Summary

The UI is losing its monopoly on access, but the system of record remains valuable. Seema Amble calls Salesforce’s Headless 360 largely a rebrand of existing APIs, yet an important acknowledgment that agents may retrieve CRM data without opening Salesforce. The durable asset remains “the data, the logic, everything stored below it.”
Incumbent enterprise software is protected less by screens than by decades of encoded business logic and exceptions. A PostgreSQL database plus APIs cannot simply replace SAP: deployments codify how a 100,000-person, 20-country company operates, complies, and decides. Steven Sinofsky’s blunt warning is that founders “wildly underestimate” the sophistication customers have built into these systems.
“Agent” obscures three economically different jobs: lookup, action, and analysis. Lookup is mostly a more forgiving interface; action raises identity, permission, credential, and paid-seat questions; analysis can span systems and models but requires verification because hallucination becomes consequential. Headless access therefore does not by itself solve enterprise deployment.
The long tail of exceptions—not the routine workflow—is the central agent challenge and a potential source of new product value. Geographic practices, account-specific judgment, permissions, and policies often live in employees’ heads rather than CRM fields. Agents can collect that context by observing calls and computer use, but “almost everything interesting in an enterprise is an exception,” so trust accumulates slowly.
Automation is more likely to expand enterprise software demand than finish a fixed quantity of work. Amazon’s automated returns created a new optimization loop; automating expense and travel workflows can produce new performance analysis; AI-assisted contracts may become longer and more sophisticated. Sinofsky’s framing: “The long tail got no shorter. It just got longer in a different way.”
The strongest startup wedge is between established categories or organizational functions, not directly against a mature incumbent. A head-on replacement inherits “8,000 things” a customer expects, while an AI-native overlay can translate between sales and finance, convert collected data into action, or capture previously invisible field activity. Sinofsky’s instruction is simple: “Aim for the middle and do things in the new way.”
Enterprise AI’s most credible network effects may form inside companies rather than across them. Compliance and security make external networks difficult, but visible wins with chat can spread among colleagues, much as advanced Excel use once did. Products connecting functions that previously needed manual integration could create entirely new categories.

Deep dive

1. Headless software changes the access layer, not the underlying asset

Amble defines headless software through the shift from human-operated workflows to agent access. Traditional applications used interfaces to capture data; an agent may bypass those interfaces, leaving “the data, the logic, everything stored below it” as the valuable layer.
Her Salesforce verdict is deliberately restrained: Headless 360 appeared to be “a marketing announcement more than anything else,” with existing APIs repackaged under a new name. The signal still matters because Salesforce is publicly acknowledging that agents will increasingly access CRM records outside its UI.
Notion may be a more natural headless use case because its users skew more technical and agentic. Burger also widens the definition beyond APIs or MCP: a Slack chatbot can become the CRM interface, and Amble recalls reading that Slack-agent usage had increased by roughly 300%.

2. Lookup, action, and analysis carry radically different risks

Sinofsky calls the market “definitional hell” and jokes that an agent is “a new word for program that takes a very long time to run and might not finish.” The humor masks a useful warning: new terminology is conflating very different capabilities.
Lookup is lightweight and many announced agent APIs merely add a forgiving interface to an old query. Doing something is materially harder: an agent must impersonate a person, hold credentials, respect authorization, and potentially resolve whether it consumes another paid seat.
Analysis best fits the agent model because it can cross systems, run without a tight time limit, route work among models, and compare results. It is also where hallucination matters most: consequential analysis needs a trail showing that “every step of that analysis was correct.”
Amble adds that read/write permissions, interactions between agents, and conflicts over who may access or write to a central source of truth are solvable but will take time.

3. Software becomes sticky after customers make it part of themselves

Amble traces stickiness to frequency, muscle memory, undocumented SOPs, downstream workflows, and external dependencies. A CRM connects sales, marketing, finance, and billing; payroll and ERP add legal and audit requirements around maintaining one source of truth.
Sinofsky’s commercially minded correction is that “the most sticky thing you could do is actually collect money from a customer.” Once software is sold and used, its true moat often emerges only when a customer threatens replacement; repeated objections across accounts reveal what is genuinely indispensable.
Outlook’s delegate access, shared calendars, and recurring-meeting exception handling became displacement barriers without being conceived as a grand retention strategy. Sinofsky says Microsoft might not displace roughly 600,000 seats at General Motors because of calendar behavior.

4. SAP’s moat is business logic, not database gravity

Sinofsky places decades-old insurance software beyond even SAP in stickiness because insurance companies wrote systems that codified external regulatory forces over 50 or 75 years. He uses payment collection similarly: Stripe had to code for countries, tax jurisdictions, currencies, border crossings, and currency exchange, turning a seemingly mundane function into formidable infrastructure.
Amble rejects the idea that PostgreSQL plus APIs can produce an SAP replacement: “That’s absolutely not true.” Multi-year implementations exist because the software captures how a particular business actually operates; the logic matters “way more” than the database containing its records.
Expense reporting illustrates the scale error. A 40-person startup can use one employee, receipt photos, and OCR; a company with 100,000 employees across 20 countries must reconcile national laws, corporate policies, and exceptional cases.
Larry Ellison once argued that businesses should accept an 80% solution instead of customizing everything. Sinofsky’s rebuttal is that operating choices distinguish Ford, Toyota, General Motors, and Daimler: decisions about materials, hedging, hiring, and product lines become ERP logic. Goldman Sachs captured the same phenomenon in telling Microsoft, “We make more money from Excel than you do.”

5. Language models turn enterprise escape valves into usable interfaces

Today’s AI opportunity around SAP is often retrieval and usability, not replacement. A user can connect tables across geographies, ask questions naturally, or generate personalized reports without navigating screens; in Amble’s phrase, “accessing the UI is optional.”
Sinofsky says the two most-used features absent from enterprise software are effectively export to Excel and export to CSV or PDF. Those formats are escape valves: twenty awkward PDFs can now be dropped into a model for cross-currency or exception analysis that previously required repetitive copying and pasting.
Ad hoc workflows matter because they become tomorrow’s products. CRM began as account managers tracking customers in spreadsheets before Siebel and Salesforce productized the process; chat interfaces to SAP or Salesforce similarly exploit models’ ability to synthesize and orchestrate unstructured information.
The limitation is context. Salesforce enforces data collection, but it does not necessarily record why an Asian prospect receives one response and a US prospect another. An outbound agent needs those exceptions, permissions, and unwritten policies—not merely clean fields.

6. Exception handling is where agents must earn trust

Sinofsky observes that no salesperson believes the default answer is right for a particular account. Even a perfectly localized overdue-payment message will be adjusted by the rep; as he puts it, “almost everything interesting in an enterprise is an exception.”
The McDonald’s kiosk supplies the physical analogy: customers start with the standardized flow, then give up when they want an exception, such as mixing two McFlurry flavors. Enterprise pricing works similarly—“How much is it per seat?” still ends with a call and remains an exception.
Amble sees voice agents and computer-use observation as ways to capture exceptions previously stored only in people’s heads. Observation over a few days cannot cover a long sales cycle; buyers must believe the system has seen enough rare cases before permitting autonomous action.
Sinofsky says Amazon demonstrates a different possibility: redefine exception policy rather than encode every old escalation. Its system can simply favor the customer by reshipping the wrong consumable, then use the resulting data to improve descriptions, warehouses, or handling. AI may eventually make such decisions more predictable and repeatable.

7. Automation moves the frontier instead of eliminating the work

Sinofsky rejects the fixed-pie assumption that AI merely replaces n people with software and leaves nothing else to do. Amble argues that even if AI helps contracts get done more quickly, contracts may become longer and cover more scenarios; Sinofsky cites radiology as a correlation-not-causation example in which radiologists embrace AI while there is a radiologist shortage.
Business travel can move from an expense sink to performance optimization and remote-work optimization once tied to how the company performs. “There’s always another layer of analysis on top,” and that analysis creates new processes by which companies differentiate themselves.
Open-source development supplies Sinofsky’s symmetry test: deciding when a release is finished still requires people to concur, preserve a trail, and explain choices. Closing financial books has the same mental model. Businesses remain groups of people deciding things; software changes the level and tools of those decisions.

8. The investable openings sit between systems, functions, and worlds

Sinofsky distrusts the clean middleware diagram. No vendor wants to become a dumb SQL store beneath another company’s interface, while customers do not want a scenario whose stability equals “the most unstable part” of a vendor chain. Large incumbents can also treat a competitive tie as victory by bundling the adjacent feature.
Amble says Workday illustrates the practical barrier: although APIs exist, documentation and access can be difficult, and not all endpoints are exposed. Incumbents are not incentivized to become dumb databases. She outlines three paths: use an incumbent’s agents, rebuild everything internally, or deploy an AI-native layer alongside the system of record.
The first path has mixed prospects because incumbents do not want to become background data stores; Amble says she is not bullish on incumbent software building great agents on top. The second resembles “open-heart surgery” on a live business. The third preserves existing logic while learning from operations, collecting new records, and potentially becoming a new system of record over time.
Amble’s startup model moves from collection into action: prioritize leads, identify churn risk, send outbound messages, observe responses, and learn which language works by geography. Construction, manufacturing, and other physical-world verticals add valuable data from what humans and machines do in the field.
Sinofsky’s preferred strategy is to sit between two incumbents or two organizational functions while legacy vendors bolt AI onto existing products. HTTP and HTML won by implementing computing differently, not by matching every client-server feature. Internally, chat can spread through visible wins—the modern echo of coworkers crowding around an early Excel user.

Adam Wyden: buying someone else's pain in Stagwell $STGW and Driven Brands $DRVN | ADW Capital Everything in Capital Markets is Downstream of Algorithms