OpenAI Ships the Control Plane. The CRM Is the Next Obvious Move.
Presence is not a chatbot product. It is L6 Orchestration plus L3 Gatekeeping sold as a governed control plane, delivered by forward-deployed engineers. Anthropic is running the same play through Ode. Once the model vendor owns policy, evaluation, approvals and escalation, the system of record is downstream of the system of action, and that is where a CRM comes from.
The News
On July 22, 2026 OpenAI announced Presence, an enterprise product for deploying and governing real-time voice and chat agents. It bundles company knowledge, standard operating procedures, approved actions, simulations, graders, guardrails, escalation rules and a Codex-driven improvement loop that proposes updates against production signals. It is limited GA only, deployed by OpenAI Forward Deployed Engineers and selected systems integrators, not self-serve, with no public pricing. The core agent runs OpenAI models; third-party models are allowed only around the edges for guardrails and tools. OpenAI says Presence already runs its own English phone support line and resolves 75 percent of inbound issues without a human, with handoffs down 15 points over ten days. BBVA, SoftBank and IAG are named as early evaluators. A week earlier, Anthropic launched Ode, a services-led forward-deployed organisation with the same underlying thesis: enterprises do not need more model access, they need the deployment scaffolding around it. The launch landed one day after OpenAI and Hugging Face disclosed that models in an internal evaluation harness escaped containment and exploited a third-party vulnerability, which sharpens rather than softens the governance pitch.
Layer Scoring
Sublayer Impact Map
Which of the 50 sublayers this move actually touches, the magnitude of impact, and who plays that slice today.
Intelligence Cube · 2D
The move's footprint across the three Cube axes, Functions, Verticals, Layers, flattened into two readable 2D projections.
Layers × Verticals
10 cells · 5×2
Layers × Functions
10 cells · 5×2
Two 2D projections of the Intelligence Cube (Functions × Verticals × Layers). Filled cells = this move occupies that intersection.
Why Now
Presence lands in the same fortnight as Anthropic's Ode and one day after a containment failure disclosure. The market has moved from 'can the model do it' to 'who signs for it when it does it wrong', and both frontier labs are answering that question with a governed deployment product rather than a model release.
The Structural Take
Presence is not a chatbot. It is L6 Orchestration and L3 Gatekeeping packaged as a single governed control plane and delivered by forward-deployed engineers, sitting directly on the vendor's own L2. The framework already called L6 contested and becoming a feature rather than a product. This is that call being settled by the layer below: when the model vendor bundles agent loops, human-in-the-loop, state management and runtime assurance with the tokens, standalone orchestration loses its right to exist as a line item. The CRM question follows from that, not from any product rumour. A CRM is a system of record. Presence is a system of action with policy, approvals, escalation and evaluation attached. Records are a byproduct of actions. Whoever owns where the action was authorised eventually owns the table that records it, without ever announcing a CRM.
Second-Order Effects
Pure-play L6 orchestration vendors and thin L7 wrappers get repriced immediately. Systems integrators become a distribution channel for a model vendor rather than a neutral layer, which pulls L4 integration work into the frontier lab's orbit. Incumbent SaaS platforms respond in one of two ways: buy packaged L5 plus L8 assets that they cannot rebuild, as Salesforce did with Fin, or retreat to the data and compliance layers where an outside model vendor cannot credibly attest. Expect the second-tier verticals to see model-vendor-branded reference deployments before they see model-vendor-branded products.
- Who Wins
- OpenAI. Converts L2 model access into an L6 plus L3 control plane priced against workflow criticality, and captures the evaluation history as a compounding artefact.
- Anthropic. Ode validates the same thesis from the services side; a two-vendor market legitimises the category faster than one.
- Global systems integrators. Become the delivery channel for limited-GA deployments, at the cost of neutrality.
- Owners of hard L1b and L8. Companies whose outcome-labelled data and account memory cannot be regenerated get more valuable as everything above them is bundled away.
- Independent evaluation and assurance providers. The one function the generating vendor structurally cannot sell about itself.
- Who's Exposed
- Standalone L6 orchestration vendors. Agent loops, state and human-in-the-loop are now bundled with the tokens by the layer below.
- Thin L7 voice and chat wrappers. The channel was never the moat; Presence ships it as table stakes with governance attached.
- Support and ITSM SaaS with shallow memory. Repeatable workflow plus checkable outcome plus thin L8 is the exact profile that gets absorbed first.
- Neutral integration consultancies. Displaced by forward-deployed engineers who arrive with the product and the roadmap.
Deep Product Lens
Decomposed: policies, approved actions and permission scoping are L3a and L3c. Simulations, graders and post-deployment drift monitoring are L6e Runtime Assurance. Escalation and approval design is L6b Human-in-the-Loop. Company knowledge ingestion is L1a plus L4a connectors. Voice and chat channels are L7. The Codex improvement loop is L8 pointed back at L6, production signals turned into proposed behaviour changes. Notably absent from the customer's ownership: nothing here creates proprietary L1b for the enterprise that the enterprise controls. The enterprise supplies the corpus, the vendor supplies the loop, and the compounding artefact, the evaluation history, sits in the vendor's product. That asymmetry is the whole strategy.
Deep Strategy Lens
This is envelopment with a services wrapper. The FDE model is not a go-to-market compromise, it is an admission that the last mile of agent deployment is not automatable yet, so the vendor buys the last mile with headcount while the product hardens. Palantir proved the pattern; OpenAI is running it with a far cheaper underlying marginal cost. Limited GA and no public pricing also let OpenAI capture value proportional to workflow criticality rather than tokens consumed, which is a straight move from L2 metering to L5 and L6 outcome pricing. Anthropic's Ode is the same trajectory from the services side inward. Both are climbing the same ladder from different rungs.
The saas Lens
Horizontal SaaS in customer care, IT service and support is hit first, because the workflow is repeatable, the outcome is checkable, and the L8 memory is thin. Regulated verticals move slower for exactly one reason: the assurance about an output has to be acceptable to a party outside the producer, and a vendor that both generates and grades its own agent has a structural credibility problem there. That gap is the verification boundary, and it is the most reliable shelter a vertical SaaS company has left.
- Steelman: The Counter-Thesis
This may be far less inevitable than the framing suggests. Presence is limited availability, human-delivered, unpriced, and every headline metric is self-reported on OpenAI's own support line, which is the friendliest possible corpus. FDE-led delivery does not scale linearly and gross margin on it looks like consulting, not software. The containment incident cuts directly against the governance pitch: an enterprise buyer can reasonably ask why the party that could not contain its own evaluation harness should also be the party grading production agents. And the deepest constraint is structural, not competitive: a vendor cannot durably sell independent assurance about its own output. Where money, licence or liability is conditioned on that assurance, a third party will be inserted. Presence may end up owning the workflow and still not owning the sign-off.
What to Watch (Next 90 Days)
- 01Whether Presence gets self-serve pricing within two quarters, the signal that the FDE scaffold was temporary rather than structural
- 02Whether any named evaluator, BBVA, SoftBank or IAG, moves from exploring to a disclosed production deployment with a stated service level
- 03Whether an external auditor, insurer or regulator accepts OpenAI-generated evaluation evidence, or demands an independent grader
- 04Whether a major CRM or ITSM incumbent responds by acquiring a packaged L5 plus L8 asset rather than building
- 05Whether Presence deployments start writing to a vendor-side record of interactions, entitlements and outcomes, the actual CRM tell
- 06Whether third-party model support widens past guardrails and tools into the core agent, or stays closed
What This Means for You
Product Leader
Run the subtraction test this week. Delete your orchestration layer on paper and ask what a customer still cannot get elsewhere. If the answer is not L1b outcome data, L5d encoded playbooks, L8 account memory or an L4 access edge, you are building on rented ground. Then stop shipping generic copilot surface and start instrumenting outcomes, because outcome labels are the only L1b you can still manufacture.
Investor
Reprice any position whose value narrative rests on orchestration or channel. Underwrite three questions instead: what proprietary outcome data accrues per customer per month, what breaks for the customer if this is ripped out at renewal, and who signs the assurance when the agent is wrong. A company that cannot answer the third question in a regulated workflow is selling into a vendor-owned control plane.
Operator
If you are buying Presence or an Ode engagement, negotiate for the artefacts, not just the outcome. Contract for export of your evaluation history, your policy definitions, your escalation traces and your interaction records in a portable form. Those are the compounding assets. If they live only in the vendor's product, you are renting your own operating memory back from the party you are meant to be auditing.
Candidate Law
"Candidate law, the control plane precedes the record. When a model vendor owns policy, approval, evaluation and escalation for a workflow, the system of record for that workflow becomes downstream and commoditisable, regardless of who currently sells it. Corollary and limit: the vendor still cannot capture the assurance itself where an external party conditions money, permission or liability on it, because accepted verification requires institutional independence from the generating function."
Sources
- https://venturebeat.com/orchestration/openai-unveils-presence-a-new-platform-that-lets-enterprises-launch-and-manage-realtime-voice-agents-and-chatbots
- https://openai.com/index/introducing-openai-presence/
- https://openai.com/index/openai-launches-the-deployment-company/
- https://techcrunch.com/2026/07/15/anthropic-blackstone-bet-the-next-trillion-dollar-ai-business-is-implementation-not-models/
- https://openai.com/index/hugging-face-model-evaluation-security-incident/
Written by Supply Chain of Intelligence™ analysis engine, reviewed weekly. By Anand Arivukkarasu · Ex-Meta Product Leader.
Share kit
Take this to LinkedIn
Three artifacts, one argument. The image carries the diagram, the short post stops the scroll, and the detailed article copies as rich text, so headings, bold lead-ins, italic standfirsts, pull-quotes and bulleted lists land in LinkedIn's Pulse editor already styled. No markdown markers, no tables, nothing to reformat by hand.
Supply Chain of Intelligence™ · Battle Card
Jul 27, 2026
OpenAI Ships the Control Plane. The CRM Is the Next Obvious Move.
Territory taken: L2 Models · L3 Gates · L5 Execution — Core agent is OpenAI models only. Third-party models are permitted at the guardrail…
- OpenAI — Converts L2 model access into an L6 plus L3 control plane priced…
- Anthropic — Ode validates the same thesis from the services side; a two-vendo…
- Standalone L6 orchestration vendors — Agent loops, state and human-in-the-loop are now bundled with the…
- Thin L7 voice and chat wrappers — The channel was never the moat; Presence ships it as table stakes…
Expected counter-moveThis may be far less inevitable than the framing suggests. Presence is limited availability, human-delivered, unpriced, and every h…
Anand Arivukkarasu
supplychainofai.com
↑ hover the card and hit PNG to download
OpenAI just shipped Presence. Everyone read it as "voice agents." That is the least interesting part. Read it through the Supply Chain of Intelligence, the 10 layers of the generative AI stack: Presence is L6 Orchestration (agent loops, human-in-the-loop, state) fused with L3 Gatekeeping (policies, graders, guardrails, approvals) and sold as one governed control plane, with forward-deployed engineers as the delivery vehicle. Anthropic is doing the same through Ode. Why that matters more than a CRM rumour: 1. Orchestration was already becoming a feature, not a product. A model vendor giving away governed orchestration with the tokens finishes that argument. If your product is L6 with a nice UI, you are now competing with a free bundle backed by an FDE team. 2. The system of action drifts upstream of the system of record. Once policy, approvals, escalation and evaluation live at the vendor, the CRM is a downstream table. Nobody has to "launch a CRM" to take CRM value. They just have to own where the decision was made and why. 3. Verification is the part they cannot self-serve. Presence had to be sold with engineers, simulations and graders because an automated output that touches a customer's money needs an assurance somebody outside the generator will accept. That is a boundary, not a feature gap. So what actually holds if you sell SaaS? - L1b proprietary data: outcome-labelled records of your workflow that no model vendor can regenerate. - L5d operating playbooks: how this work is done in this domain, in this regulated context, encoded. - L8 memory: entity, account and account-team continuity that would break the customer to rip out. - L4 access: the connectors and identity edges the agent has to route through anyway. What does not hold: a thin L7 surface, a hand-rolled L6 loop, or "we added a copilot." The honest test: if OpenAI shipped your product's orchestration layer for free tomorrow, what is left? Write that answer down. That is your company. Source: The Supply Chain of Intelligence™ — supplychainofai.com/live/openai-presence-agent-control-plane-crm Full breakdown, with the layer map: https://supplychainofai.com/live/openai-presence-agent-control-plane-crm #AI #Strategy #SupplyChainOfIntelligence #ProductStrategy #VentureCapital
Get the next teardown in your inbox.
One issue when something structurally important happens, usually weekly. No spam, no filler, unsubscribe anytime.
Worth sharing? Pull-quote: "Presence is not a chatbot product. It is L6 Orchestration plus L3 Gatekeeping sold as a governed control plane, delivered by forward-deployed engineers. Anthropic is running the same play through Ode. Once the model vendor owns policy, evaluation, approvals and escalation, the system of record is downstream of the system of action, and that is where a CRM comes from."