OpenAI's Two-Front Expansion: Your Medical Chart Meets the Enterprise Agent:
One week, two launches — ChatGPT starts reading patient health records while OpenAI Presence starts selling agents with engineers attached. Here's what both moves reveal about where enterprise AI is actually headed.
300M+: Weekly health-related ChatGPT queries, per OpenAI
75%: Support tickets Presence resolves without a human
40%: Agentic AI projects Gartner expects cancelled by 2027
From Health Records to Enterprise Agents: Inside OpenAI’s Latest Double Launch.
1: ChatGPT Goes Inside the Medical Record:
OpenAI has moved ChatGPT's Health feature out of its sandbox and into every conversation.
The company is now letting logged-in users 18 and older connect Apple Health data and, where supported, records from US hospital systems, One Medical, or Function Health, directly to ChatGPT. It's live on web and iOS across Free, Go, Plus, and Pro tiers. Once synced, ChatGPT can pull medications, lab results, recent visits, sleep data, and activity logs into whatever conversation is already happening — a dinner recommendation that accounts for a dietary restriction, or a weekend-activity suggestion adjusted for a recent injury.
That design followed the data. OpenAI's earlier, sandboxed version of this feature required users to open a dedicated health area to get grounded answers, but the company found that more than 70 percent of health-related conversations among its test group happened somewhere else entirely — buried inside meal planning or an unrelated symptom question, never in the space built for it. The redesign simply follows the user instead of asking the user to follow the tool.
2: What Early Users Are Actually Saying:
The accounts OpenAI published range from genuinely useful to a reminder of what's at stake. Testers described turning scattered diagnoses and imaging results into a single usable timeline, spotting patterns across years of labs that a single doctor's visit wouldn't surface, and replacing a habit of manually exporting Apple Health spreadsheets into every new chat.
One tester, a nurse, found an unexpected entry in her own chart through the feature — a moment OpenAI is careful to frame as the tool surfacing something for a human to follow up on, not the tool making a clinical judgment.
ChatGPT can access my existing labs because I've connected everything, and it can look over time — that's the big advantage. It's like having a research analyst. It allows me to be more proactive and own more of my health journey.
— Reweti, portfolio manager and early Health tester
OpenAI credits the jump partly to newer models — GPT-5.5 Instant for Free users and GPT-5.6 Sol for paid tiers — validated against HealthBench Professional, an internal benchmark built with hundreds of physicians scoring accuracy, safety, communication, and appropriate escalation. Every GPT-5.6 model reportedly outperformed GPT-5.5 on that benchmark, though the physician testing done on the live product hasn't had its methodology published in detail.
3: The Permission Architecture Behind the Feature:
The access model is opt-in, reversible, and deliberately narrow.
The Hidden AI War
Nobody Is Telling You About
Our latest documentary deep-dive into the geopolitical struggle for machine intelligence dominance. Explore the two paths of AI development: open source vs. closed architecture.
Anthropic’s Surprise Double Launch Directly Targets OpenAI and Google’s AI Dominance
● Connected health data and conversations that draw on it are excluded from foundation model training and ad targeting, regardless of a user's broader ChatGPT training settings.
● ChatGPT asks before using connected data to personalize a response by default; users can switch to "always allow" in Settings > Plugins > Health, reversible at any time.
Support our research
Independent analysis fueled by you.
● Disconnecting a data source triggers deletion within 30 days, though anything already surfaced in existing chats persists until those conversations are deleted manually.
● Memory can be built from health conversations but never directly from raw connected records — Temporary Chat or disabled memory avoids it entirely.
OpenAI also flags the harder edge case: actions that could expose health data through other connected plugins, like sending an Apple-Health-derived training plan to a running partner. Additional checks and, in sensitive cases, explicit confirmation apply — backed by red-teaming the company says it runs but hasn't published results from.
4: OpenAI Presence: Enterprise Agents Built By Hand:
The same week, OpenAI took the opposite approach to selling AI: not self-serve at all. OpenAI Presence, announced July 22, is a managed product delivered through limited general availability, led by OpenAI's own Forward Deployed Engineers and a set of selected systems integrators. Each engagement starts with one job — a billing dispute, an insurance claim, an IT service request — and the agent is scoped to only the access that job requires, with the customer defining escalation rules.
OpenAI's own documentation lays out a six-stage rollout running from scoping through security and legal review, simulation testing, staged deployment, and post-launch iteration by design, not as an afterthought.
The bet behind Presence is a direct response to why agentic AI projects fail in the first place. Gartner has projected that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, and attributes that to governance gaps and undefined business value rather than to model capability. Presence's simulations, guardrails, audit trails, and controlled rollout with rollback are aimed squarely at that failure pattern.
5: Where the Constraint Actually Sits:
Delivery capacity, not model capability, is the bottleneck OpenAI is rationing. Access to Presence depends on workflow fit, implementation readiness, and available delivery capacity — and that last constraint is a consulting problem, not a software one. Software scales; engineers cleared into a bank's core systems do not.
OpenAI's strongest disclosed proof point is its own English-language support line, which the company says now resolves 75 percent of inbound issues without human help. The three named early customers — BBVA, SoftBank, and IAG — are described as design partners exploring specific use cases, not yet running Presence at scale. Pricing and the underlying model configuration also remain undisclosed, set per deployment rather than published.
You Don't Need a Team of Forward Deployed Engineers to Get This Right.
OpenAI's own roadmap makes the point for us: personalized AI and production-grade agents both live or die on permissioning, escalation logic, and human-in-the-loop guardrails — not on model size. Most businesses can't hire a Forward Deployed Engineering team or wait for limited GA. They need that same discipline built in from day one, at a price that doesn't require an enterprise procurement cycle.
Agent+, Otherworlds AI's Business AI Platform, brings that same governed-agent approach to any company for $297/month — powered by Google Opal automated workflows, with guardrails and escalation paths built in, not bolted on later. Need something more tailored? Our custom enterprise AI builds start where Agent+ leaves off.
See what Agent+ can do for your business at otherworldsai.com
