at a bank operating in 22 countries warned against a setup where 20 of the countries are impacted because two of the countries do not allow certain things. The report treats this as a design constraint, not an edge case.<\/p>\n
Usage figures say little about results. One organization observed early in its AI journey appeared 88 percent adopted, but produced better work in fewer than one in 5,000 sessions. Saved time also needs a plan. If a worker can do four hours of work in one hour and there is no structure for how the three hours saved will be used, there is no benefit to the company.<\/p>\n
The junior talent gap has no proven solution. AI removes the repetitive work that once built judgment in junior employees, and no organization has reported a fix that closes the resulting skill gap.<\/p>\n
It is an AWS report based on confidential 45 to 60 minute interviews with 154 executives at 128 organizations in 23 industries over nine months. The focus is how enterprises should govern AI agents, keep humans accountable, and handle risk as deployments scale.<\/p>\n
The report flags security, privacy, and data leakage as the top technical risks. Proprietary data, customer personal information, and confidential business intelligence can leak through AI systems, often invisibly, including through third-party tools that employees use without oversight.<\/p>\n
Some sort projects by risk before they start. Bradesco uses a classification tree that asks whether a use case involves personal data, whether it runs live or in batches, and whether a human must stay in the loop, then maps each answer combination to a risk level and a set of controls.<\/p>\n