The two most valuable AI companies on earth shipped products within eight days of each other this month. Neither product was a model. Neither was a chatbot. Neither was a capability upgrade.

Anthropic launched Ode on July 15, a consulting organization built around forward-deployed engineers who embed inside companies to make Claude work in production. OpenAI launched Presence on July 22, an enterprise platform for deploying and managing AI agents across customer-facing and internal workflows, also staffed by forward-deployed engineers who lead every implementation.

Both companies looked at the same problem and arrived at the same answer: the bottleneck is not the AI. It is getting the AI to function inside the mess of policies, permissions, legacy systems, and human workflows that make up a real business.

That convergence tells you more than any benchmark ever will.

What Presence actually does

Presence is not an API. You cannot sign up for it online. OpenAI has not published pricing. Every deployment starts with OpenAI’s own engineers sitting down with your team, identifying a specific workflow, connecting the agent to your internal systems, defining what it can and cannot do, testing it against edge cases, and then monitoring it after launch.

Each agent gets a defined job. Resolve a billing dispute. Handle an insurance claim. Process an employee IT request. The agent receives only the data and system access required for that task. Your team decides which actions the agent handles alone, which require approval, and when a human takes over.

Before anything reaches production, teams run simulations. The system tests how the agent handles routine requests, unusual scenarios, and policy changes. Graders evaluate whether it followed the rules, used the right tools, and escalated when it should have. After launch, the system keeps monitoring. When something drifts, it flags the issue and proposes a fix that gets tested against the current version before rolling out.

OpenAI already runs this system on its own customer support line at 1-888-GPT-0090. According to the company, it resolves 75 percent of inbound issues without a human ever getting involved. Over a 10-day period, their automated improvement loop cut human handoffs by another 15 percentage points (VentureBeat, July 22, 2026).

Those are company-reported numbers, not independently verified. But even discounted, they reveal something important: OpenAI tested this on itself before selling it. They ate their own cooking.

The Palantir parallel nobody is talking about

If the delivery model sounds familiar, it should. Forward-deployed engineers embedding inside client organizations to make proprietary software work in complex environments is the model Palantir pioneered years ago for government and defense contracts.

OpenAI and Anthropic both independently adopted the same playbook. Not because they admire Palantir’s stock price. Because the problem is the same. Enterprise software that touches real operations, real compliance requirements, and real data cannot be deployed remotely with a README file. Someone has to sit in the room, understand the existing workflow, and wire things together.

This is not what most people expected from AI companies. The narrative for years was that intelligence would scale through APIs. Build a smarter model, expose an endpoint, let customers figure out the rest. That story is over. Cisco data shows 85 percent of enterprises are now piloting AI agents, but only 5 percent have shipped them to production. The gap between those numbers is not a technology gap. It is an operations gap.

What this means if you run a team

Three things to take from this convergence.

First, stop waiting for the model to get better before you deploy. The companies that build the models just told you the model is not what is holding you back. If OpenAI needed forward-deployed engineers to make its own AI work in production, your team is not going to solve deployment by upgrading to a newer API tier.

Second, ask who owns deployment in your organization. Not who picked the vendor. Not who ran the pilot. Who owns the workflow redesign, the permissions model, the escalation rules, the monitoring after launch, and the process for updating the agent when your policies change? If nobody owns that, you do not have an AI deployment. You have a demo.

Third, budget for the boring part. The exciting part of AI is the capability. The expensive part is making it work inside your business. Presence and Ode both exist because companies kept buying API access and then discovering that nobody on the team knew how to turn it into a working system. The forward-deployed engineer model is a direct response to that pattern.

The convergence is the signal

When two competitors who agree on almost nothing both arrive at the same structural conclusion within days of each other, that conclusion is probably real.

OpenAI and Anthropic are not shipping operations products because operations is exciting. They are shipping them because their enterprise customers kept failing at deployment and blaming the model. The model was fine. The operations were not.

BBVA in Mexico, SoftBank in Japan, and IAG in Australia are among the first organizations evaluating Presence. They are not evaluating a new model. They are evaluating a new way to get the model they already have into the workflows their teams actually run.

If you are a business leader still treating AI adoption as a technology decision, this month gave you a clear signal. The companies that make the technology just told you the technology is not the hard part. The hard part is the work nobody wants to do: mapping workflows, defining permissions, testing edge cases, monitoring performance, and updating the system when the business changes.

That is not the exciting part of AI. It is the part that determines whether AI works at all.