Agents make the UI optional, but business logic stays sticky

1. Value sits below the UI — ▶ 2:04

   Headless software places value in the data, logic and storage beneath the interface, not in workflow screens built for humans. Amble judges Salesforce's Headless 360 mostly a marketing announcement: as far as she could tell, it repackaged APIs that already existed. The broader trend matters more. Agents increasingly reach systems of record directly, and she cited a reported 300% rise in Slack agent usage, something she recalled reading, as a sign that users prefer data delivered to them over logging in. Notion's headless product makes more sense to her because its users are more technical and more likely to build agents.

2. Lookup, act, analyze — ▶ 7:42

   Sonowski argues that headless and agent are conflated, and that three different tasks need separating. Lookup is lightweight, and many new headless APIs are just a more forgiving interface to old queries. Acting on a system of record raises issues of impersonation, credentials and whether the agent consumes a paid seat. Analysis spans multiple systems and suits agents because it is not time-bounded and can iterate or compare answers across models. It is also where hallucination is most damaging, because every step must be verifiable. Deployment maturity differs across the three.

3. Stickiness is accidental — ▶ 11:53

   Sonowski says the stickiest software is software that collects money, and that its causes of stickiness are discovered, not designed. Different stakeholders cite different reasons: HIPAA compliance, administrator onboarding, user muscle memory. A rep learns the real reason when a customer threatens to leave. His example is Outlook: delegate access and shared calendars, which no one deliberately built as a moat, kept customers like General Motors from displacing 600,000 seats. Amble adds inertia, undocumented standard operating procedures, and legal and audit requirements in ERP and payroll.

4. SAP codifies the company — ▶ 17:57

   Sonowski argues that SAP is not where a company's data happens to sit; the company's business rules are codified in it. Ford, Toyota and General Motors buy from the same suppliers, so what separates them is which screens and customizations managers choose. Expense reporting scales from one person handling 40 employees to 100,000 people in 20 countries with differing laws. Larry Ellison's multi-year rant for an 80% solution ran into this reality. Amble calls the notion that Postgres plus APIs can replace SAP absolutely untrue, since logic matters far more than storage.

5. Vibe-coding misses the hard part — ▶ 22:50

   Amble described a head of revops at a company of more than 1,000 people tasked with rebuilding the firm's Salesforce internally, who assumed the job was importing known fields and data. She argues the hard parts are deciding what gets captured, how it maps to organizational lines, and who maintains it over time. Vibe-coded projects, she notes, often go stale. Sonowski adds that customers apply great sophistication to their tools. A Goldman Sachs banker once told Microsoft, "We make more money from Excel than you do," referring to the add-ins and code the bank had built.

6. LLMs unlock the export button — ▶ 25:25

   Sonowski says enterprise software can usually produce any report, but users cannot configure it. The most used features are export to Excel, CSV or PDF. Language models now consume those exports, so 20 PDFs of declined expense reports across odd periods or currencies can be analyzed in a model rather than by copy-paste. He sees ad hoc processes as the source of new products, noting CRM began as spreadsheets and its predecessor was Siebel. Chat interfaces to SAP or Salesforce exploit models' strength at synthesizing unstructured information, and Amble describes retrieval without SQL as the main use today.

7. Exceptions are the context — ▶ 28:45

   Amble defines the context agents need as exceptions, edge cases, permissions and policies that live outside CRM fields, such as responding one way to customers in Asia and another in the US. Sonowski adds that almost everything interesting in an enterprise is an exception. Capture routes include voice agents, such as one a portfolio company uses on freight compliance calls, and computer-using agents observing humans. It is slow: exceptions occur infrequently and sales cycles are long, so vendors must observe enough interactions, and buyers must trust that the software has captured the context to act.

8. Automation lengthens the tail — ▶ 32:43

   Amazon's chatbot resolves wrong-item complaints in the customer's favor, then uses the data to improve shipping, warehouses and product descriptions. Sonowski argues productivity drives new scenarios, so the long tail "just got longer in a different way." Expense reporting grew into travel analysis, and he predicts AI-assisted contracts will become longer and more sophisticated, with more litigation. His radiology example, where radiologists like AI yet a shortage exists, he concedes is correlation not causation. The evidence is analogy, and nothing shows new work will outweigh displaced work in any given function.

9. Middleware rarely stays neutral — ▶ 46:09

   Sonowski argues MCP reflects an engineer's ideal architecture, but no vendor accepts being reduced to a back end: storing SQL for expense reports is a decaying business. Customers also resist assembling stacks from many providers because a system is only as stable as its least stable part. Amble cites Workday, whose APIs exist but whose data is hard to extract cleanly. She sees three paths: build agents on the incumbent, with mixed results; DIY, which she likens to open-heart surgery on a live patient; or build alongside existing systems, gradually observing the business.

10. Aim between incumbents — ▶ 52:25

    Sonowski calls head-on competition the dumbest move, because customers demand thousands of features before a startup enters the door. During a technology shift, he argues, incumbents will not disturb their product line or go-to-market; they will bolt AI on. His advice is to aim between two big players and work a new way, as HTTP and HTML displaced client-server without replicating it. Amble's version is functional: move from collecting data to acting on it through lead prioritization, churn flags and agentic outbound loops that generate benchmark data, or capture physical-world data in construction and manufacturing.

11. Network effects inside firms — ▶ 57:26

    Sonowski says network effects across companies face compliance and security obstacles, but the biggest enterprise network effect is inside a company. Most employees are not trying to streamline their jobs; a small motivated group, like the Goldman bankers using Excel, spreads tools to others. He sees chat doing this in 2025, and believes a white paper he generated for an SAP friend may have started such a loop in her team. Software that uses AI to connect functions that do not normally communicate, as Figma did for design and product development, could be a new category.

Enterprise software is not dying; its interface is losing importance. That is the narrower claim beneath Sema Amble's essay, and it is more defensible than the apocalyptic version. When agents read and write to systems of record directly, the screens built for human workflow become optional, and the data and logic below them carry the value. Salesforce's Headless 360 announcement is thin evidence for this. Amble thinks it was mostly marketing, a rebrand of APIs that already existed. The underlying behavior is better evidence. She cited a reported 300% rise in Slack agent usage, a figure she recalled reading somewhere, as a sign that people would rather have information delivered to them than log in to retrieve it.

But removing the interface does not turn a system into a commodity database, and Steven Sonowski spent much of the conversation explaining why. In his account, SAP is not a place where a company's data happens to live. It is the company, codified. Ford, Toyota and General Motors buy steel and aluminum from the same suppliers. What separates them is which screens managers choose to look at and which customizations they make. Expense reporting shows the mechanism: one person can handle receipts for 40 employees, but 100,000 people across 20 countries bring national laws and corporate policy that harden into process. Amble made the same point about the idea that Postgres plus APIs can replace SAP: the logic matters far more than where the data sits.

The same reasoning explains why vibe-coding a replacement is harder than it looks. Amble described a head of revops at a company of more than 1,000 people who had been asked to rebuild the firm's Salesforce instance internally and believed that importing fields and data was the job. The difficult parts, she said, are deciding what gets captured, how it maps to organizational lines, and who maintains it afterward. Stickiness, in Sonowski's telling, is similarly hard to engineer. It emerges when a customer threatens to leave and a rep discovers what actually holds them: HIPAA compliance for one stakeholder, onboarding for administrators, muscle memory for users, or Outlook's delegate access and shared calendars, which he said no one deliberately designed as a moat.

Editorially, this suggests two kinds of stickiness that agents treat very differently. Habit-based stickiness, built on the way humans click, read and write, is exposed when an agent does the clicking. Stickiness that comes from codified rules is not. Sonowski pointed to decades-old insurance software, and to Stripe, which he said encoded tax rules across jurisdictions, as systems that captured an external force such as regulation. Amble noted that ERP and payroll carry legal and audit requirements that make loose tracking untenable. The distinction between the two kinds is the editor's inference, not a claim either speaker made in those terms, but it fits their examples. It implies that incumbents' exposure may lie more in user-facing workflow than in their cores.

If the logic is what matters, the hard problem for agents is the logic that was never written down. Amble calls this context: exceptions, edge cases, permissions and policies that live outside CRM fields, such as responding one way to a customer in Asia and another in the US. Sonowski argued that almost everything interesting in an enterprise is an exception, and that most people cannot articulate the process they use to resolve a customer issue. Capturing it is a slow, observational task. Amble pointed to voice agents on freight compliance calls, used by one of her portfolio companies, and to computer-using agents watching how humans work. But exceptions are infrequent and sales cycles long, so a vendor must observe enough interactions, and buyers must trust that coverage, before an agent can act.

Sonowski added a taxonomy that clarifies how much agents can do today. Lookup is lightweight, and many new headless APIs are simply a more forgiving front end to old queries. Acting on a system of record raises enterprise questions about impersonation, credentials and whether the agent consumes a paid seat. Analysis, which spans multiple systems, suits agents because it is not time-bounded and can iterate or route across models to compare answers. It is also where hallucination hurts most, since every step needs verification. Amble added permissioning, read versus write rights and interactions between agents, all solvable but slow. The practical lesson is that headless and agent are often conflated, and a claim about one may say little about the other.

Sonowski's strongest rebuttal to automation anxiety is that productivity creates new work. Amazon's customer service chatbot resolves wrong-item complaints in the customer's favor, and the resulting data feeds improvements to shipping, warehouses and product descriptions. The work did not vanish; it moved to analysis. In his words, "It just got longer in a different way." Expense reporting led to travel analysis, and contracts, he predicted, will grow longer and more sophisticated with AI. His evidence is analogy rather than measurement, and he conceded that the radiology example, where radiologists like AI yet a shortage persists, is correlation rather than causation. The argument is persuasive as a pattern, but nothing in the discussion shows that new work will exceed displaced work in any given function.

The conversation's sharpest strategic claim concerns incumbents. Sonowski argued that no software vendor will accept being abstracted into a back end, because reducing a product to storing SQL for expense reports is a decaying business. Customers also resist assembling workflows from many providers, since a system is only as stable as its least stable part. Amble offered Workday as a case in point: it has APIs, but extracting all its data cleanly is difficult, and the vendor has little incentive to become a dumb database. Hence her three paths. Build agents on the incumbent, which she expects to yield mixed results. Do it yourself, which she compares to open-heart surgery on a live patient. Or build alongside the existing system and gradually observe how the business operates.

That leaves startups two kinds of openings. Sonowski's is positional. Competing head-on forces a vendor to meet thousands of requirements before entering the door, while incumbents bolt AI onto existing products because they will not disturb their product line or go-to-market. Aim between two big players, he advised, and do it in a new way, as HTTP and HTML displaced client-server without replicating what it did. Amble's is functional: move from collecting data to acting on it, by prioritizing leads, flagging churn risk and running outbound loops that gather benchmark data, or by pulling in physical-world data from construction and manufacturing. Both see value in bridging functions, and Sonowski argued that the largest network effects sit inside companies, where chat is spreading as Excel once did among Goldman bankers.

The unresolved tension lies between these two halves. The speakers argue that incumbents' logic is nearly irreplaceable, that incumbents will absorb whatever ecosystems build around them, and that startups can still win. Reconciling those claims depends on who ends up owning context. Amble allows that AI startups might one day replace back-end systems of record by observing how a business operates. Sonowski says SAP will simply do what the ecosystem does, though not all of it well. Neither offers systematic evidence of which dynamic is prevailing, and the discussion rests largely on anecdote and portfolio observation. If exceptions can only be captured slowly, the race may turn on whether startups accumulate that context before incumbents bundle something good enough.