On distribution, governance, and what Salesforce’s Headless 360 reveals about the next five years of enterprise AI
TL;DR: The “AI-native startups will flatten incumbents” take is intuitive however a bit incomplete. In enterprise, incumbents with proprietary data, entrenched workflows, and compliance posture have a structural edge - If they stop bolting chat widgets onto dashboards and start treating their own platforms as governed infrastructure for agents. Opening the platform is necessary. But the deeper question - who governs 1000s of agents calling into the mesh? - is where the real lock-in forms.
The Take That Feels Obvious
Every few months, the same argument surfaces in VC memos and tech Twitter/X:
“Enterprise incumbents are toast. AI-native startups will rebuild every category. Distribution doesn’t matter when the product is 10× better.”
It’s intuitive. It’s fasionable. It’s the thesis on which billions have been deployed! It’s been roughly right for fifteen years of SaaS - Workday beat PeopleSoft, Stripe beat a generation of payment gateways, Figma beat Adobe in a category Adobe was supposed to own!
So applying the pattern to the AI era feels like an obvious pattern recognition. However, at a risk of being contrarian ~ I think it’s not a complete picture. Especially, in a specific, economically enormous slice of enteprise software or rather agentic world now!
In the enterprise systems of record, regulated workflows, and proprietary context matters more than generic capability. The incumbents with real distribution have a better shot this cycle than last. Most of them will still fumble it. But the ones who don’t will look, five years from now, less like Kodak and more like Microsoft post-Satya.
Context is harder to build from scratch than software.
What Distribution Actually Means Now
When people say “incumbents have distribution,” they usually mean “they have customers.” That’s the thin version.
The thick version - the one that matters - is primarly four things:
-
Proprietary data no one else can reach. Salesforce has 27 years of deal flow, service case data, and customer graphs. Microsoft has every meeting, doc, and email inside half the Fortune 500. Stripe has the ledger of a meaningful share of internet commerce.
-
Entrenched workflows that encode domain knowledge. The reason a Salesforce implementation takes six months isn’t that the software is hard to install - it’s that every customer has decades of business-logic sediment. That sediment is the actual system. Ripping it out is an organizational project, not a technology one.
-
Compliance posture. SOC 2. HIPAA. FedRAMP. GDPR. The EU AI Act. Each is a multi-year investment. Startups selling into regulated industries routinely die at the compliance review.
-
Buyer trust with real switching costs. Not “our data export is hard” The deeper lock-in: the CIO already spent political capital standardizing on a vendor and doesn’t have another such fight in them.
Here’s the twist: these four things are exactly what AI agents need to do anything useful.
| An agent without… | Is a… |
|---|---|
| Real company data | Trivia machine |
| Knowledge of your workflows | Prompt generator |
| Compliance posture | Liability |
| CIO trust | Demo |
Startups are excellent at producing the first two or probbaly three. Very few have a credible answer to the fourth.
The Trap vs. The Move
So if incumbents have the structural advantage, why does the “incumbents are doomed” take feel so persuasive?
Because most incumbents, when asked to respond to AI, do the wrong thing.
The trap: Bolt a chat widget onto the existing product. Put “AI-powered” in the release notes. Run a press cycle. Declare victory. The workflow is unchanged, the data is still trapped in the same UI paradigm, and the startups competing are building the whole product around AI. No amount of bolted-on chat UX closes that gap.
The move:
Stop treating your product as a UI. Start treating it as a set of primitives any agent can call.
Open the APIs. Support open protocols (MCP, A2A, REST). Expose your proprietary context - data, workflows, rules, permissions, audits - as callable infrastructure. Let Claude hit it. Let GPT hit it. Let the customer’s internal agents hit it. Keep the opinionated runtime, the governance, the audit trail, the business logic on your system. The access layer is open. The value layer stays closed.
This is the AWS 2006 playbook applied to AI. Amazon’s insight was that its infrastructure was more valuable if anyone could use it than if only Amazon could. That move cannibalized internal optionality and compounded into a trillion-dollar business.
The enterprise AI equivalent: your platform is more valuable if any agent can call it than if only your own AI features can.
Case Study: Salesforce’s Headless 360
Salesforce made this bet unusually legibly at TDX 2026, under the banner Headless 360. The products announced matter a bit less than the posture:
“We assume AI agents are first-class users of this platform now. We have digital users working alongside human users and why should digital users be logging on salesforce UI”
That sentence sounds innocuous. It implies architectural, pricing, and organizational commitments most enterprise companies argue about for two years before making.
| Commitment | What it signals |
|---|---|
| Agents can access the full platform without a browser (REST, GraphQL, MCP endpoints GA) | The UI is no longer the privileged interface |
| Agent Script (deterministic control DSL) is open-sourced | Value isn’t in the logic language - it’s in the runtime executing it |
| Experience Layer decouples the agent from the surface (Slack, Teams, voice, ChatGPT, React) | Not defending the UI as a moat anymore |
| Testing Center, Evals, Session Tracing | Betting on measurement - what regulated buyers need before deployment |
None of these is revolutionary alone. Together, they describe a company that has internalized the move - stop being a product, start being a platform for agents or in the case of CRM - be agentic CRM infrastructure provider.
The Pattern Is Wider Than Salesforce
Microsoft. Copilot started as a chat widget - the exact trap. But the real bet is Microsoft Graph + Fabric + MCP support, which lets any agent query across the M365 substrate. That distribution moat is unreplicable.
Stripe. The Agent Toolkit - tool-callable transaction APIs, agent-friendly auth, observability hooks - that make any agent a first-class client of the Stripe ledger. No AI-powered dashboard. Just: we’re making our moat callable by AI.
Adobe is the undecided case. Firefly has a legally clean licensed training corpus plus deep Creative Cloud integration. Open question: will they let non-Adobe agents access that substrate? If yes - they win comfortably. If no - they become an AI-features company in a market that has commoditized AI features.
The pattern: the winners aren’t the ones with the best model. They’re the ones whose proprietary context becomes callable infrastructure for an agent ecosystem they don’t fully control.
The Next Question Nobody’s Answering: Who Governs the Agents?
Opening the platform is necessary. However It’s insufficient.
Here’s the question that lands one layer up:
If a financial instituion say a bank deploys 1000s of agents - across Salesforce, ServiceNow, custom systems, LangGraph, and three other vendors - who controls that fleet? Who decides what each agent can do? Who revokes access when policy changes?
This isn’t hypothetical. In the next 24 months, agent-to-agent traffic in large enterprises will likely exceed human-to-agent traffic. And here’s the problem: existing enterprise governance was built for applications with deterministic code paths and human-granted credentials. It does not extend to agents.
Agents are a fundamentally different resource class. They act with judgment, not deterministic logic. They hold credentials and authority to move money and data. Their behavior changes without code changes - a model update, a prompt drift, a retrieval shift. No single team “owns” them the way someone owns a Salesforce org.
What every enterprise will need - and what almost no one has yet - is an Agent Control Plane: a horizontal governance layer that sits above every vendor and answers five questions:
- Where are all my agents? (A registry - the CMDB for agents)
- Who is this agent? (Verifiable, revocable identity - the Okta for agents)
- What is this agent allowed to do? (Machine-readable policy, enforced at runtime)
- What did this agent actually do? (Immutable audit trail, attributable to the agent)
- Is this agent still certified? (Lifecycle management - create, certify, re-certify, decommission)
This is a new category, not a feature. The same way identity management was a new category that Okta created by sitting above every application rather than inside any one.
Why this matters for the incumbent thesis
The incumbents who understand this won’t just open their platform to agents. They’ll make themselves the most governed, auditable, and interoperable node in the enterprise’s agent mesh.
Salesforce’s positioning here is instructive: don’t pretend to own the control plane - instead, expose every primitive (identity, policy, audit) over open protocols (MCP, A2A, OpenTelemetry) so the bank’s central control plane talks to Agentforce the same way it talks to ServiceNow or a custom in-house developed agent.
The platform that is easiest to govern becomes the platform that CISOs let agents call first. In regulated industries, governability is the unlock - not features, not price, not even data quality.
That’s a moat that compounds. And it’s one that startups - who struggle with compliance at the application layer - will find nearly impossible to replicate at the agent governance layer.
The Honest Caveats
I’ve argued the bull case. Here’s where it breaks.
1. Opening an API/MCP isn’t the same as opening a good one. Inconsistent schemas, bad auth, punitive rate limits, half-hearted MCP implementations - you’ve “opened” the platform and ensured nobody will use it. Developer experience is the difference between infrastructure and shelfware.
2. The governance layer doesn’t exist yet. I’ve argued incumbents should position as governed nodes. But the control plane category is pre-market. Whoever builds it (hyperscaler, independent startup, or incumbent consortium) shapes the rules of multi-vendor agent governance. If incumbents wait to be governed rather than proactively making themselves governable, the window closes. If agent has made an error, who is penalised for it? Customer? Vendor? Model Company?
3. The window is 18 months, not 5 years. Architectural bets compound slowly and then suddenly. Signaling bets compound fast. By mid-2027, either you’ve made it clear agents are first-class citizens of your platform - or you haven’t, and that perception drives renewal decisions for the next five years.
The Scorecard: Serious vs. Theatrical
If you’re trying to read whether an incumbent is serious or theatrical:
| Serious | Theatrical |
|---|---|
| Open protocol support (MCP, A2A) | New chat widget in the existing product |
| First-class agent-as-user auth (not a human token with a wig on) | “AI” prefix on a pricing tier |
| Machine-readable policy and audit infrastructure | Model partnership with no architectural implication |
| Open-sourcing the control/logic plane | Rebranded feature roadmap |
| Making itself governable by external control planes | Press flurry with no changed code |
| Public commitments the CEO can’t quietly walk back | Demo-only “AI capabilities” |
Final Take Away
-
Distribution in the thick sense (data, workflows, compliance, buyer trust) is exactly what AI agents need - and exactly what startups can’t cheaply replicate.
-
Opening the platform is necessary but insufficient. The deeper lock-in forms at the governance layer: the incumbent that is easiest to govern in a multi-vendor agent mesh becomes the one CISOs greenlight first.
-
The category that matters next isn’t “AI features” - it’s “Agent Control Plane.” Whoever builds the horizontal governance layer (registry, identity, policy, audit, lifecycle) for enterprise agents will shape this era the way Okta shaped identity. Incumbents with compliance DNA have a head start. Whether they use it is the open question.
This is one reading of a moving target. The governance layer is pre-market and will evolve fast. If you’re building at an incumbent, inside an AI-native startup, or buying in this space - I’d be interested to hear where this is wrong.