The interface is disappearing.
The record is not.
At the September 15 close, Salesforce clears the proposed three-year return hurdle and bear-downside rail in the first explicit model. Build the complete Call packet; do not promote this working paper into a position.
Ben Thompson argues that agents will make enterprise software interfaces less defensible. HOUSE agrees with the mechanism but not with the shortcut. The investable question is which system of record still controls governed data, permissions, and action when the agent owns the screen.
Thompson supplies the product observation: “Salesforce should be headless.” He does not recommend CRM, specify an entry, set a horizon, or define a loss condition. Every security choice and number below is HOUSE inference.
INVESTMENT CONCLUSION FIRST
CRM clears. The basket does not.
Agents can bypass interface friction. UI familiarity alone is becoming a weaker moat.
Systems of record may retain power through governed data, permissions, identity, and write access.
The modeled base case clears a proposed 13% annualized hurdle without requiring a heroic exit multiple.
No live HOUSE position, sealed basis, or AI panel exists. Advance one name, not a correlated software basket.
From a future eligible close, CRM should outperform IGV held unchanged over 36 months if Salesforce converts its system-of-record position into governed agent actions faster than agent interfaces erode seat pricing.
HOW THE THESIS MAKES MONEY
Lose the face. Keep the authority.
Users ask a model to complete work instead of navigating each vendor's interface.
Customer, workflow, entitlement, and transaction data still require an authoritative home.
The valuable layer authenticates the agent, limits its actions, records its work, and reverses mistakes.
Winners monetize governed reads, writes, workflows, and data movement faster than UI revenue decays.
A correct product observation can still produce the wrong stock.
- Agents must act, not merely summarize. Copilots that stop at answers do not create a new transaction layer.
- Customers must keep the record vendor. Open formats, warehouses, and agent memory cannot make the application database interchangeable.
- Control must remain monetizable. API and action revenue must replace lost seat value instead of becoming a free compatibility feature.
- Growth must reach each share. Acquisition spending, stock compensation, infrastructure cost, and buybacks decide the equity result.
MAP THE SECURITIES
Four names. Three different trades.
CRM, NOW, and SNOW share enterprise-AI exposure; owning all three does not diversify this thesis. ORCL's present return mechanism is even further away. HOUSE ranks the claims instead of hiding them in a basket.
Sales, service, customer data, permissions, and workflow can remain authoritative even when another agent owns the interface.
WHAT MUST PROVE · AGENTFORCE + DATA 360 GROW PER-SHARE CASH FASTER THAN SEAT VALUE DECAYSWorkflow authority and Action Fabric may make ServiceNow a cleaner cross-agent execution layer.
WHAT MUST PROVE · THE SUPERIOR GROWTH SURVIVES A FULL VALUATION AND DILUTION MODELGoverned access across application silos is strategically useful, but adjusted cash flow cannot erase dilution.
WHAT MUST PROVE · PER-SHARE ECONOMICS CATCH THE PRODUCT-REVENUE STORYDatabase gravity is relevant, but the current earnings debate is AI infrastructure, backlog, and capital intensity.
GRADE IT AS AI INFRASTRUCTURE, NOT AS HEADLESS-SAAS DIVERSIFICATIONPRICE IS PART OF THE THESIS
CRM earns the next round.
The model uses reported FY26 free cash flow, the midpoint of FY27 cash-flow growth guidance, guided diluted shares, explicit per-share growth, and an exit free-cash-flow multiple. It is intentionally simple enough to attack.
PRICE BASIS. CRM $255.65, NOW $141.90, SNOW $322.98, ORCL $140.35, and IGV $105.55: September 15, 2026 daily closes from Yahoo Finance via the VPS, queried September 16. NOT LIVE.
CRM INPUTS. FY26 free cash flow · FY27 guidance and diluted shares · Q2 Form 10-Q.
BENCHMARK. IGV tracks North American software and selected interactive-media companies. The 10% base assumption and +3-point hurdle are HOUSE proposals, not iShares forecasts.
THE OPPOSITION CASE
Headless can be right and CRM can still lose.
The model owns the customer.
Third-party agents reduce Salesforce to a replaceable database and pressure API access toward zero-margin compatibility.
Usage does not replace seats.
Premium seats and UI bundles contract before Agentforce consumption becomes large or durable enough to compensate.
The record moves.
Warehouses, open tables, MCP servers, and model memory make enterprise context less dependent on the application vendor.
The company wins. Owners do not.
Acquisitions, stock compensation, infrastructure cost, and buybacks consume the operating benefit while IGV compounds elsewhere.
REALITY GRADES
Usage is real. Capture is unresolved.
Agentforce plus Data 360 ARR, up more than 210% year over year.
OFFICIAL RESULTS -> AGENTIC WORK7.0BCumulative Agentic Work Units; 3.2B in Q2, up 97% from the prior quarter.
ACTIVITY, NOT PROFIT -> SERVICENOW · Q2 2026+24.5%Subscription revenue growth, with AI crossing $1B of annual contract value.
SECOND SCREEN -> SNOWFLAKE · Q2 FY27126%Net revenue retention, beside 37% product-revenue growth and 30% RPO growth.
DATA-LAYER EVIDENCE -> SNOWFLAKE · Q2 FY2729%Stock-based-compensation-related charges as a share of revenue in the quarter.
PER-SHARE WARNING -> ORACLE · Q1 FY27$664BRemaining performance obligations, with cloud infrastructure revenue up 121%.
DIFFERENT RETURN ENGINE ->These facts support rapid agent adoption, governed-data demand, and enterprise willingness to spend. They do not yet prove that record owners capture more economics than model vendors, warehouses, integrators, or customers.
RTI · REAL-TIME INTELLIGENCE
Watch the value migrate.
Listen for headless-platform pricing, Agentforce consumption, Data 360 attach, and explicit evidence that usage replaces rather than cannibalizes seats.
Track organic cRPO and subscription growth excluding Informatica, Agentforce ARR, cash conversion, stock compensation, and diluted shares.
Which layer records agent actions, controls permissions, and charges for them? Product announcements matter only when revenue and per-share cash follow.
New free tiers, usage units, API charges, bundled access, and customer migration reveal where bargaining power is moving.
RTI means real-time intelligence: recurring public evidence checked against the same written mechanism. It is not a claim that market data or the page updates tick by tick.
WHAT WOULD CHANGE OUR MIND
Make the thesis mortal.
Dates, series, and exact thresholds must freeze with the future Call packet. These are underwriting proposals, not a graded record.
- ORGANIC GROWTH FAILS. By the FY28 year-end report, Salesforce's organic subscription and support growth has not reaccelerated despite material Agentforce and Data 360 usage.
- CASH CAPTURE FAILS. Three-year free cash flow per share grows below 7% annualized from the future lock basis.
- HEADLESS PRICING FAILS. Governed agent reads, writes, or actions remain bundled or free and do not replace measured seat pressure by FY28.
- THE RECORD MOVES. Disclosed customer migrations show warehouses, open data layers, or model platforms replacing Salesforce as the authoritative customer record.
- PER-SHARE DISCIPLINE FAILS. Acquisitions and stock compensation offset operating cash growth, leaving diluted per-share economics below the sealed path.
- THE TRADE LOSES. At 36 months, CRM fails to outperform IGV held unchanged. The money verdict stands even if headless architecture wins.
WHAT REMAINS BEFORE CAPITAL
Turn the screen into a record.
Separate Informatica, model seat pressure, action consumption, infrastructure cost, stock compensation, buybacks, and acquisition spending.
Normalize free cash flow, dilution, and valuation instead of awarding the trade to the fastest current growth rate.
Same close, no rebalance, corporate actions, benchmark treatment, and exact kill-condition settlement.
Do not ask frontier models to vote on an article. Ask them to challenge the security, entry, cases, and falsifiers.
Never backfill September 15 as a basis. If price or facts change first, rerun the model.
The source observation, HOUSE inference, machine attack, and future evidence remain separate on the permanent record.