The enterprise AI coworker arrives with a job description and an audit trail

A job description for agents

Atomicwork has launched six Certified AI Coworkers for IT service management, and the pitch is that certification should mean something testable. The company says each coworker is evaluated against role-specific benchmarks that include simulated employees and seeded company data, rather than generic capability claims. The Atom AI Coworker was tested against 638 scenarios across 12 support domains, and the Access Ops Manager bench includes 127 scenarios covering various access request inputs. m:…2585

Certification also covers governance. The company says its certified coworkers include deterministic controls for tool and policy scoping, support Cross App Access for agent-to-app connections, and let IT admins set a monthly budget to manage costs. The platform performs weekly drift checks comparing a coworker's runs against expected behavior, and third-party contribution to the marketplace is currently limited to the skill layer. m:…2585

The context is a market full of unproven claims. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, and Gartner estimates that only about 130 of the thousands of vendors calling themselves agentic actually deliver agentic capabilities. Atomicwork's argument is that the truth is in the evaluation, and that buyers should demand it before deployment. m:…2585

Restricted autonomy is already here

ModelOp chief technology officer Jim Olsen makes a complementary point: enterprises are not currently close to achieving full AI agent autonomy, but restricted autonomy is already in use for tasks such as processing refunds and credits. Most IT leaders do not currently trust AI agents for business-critical work, and granting agents full access to email and full autonomy presents significant risk exposure. m:…3319

Olsen's answer is policy, tied to specific agents and use cases. Building trust requires clear policies, and ModelOp's Full-Stack AI Policy Enforcement, integrated with NVIDIA OpenShell, manages policy in a single system of record and enforces it at the process, execution, operating system, and network levels. With OpenShell, leaders can view the compliance status of every agent version in one location. m:…3319

Escalation, not volume

Genie AI's account of its own engineering practice sharpens the same idea: the number of tasks an agent completes is a poor metric for evaluating its utility. What matters is whether the agent stops and asks at the right moments, and Genie programs its agents to identify when they reach the limit of their autonomy. Escalation rules cover cost, architectural decisions, and legal quality, and are stored in a document agents read before each job. m:…1394

The practice is concrete. In October, Genie's Software Factory merged 18 feature flags into a major release and escalated only three or four questions during the merge. The factory switches to night mode after 9pm, deferring non-autonomous decisions to the morning, and its Agentic Legal Specialists system was built by 75 agents working over an afternoon and overnight. m:…1394

The discipline comes from failures. A previous version of Genie's document agent failed by misrepresenting a contractual right as mutual, and by presenting a rewritten clause as a direct quote. Legal agents require firmer enforcement of rules than prompts to prevent silent failures, a warning that applies well beyond legal work, since a confident misquote is the most dangerous kind of agent error. m:…1394

The audit trail closes the loop

Leena AI's Forge shows what the operational side of this looks like. The company says Forge reduces the time to trace failed AI runs from 25 minutes to 15 seconds, by reading procedure logs, run history, and tool responses to explain errors in plain language. For a team handling 20 issues a month across five AI Colleagues, the company says Forge returns roughly 40 hours of work time per month. m:…9311

The underlying workload is real. A security team at a global enterprise uses an AI Colleague on Leena AI for first-pass phishing triage, handling approximately 19 reported emails per day and about 1,100 every two months. Maintaining AI accuracy is difficult because environments, connected systems, and model updates keep changing, which is why Forge makes changes to automations only once the builder approves the proposed fix. m:…9311

The audit world is moving the same way. Unframe's review of internal audit tools notes that 84% of audit functions have adopted audit management software, and that modern tools can extract facts from unstructured documents and compare them against a source of truth. Traditional testing relies on samples of 25 to 40 items to represent thousands; full-population testing analyzes the whole population instead. The certified coworker, the restricted autonomy, the escalation rule, and the audit trail are four answers to the same question: not how much an agent can do alone, but how an organization can prove what it did. m:…5221