Leah said on August 13 that its agentic operating system runs legal, contracting, procurement, and finance workflows end to end for more than 400 enterprise customers. It also said its Oracle Cloud move is expected to improve platform performance by up to 30 percent and reduce cloud operating costs by up to 40 percent.
The number is not the interesting part. The workflow claim is.
Most companies still talk about AI as something that assists a person with a task. Leah is describing something different: agents that operate inside commercial workflows, coordinate across systems, follow customer-defined access rules, and hand work back to humans when oversight is required. That is closer to a new operating layer than a better chatbot.
For leaders, the question is not whether this specific platform is the right one. The question is what happens to your organization when AI stops waiting at the edge of the work and starts carrying pieces of the work itself.
The assistant frame is too small now
The easiest way to misunderstand agents is to keep calling them assistants.
An assistant helps a person do something. It drafts a contract clause. It summarizes a vendor negotiation. It finds a missing purchase order. Useful, but still centered on a human operator who knows what needs to happen next.
A workflow owner has a different role. It watches for the condition that starts the process. It knows which system matters. It understands who can approve what. It routes exceptions. It records what happened. It keeps the work moving when no one is staring at the queue.
That is why the Leah announcement matters. The company is not claiming a writing tool for legal or finance teams. It is claiming autonomous execution across legal, contracting, procurement, and finance. Oracle’s release says Leah’s agents securely access authorized enterprise data and operate within customer-defined security controls, access policies, monitoring requirements, and human-oversight processes.
That sentence is dense, but the business implication is plain: agents are moving into the lanes where work actually gets done.
The org chart will lag the work
When AI starts owning workflow steps, the first place leaders look is the org chart. That is usually the wrong first move.
The org chart tells you who manages people. It does not tell you how work moves. A contract review might pass through legal, sales, finance, procurement, and an executive approver before the customer ever sees the final version. No single box on the chart owns the whole path. Everyone owns a piece, which often means nobody owns the delay.
Agents expose that problem because they need the workflow described clearly. What starts the process? What data can be used? Which decisions are safe to make automatically? Which exceptions need a human? Who is accountable if the agent routes the work incorrectly? Where does the audit trail live?
If those answers are fuzzy, the agent does not fix the workflow. It reveals the mess.
I have watched this pattern enough times to be wary of any AI project that starts with the tool instead of the handoff. A leader sees the demo, the team buys the platform, and then everyone discovers the actual work was never documented cleanly enough for an agent to run it. The blocker was not intelligence. It was ownership.
Collaboration means delegation with boundaries
John’s core point on agents is simple: agents are co-collaborators, not tools. That does not mean giving AI vague autonomy and hoping it behaves. It means treating the agent like a participant in the operating model, with a defined lane, permissions, review points, and a memory of the work.
That frame changes the implementation conversation.
If AI is a tool, adoption means training people to prompt it. If AI is a collaborator, adoption means deciding what responsibility can be delegated, what context the agent needs, and how humans stay in the loop without becoming a bottleneck again.
This is where most enterprise AI programs get stuck. They either keep the AI too far from the work, where it can only generate suggestions, or they push it too deep without enough process clarity, where governance teams rightly get nervous. The useful middle is narrower and more practical: delegate a contained workflow step with clear boundaries, then expand only after the team can explain what happened without guesswork.
For example, do not start by asking an agent to “handle procurement.” Start with intake triage for vendor requests under a certain dollar amount. Let it collect missing information, compare the request against policy, route exceptions, and prepare the approval packet. The human still decides. But the human no longer chases the basic information across five systems.
That is collaboration. Not magic. Not replacement. Delegation with boundaries.
What to ask before buying anything
Before a company buys an agentic workflow platform, the better exercise is to pick one recurring process and answer six questions in plain language.
What is the trigger? What outcome should the workflow produce? Which systems contain the truth? Which decisions can be made automatically? Which decisions require a person? Who reviews the agent’s work when something goes wrong?
If the team cannot answer those questions, they are not ready for the platform yet. They are ready for workflow mapping.
The Oracle and Leah announcement gives business leaders a useful signal because it shows where the category is going. Agents are being packaged less like blank AI helpers and more like operating components for real business processes. Oracle made the same move in HR earlier in August with Fusion Agentic Applications that connect work, skills, learning, and workforce planning inside HCM.
That is the direction: AI closer to the process, closer to the system of record, closer to accountability.
The risk is that companies will treat this as another software purchase. They will buy agentic applications and assume the operating model comes included. It does not. The vendor can provide the agent. The company still has to define the work.
The first step this week is not to shop for an agentic operating system. It is to find one workflow that crosses departments and write down who owns every handoff. If that exercise feels harder than expected, good. You found the real work.
Agents are starting to own parts of the workflow. The companies that benefit will not be the ones that give AI the broadest mandate. They will be the ones that know exactly which responsibility they are handing over, exactly where the boundary sits, and exactly who remains accountable when the work moves without waiting for a prompt.