OpenAI’s Presence Turns “AI Agents” into a Production System—And the Enterprise Agent Stack Just Got a New Interoperability Toll Booth

What changed, and why it matters

Presence is OpenAI’s first “infrastructure-like” agent product—real-time, deployed, governed, and updated in production

OpenAI launched Presence on July 22, 2026 as a deployed enterprise product for “trusted AI agents.” Unlike a chatbot wrapper, Presence is positioned as an end-to-end operating layer for real-time voice and chat agent experiences that can answer questions, resolve issues, use company systems, take approved actions, and escalate to people.

Presence deployments are job/workflow-scoped (e.g., billing issues, insurance claims support, employee IT service requests).
Each agent is given only the required knowledge and system access for that job.
Companies set the policies: what the agent can do, when approval is needed, and when to hand off to a person.
Presence includes a pre-launch verification loop (simulations/evaluation) and a post-launch improvement loop using Codex proposals that teams test and approve before rollout.

The key shift is architectural: OpenAI is no longer selling only intelligence (models). It’s packaging a deployable control surface that sits between agent execution and enterprise systems—exactly the layer incumbents monetize as “workflow platforms.”

Load-bearing metrics from the primary source

OpenAI claims measurable productivity lift: 75% of inbound issues resolved without humans, plus faster reductions in handoffs

Automation rate (inbound issues)

75%

Presence claim: “now resolves 75% of inbound issues without human assistance (within weeks)”

Handoff reduction speed

15 pts

Presence claim: “reduced human handoffs by 15 percentage points in just 10 days”

These two numbers are the economic heart of the product narrative: they frame Presence as a system that can (1) reduce human labor by increasing correct autonomous resolutions and (2) learn quickly in the real environment via evaluation + Codex-driven updates.

What the metrics imply about the “agentic toll booth”

Evidence-backed capability

Presence is designed for verified outcomes + policy adherence (pre-launch evaluation + post-launch guarded updates).

This is what makes “real-time agent action” commercially credible.

Economic bottleneck created

Enterprise AI vendors now need a path to get their agents through OpenAI’s governed deployment surface (or replicate the same controls).

That’s the interoperability toll booth—not the model itself.

Supply-chain map (who provides what, and where Presence inserts)

Presence reroutes the agent supply chain: models → governed agent runtime → enterprise systems access → monitored rollout

Presence as a control-plane insertion point across the agent supply chain

A simplified, production-focused view of where Presence sits between “agent intent” and enterprise action.

Unit: component

1) Model intelligence

OpenAI models are the intelligence layer used inside Presence.

2) Deployment control plane (Presence)

Policies, guardrails, simulations/evaluation, and post-launch Codex-proposed updates.

3) Enterprise systems access

Agents receive only job-required knowledge + system permissions.

4) Execution + escalation

Approved actions, plus escalation to people when outside policy boundaries.

That insertion point matters because it defines which layer becomes the monetizable bottleneck in the “agentic enterprise stack.” If Presence becomes the default path for governed real-time agents, then workflow platforms (and enterprise CX/ITSM suites) may need to either adopt Presence’s deployment pattern or risk losing the most defensible part of the stack: production-grade reliability.

Competition: why incumbents should care even if they keep their UI

Presence threatens incumbents on the layer they can’t fully abstract away: policy-verified real-time execution

Presence directly contests the distribution and switching-cost economics of agentic platforms—because its value is not only “chat capability.” It’s the production governance layer (pre-launch evaluation + post-launch guarded updates) that determines whether agents can safely perform work in real time.

For ServiceNow, the vulnerability is workflow automation being perceived as “less reliable” than a governed agent runtime.
For Salesforce, the vulnerability is CRM service/agent experiences being judged on resolution quality and safe action—not just conversation UI.
For Intuit (and SMB accounting/marketing ecosystems), the threat expands beyond enterprise workflows into small-business adoption via a dedicated ChatGPT program.

If customers can reach measurable labor reduction (75% inbound autonomy; 15-pt handoff reduction in 10 days) with a governed runtime, then platform vendors may face a pricing squeeze unless they match the same operational reliability controls.

SMB distribution shock: OpenAI uses “small business” as a second funnel

The small-business program looks like a distribution moat—potentially redirecting budget away from SMB SaaS tooling

OpenAI also launched a ChatGPT for small business program (tied to ChatGPT Work) with hands-on training, in-person “academies,” and guides designed to be uploaded into ChatGPT Work. It explicitly references GPT-5.6-backed agent work available across subscription plans for small businesses.

Program statistics (from an earlier Small Business AI Jams-style event): 78% built a functional AI workflow in a single day; 42% saved more than five hours a week.
Curated integration/partner tool list explicitly includes Dropbox, Shopify, Intuit, Slack, Atlassian, Wix and more.

For incumbents like Intuit, the distribution risk is not that ChatGPT replaces accounting software overnight. It’s that an easier-to-adopt agent stack can reclassify budgets: more “AI workflow subscriptions” and less incremental spend on add-on automation modules.

Fundamentals check on exposed public peers (directional, not a verdict)

Public platform peers show different financial sensitivity profiles—raising the stakes of agentic switching-cost defense

Selected peer revenue trend (annual)

Directional revenue scale provides context for how much agentic product loss or repricing could matter (not a claim about Presence impact).

Unit: USD

ServiceNow revenue 2023

8,971,000,000

ServiceNow revenue 2024

10,984,000,000

ServiceNow revenue 2025

13,278,000,000

Selected peer revenue trend (annual)

More context for CRM and Intuit scale.

Unit: USD

Salesforce revenue 2023

31,352,000,000

Salesforce revenue 2024

34,857,000,000

Salesforce revenue 2025

37,895,000,000

Selected peer revenue trend (annual)

Intuit’s scale and growth trajectory matters because it anchors SMB spend.

Unit: USD

Intuit revenue 2023

14,368,000,000

Intuit revenue 2024

16,285,000,000

Intuit revenue 2025

18,831,000,000

The reason this matters to the Presence thesis is incentive alignment: larger platforms have more to defend in enterprise governance and in SMB adoption funnels. But because we don’t yet have disclosed financial impact from OpenAI’s launch, these fundamentals are contextual—what to watch next, not what to conclude today.

Causal chain: Presence → reliability → spend reclassification → competitive repricing risk

Presence shifts the winning criterion from “agent demos” to “production outcomes,” which can force repricing

Presence defines agent success as policy-correct outcomes in real-time voice/chat deployments.
The claimed improvements (75% autonomy; 15-pt handoff reduction in 10 days) are outcome metrics tied to labor reduction.
When outcome metrics beat legacy toolchains, buyers can reallocate budgets from point solutions to the governed runtime (or require vendors to bundle it).
That creates repricing pressure on platform components that used to sell reliability indirectly (e.g., via workflow tooling and human-in-the-loop processes).

Upside for buyers: if OpenAI can consistently deliver evaluated, policy-safe autonomy faster than incumbents, enterprises may get measurable ROI sooner than with “LLM pilots.”

Risk for incumbents: UI-layer defenses (CRM/ITSM screens, copilots) may not be enough if Presence becomes the default path to safe, updated real-time execution.

Horizons (what moves first vs. what takes 1–3 years)

Near-term: pilots become deployments; long-term: governance interoperability becomes a product category

What to watch to validate the “toll booth” thesis
Horizon	First observable sign	What it would mean for investors
Days–quarters	Enterprise customers adopting Presence via limited GA rather than self-serve agent builds	Evidence that governed runtime is the bottleneck; potential repricing pressure for workflow components.
Days–quarters	SMB adoption signals (training funnels, partner tool integrations, and usage of GPT-5.6-powered ChatGPT Work)	Evidence that SMB spend reclassifies toward agent workflow subscriptions.
1–3 years	Incumbent platform vendors bundling “OpenAI-style” evaluation/guardrail control planes or offering direct interoperability	If not, customers may demand Presence-like governance for real-time actions.
1–3 years	Convergence of agent monitoring, escalation, and policy evaluation into standardized interoperability layers	New winners likely include vendors who own the control-plane experience, not just the UI.