Ryan Johnson @RyanWorkforceAI

A simple test before you launch one: can someone on your team explain what the agent is allowed to do, who approves its work, and how to shut it off? If not, that's your training plan.

This is where L&D and HR have real work to do. Deloitte found only 1 in 5 leaders say their org is ready to redesign processes around autonomous agents. The blockers they named: poorly documented processes, messy data, old habits. Those are ours to work on.

An AI agent can get a badge, a job scope, and a probation period. It shouldn't get the blame. A few notes on borrowing from HR to manage agents, without pretending they're people.

Talk about AI in terms of jobs, and people hear replacement. Talk about tasks and skills, and they can see what changes, what stays, and what they need to learn next. For L&D and HR teams, the job title is the wrong unit for AI planning. The task list is the right one.

autonomaintelligence.com/signit/runbook

Some of your organization’s most valuable knowledge is not in the knowledge base. It is in the heads of experienced people. A new direction for enterprise AI is emerging: Agents that do not just retrieve what experts wrote, they interview experts to uncover what they know. Things like: • which signals they notice first • when the standard process does not apply • which exceptions actually matter • what tradeoffs they make • what would cause them to change their decision That changes knowledge management. The goal is no longer just: “Find what the organization documented.” It becomes: “Capture what the organization knows but never documented.” The hard part will be trust, confidentiality, validation, and making sure captured judgment does not become unquestioned “truth.” Before your next expert retires, there is a valuable question to ask: What do they know how to decide that no document explains?

autonomaintelligence.com/brief/016/

autonomaintelligence.com/brief/015/

Its own examples contain questionable edges The article says: Research + Check sources in parallel That only works when the checker already has a predetermined source set or an independent acquisition mandate. Ordinarily, fact-checking needs to know which sources and claims the research process found. The more accurate graph is: discover candidate sources → retrieve and authenticate sources in parallel → extract claims and passages in parallel → verify each claim → synthesize Likewise, it proposes: Write + Format in parallel Final formatting normally depends on final text. What can run in parallel is preparation of a format scaffold from a locked outline: locked outline ├─ draft prose └─ prepare section/layout scaffold ↓ merge final prose into scaffold This is not nitpicking. False assumptions about independence create duplicate work, stale outputs, or subtle merge errors.

AI can remove busywork. It can also remove the practice that creates experts. Early-career tasks often look inefficient: • first drafts • routine analysis • basic troubleshooting • documentation • research and synthesis But those tasks also teach context, pattern recognition, judgment, and when to escalate. If AI performs all of the foundational work, an organization may preserve senior expertise today while weakening the pipeline that creates its next experts. The workforce question is not only: “What can we automate?” It is also: “What practice must we preserve—or redesign—so people can still become capable?” Automation needs a capability plan.

x.com/i/article/2083…

The final answer is not the audit trail. An AI agent can produce a clean result after taking a messy, risky, or incorrect path. For consequential work, organizations need a record of: • what the agent was asked • which data and context it used • which tools and systems it called • what actions it took • where a human approved or intervened • what failed, retried, or changed That is more than system logging. It is the evidence needed to investigate mistakes, assign accountability, improve the workflow, and decide when the agent should be trusted again. In agentic work, the process is part of the output. #WorkforceAI #AgenticAI #AIGovernance #AIEnablement

An AI agent can have permission to act. It can access accurate data. It can still make the wrong decision. The problem is meaning. Enterprise systems often store values without clearly defining: • which entity the value belongs to • how two records are related • what a metric actually measures • when the information was valid • which policy applies in this situation Access controls answer: May the agent act? They do not answer: does the agent correctly understand what it is acting on? For consequential workflows, organizations need both authorization and machine-readable context. Otherwise, an agent can follow its instructions, use accurate data, and still be organizationally wrong.

autonomaintelligence.com/brief/014/

autonomaintelligence.com/brief/013/

@gdb It is also wired up to OpenAI API/gpt 5.6 terra to read the hand written notes, and it worked really well after a few training runs on various versions of the notes.

@gdb I love that I was able to build an app for a family member - that works 😂 - that ingests handwritten notes and converts them to a CSV that they need, taking what took them hours to minutes.

@thsottiaux @ChatGPTapp I am on a $100 pro plan.

@thsottiaux My usage with @ChatGPTapp is being eaten up when I am doing no coding, or using Work, and I have no scheduled tasks running, it dropped 20% on relatively small Chat usage - what is going on?