OpenAI and Google both spent the last ten days making the same argument from different angles: enterprise AI is no longer about whether an agent can answer a prompt. It is about whether that agent can finish real work, stay inside policy, and improve after launch without turning every deployment into a custom science project.
That is why OpenAI Presence, announced on July 22, matters more than a typical product drop. It is not framed as a new model or a better chatbot. OpenAI positions it as a deployment system for voice and chat agents that can use company systems, take approved actions, and hand work back to people when needed. On the same week, Google said its Agent Development Kit had reached nearly 70 million total downloads and that nearly 90% of the Fortune 100 were using Gemini Enterprise.
Here is the practical read on what changed.
FAQ: What is the actual news here?
The news is that both vendors are shifting the conversation from model quality alone to production operations.
OpenAI Presence packages policies, guardrails, approved actions, simulations, evaluation tools, and a Codex-powered improvement loop into one enterprise offer. Google, in its July 22 earnings remarks, emphasized almost the same ingredients for Gemini Enterprise: agent-building tools, connectivity to enterprise systems, cost management, and governance tools.
The industry is moving from "our model is smart" to "our system can run a workflow without making your compliance team panic."
FAQ: Why does OpenAI Presence feel different from earlier agent launches?
Two reasons stand out.
- It starts with a job, not a general assistant.
OpenAI says each Presence deployment begins with a specific workflow like billing issues, insurance claims, or employee IT requests. That is a stronger framing than launching a blank agent canvas and hoping a team invents its own operating model.
- It treats improvement as part of the product.
Presence is built around post-launch signals, escalations, and proposed updates that teams can test and approve.
The clearest proof point is OpenAI's own support line. The company says Presence now resolves 75% of inbound issues without human assistance, and that its Codex-powered improvement loop cut human handoffs by 15 percentage points in 10 days. Those are the kinds of numbers enterprise buyers can take into a budget meeting.
FAQ: What does this say about how enterprises will judge AI spend now?
The July 17 OpenAI piece called "A scorecard for the AI age" gives away the new buying lens. The core idea is not cost per token. It is "useful intelligence per dollar," which boils down to four questions:
- Did the system complete work that matters?
- What did each successful task cost?
- Can people depend on the result?
- Does value improve as usage scales?
That framework matches the Presence launch almost perfectly. Presence is promising controlled outcomes, testable policies, and a way to lower the full cost of a successful task by reducing retries, reviews, and escalations.
If you are building internal agents, this is the metric shift to copy. Count completed work, not just interactions.
workflow_success_rate = successful_tasks / total_tasks
cost_per_success = total_run_cost / successful_tasks
handoff_rate = human_escalations / total_tasksThose three numbers will tell you more than a dashboard full of token graphs.
FAQ: Why is this happening right now?
Because usage has moved beyond role-locked copilots.
OpenAI's July 27 work research says 43.5% of occupation-specific AI use falls outside the user's own occupation. In plain English, people are already using AI to do tasks that used to require a handoff to another team. That creates pressure for better controls. Once sales, support, ops, and marketing teams all start borrowing technical or analytical work, the weak point is no longer raw model capability. The weak point is how safely that work touches systems, policies, and approvals.
This is also why Google's enterprise numbers matter. High download counts and Fortune 100 penetration do not prove success by themselves, but they do show that large organizations are standardizing around agent platforms instead of isolated experiments.
FAQ: What should product and engineering teams do with this signal?
Do not start by building a broad "company agent." Start by picking one workflow with a visible finish line.
Good first candidates look like this:
- The task already has a human playbook.
- The task requires a small, approved set of system actions.
- The quality bar can be evaluated before full rollout.
- The team can define when a human must take over.
The winning path is not autonomous everywhere. It is governed autonomy in narrow lanes that can widen over time.
FAQ: What is the catch?
OpenAI is explicit that Presence is in a limited general availability program and is not self-serve yet. Deployments are led by OpenAI Forward Deployed Engineers and select global systems integrators.
That means most teams should not read this launch as "everyone should rebuild support next week." The better read is that the enterprise agent stack is consolidating around a repeatable pattern:
- Narrow workflow scope
- Policy and approval boundaries
- Simulation before rollout
- Production telemetry after launch
- Controlled improvement loops
That pattern is durable even if the vendor changes.
Bottom line
The last two weeks gave us a cleaner definition of the category. Enterprise agents are becoming operational systems for completed work, not conversational demos for impressed stakeholders.
If that framing holds, the next winners will not be the teams with the flashiest agent prototype. They will be the teams that can prove a lower cost per successful task, a lower handoff rate, and a safer path from pilot to production.