Salesforce sold an ontology. The screens were the mechanism that forced sales teams to comply with it. Accounts, Opportunities, Stages, Forecasts, Territories, Approvals: entire go-to-market organizations reorganized around those objects, and the UI existed to make sure every rep typed structured data into the right field. AI agents weaken the interface advantage. They leave the organizational standardization underneath it largely untouched.
That distinction sits at the center of Mukund Mohan's response to the "Postgres plus APIs plus agents replaces SaaS" argument now circulating through venture and enterprise circles. His verdict: LLMs won't kill Salesforce, SAP, or most other systems of record. They will change what those companies are, how they charge, and who sits on top of them.
The timing matters. Vendors are rebranding as agent platforms, seat counts at large software buyers are under scrutiny, and a generation of AI-native startups is pitching the replacement of CRM and ERP with a database and a model. Whether incumbents survive that pitch depends on where their value actually lived, and Mohan's argument is that most observers have been looking at the wrong layer.
Headless is a shift in who uses the software, and the APIs were already there
Most enterprise software exposed APIs years ago. Calling Salesforce "headless" today overstates the technical novelty, and Mohan thinks the whole industry is leaning too hard on the word.
What has changed is the primary consumer. Historically that was a human logging in. Now, more often, it's an agent calling an endpoint. Mohan frames this as a distribution and interaction shift rather than an infrastructure shift, and the framing changes the story Salesforce gets to tell about itself. The company is repositioning from an application employees log into toward an operational substrate that agents execute against.
That's a very different narrative than "Salesforce went headless," and it's one an incumbent can survive.
The policy layer is harder to copy than the database
Here's where the Postgres argument loses ground. The database is the easy part. A large share of SaaS value lives in encoded institutional policy, which no CRUD layer captures:
- who can approve what
- escalation trees
- exception handling
- rollback rules
- auditability
- compliance workflows
- reconciliation logic
- permissions
- temporal sequencing
- coordination between departments
Mohan calls this the operational state machine, and it's the part no schema generator produces.
AI lowers the cost of recreating roughly the first 80 percent of a system of record: schema generation and workflow reconstruction. The remaining 20 percent is operational entropy, the accumulated edge cases of 15 years of real customers doing unexpected things with the system. That 20 percent is the moat, and it's where incumbents still hold the advantage.
The next moat is observing the action loop
The agentic stack may collapse the old distinction between system of record and system of execution. Under the old model, CRM stored intent, ERP stored transactions, ticketing stored requests, and humans did the work. In the new one, the system stores context, agents execute, outcomes feed back as new training and context, and the exhaust of execution becomes proprietary data.
That closed loop matters enormously. The defining question of the era becomes "who observes the action loop?", displacing "who stores the record?"
Consider a vertical AI-native field-service platform that dispatches technicians, observes outcomes, tracks timing, handles exceptions, coordinates payments, monitors resolution quality, and captures edge-case execution traces. It's building a stronger moat than a passive CRM because it generates new proprietary operational data every day. The transition Mohan describes runs from static data to dynamic execution exhaust.
DoorDash shows how far this goes. Its defensibility comes from coordinated operational execution: drivers, routing, logistics, dispatch, fulfillment timing, marketplace balancing, payments, exception recovery. Restaurant data on its own would be worth little. The durable enterprise platforms of the next decade probably look more like that kind of coordination network than like a dashboard.
Mohan singles out one line from the piece he's responding to as its most important: "The data is the context now." He agrees, then argues it understates the case. In agentic systems, context is memory, memory drives action, action generates exhaust, and exhaust becomes proprietary context. The loop compounds recursively. The strongest AI-native systems will observe, model, predict, execute, optimize, and benchmark workflows, which is a qualitatively different business from storing records.
Organizational sediment was never a network effect
What most enterprise SaaS built was switching-cost lock-in that looked like a network effect from a distance.
Another company adopting Salesforce added nothing to the value of your instance. Leaving got harder anyway, because integrations accumulated, process accumulated, training accumulated, reporting accumulated, and management rituals accumulated. Mohan's term for this is organizational sediment, and it's an honest description of what most "sticky" systems of record actually have.
Agents may change that for the first time, because agents transact across organizational boundaries. Procurement agents negotiating with supplier agents. Insurer agents settling with provider agents. Auditor agents reconciling against accounting agents, logistics agents coordinating with manufacturer agents. Once multiple counterparties run through the same execution rails, the software stops being a database and becomes market infrastructure.
The comparables shift accordingly: Stripe, Visa, Flexport, DoorDash, even Bloomberg, rather than traditional SaaS.
Seat pricing breaks before the software does
The technical transition may be the smaller threat. The commercial one is the seat.
If one agent replaces 20 UI users, workflows automate, and humans touch the system less often, then charging per employee seat becomes an artifact of a prior era. For large enterprise vendors this is existential, because their valuation multiples were built on predictable seat-expansion economics. Mohan considers this pricing transition potentially more dangerous to incumbents than the technology itself.
Pricing will likely migrate toward workflow volume, execution outcomes, API consumption, agent transactions, orchestration complexity, governance and compliance layers, and coordination rails. Every one of those is harder to forecast than headcount, and every one forces a vendor to prove value per action rather than per login.
Enterprises can't run the alternative yet
Mohan's one direct pushback on the replacement thesis concerns speed. "Postgres plus APIs plus agents" is compelling in a slide deck. In practice, most enterprises can't keep their Salesforce objects clean. The average company is nowhere near ready to safely operate custom ontologies, agent orchestration, policy engines, exception-handling systems, governance layers, evaluation pipelines, permission frameworks, and audit systems on its own. Outside elite tech companies, that capability barely exists.
So the near-term outcome is incumbents turning infrastructural rather than disappearing. Salesforce may become the policy layer, the permission layer, the audit layer, and the canonical customer graph, while AI-native systems own execution and workflow intelligence above it.
The startups that matter will probably parasitize the incumbents before they replace them. That's how platform transitions have historically played out.
What defensibility looks like now
Mohan accepts the core of the original thesis: defensibility is leaving UI habit and moving toward execution loops, operational context, policy orchestration, proprietary exhaust, multi-party coordination, real-world execution, and trust infrastructure.
The implication for anyone building or buying is that the label "SaaS" may soon describe the wrong category. The winners of the agentic era will look like operational networks: systems that coordinate action between parties and learn from every transaction. Selling seats to look at records was a different business.
