What is the right Cortex attach goal for 2027?

35-45% Cortex attach by end of FY27 — defined as the percentage of paying Snowflake customers running at least one Cortex feature (LLM Functions, Cortex Search, Cortex Analyst, Cortex Agents, or fine-tuning) in production, not just trial. The math: Snowflake exited FY26 disclosing roughly 4,000+ accounts using AI/ML features weekly against a customer base of ~11,000+ — call that ~35% "touch" attach.

The bar to clear in FY27 is converting that touch into *production* attach, which should land 35-45% if Cortex Agents lands and the consumption-pricing motion holds. Anything below 30% by FY27 close means Cortex pricing, packaging, or partner economics are broken — and that should fire the CRO.

Anything above 50% means Snowflake either cannibalized the partner-routing margin or is counting trial seats as attach (i.e., the metric itself is being gamed). *Disclosure: Snowflake has not published a single canonical "Cortex attach" definition; the 35-45% range assumes the logos-in-production framing, not the revenue-share or query-share variant.*

What Cortex Attach Actually Means

• Logos attach (the version Snowflake leans on publicly) — % of paying customers using at least one Cortex feature in a production workload. Easiest to inflate, easiest to communicate, the version most likely to anchor the FY27 narrative.
• Revenue attach — Cortex consumption credits as a % of total product revenue. Closer to truth, harder to game, but Snowflake has not broken this out as a standalone line on earnings calls — it gets folded into total consumption.
• Query attach — % of total queries that touch a Cortex function (LLM Functions, Cortex Search, embedding generation, Cortex Analyst). The cleanest engineering metric, the worst marketing metric.
• Workload attach — number of distinct production workloads per Cortex-using customer. This is the depth metric that separates real attach from one-script-in-a-notebook attach.
• Pick one and publish it. The CRO's worst outcome in FY27 is shipping three different attach numbers across three different earnings calls and letting the analyst community pick the one that hurts the most.

What Comparable AI Attach Rates Look Like

• Salesforce Einstein / Agentforce — Marc Benioff has cited Agentforce closing thousands of paid deals since launch against a customer base of 150,000+, implying low-single-digit to low-double-digit paid attach in FY26, with attach acceleration as the lead investor narrative for FY27.
• ServiceNow Now Assist — Bill McDermott has repeatedly framed Now Assist as the fastest-ramping product in ServiceNow history, with Now Assist deals appearing in a meaningful share of new and renewal contracts; public commentary suggests roughly 20-30% deal attach on large renewals, not full installed-base attach.
• HubSpot Breeze — HubSpot has positioned Breeze as embedded across the suite (Breeze Copilot, Breeze Agents, Breeze Intelligence) rather than as a separately metered SKU; effective attach is high by design but revenue attach is intentionally muddied.
• Microsoft Copilot for M365 — analyst estimates and Microsoft's own commentary point to roughly 5-15% paid Copilot seat attach against the M365 commercial base through 2025, climbing as Copilot Chat and the consumption-credit motion expand the funnel.
• Adobe Firefly / GenStudio — Adobe has cited billions of generations and broad Creative Cloud penetration but has been deliberately careful not to publish a clean Firefly seat-attach number, again because the metric is definitionally fuzzy when generative features are bundled into existing seats.
• Pattern across all five: when AI is bundled, attach looks great and revenue is opaque; when AI is metered (Cortex's path), attach looks lower but revenue is honest.

Why Snowflake's Attach Should Be Higher Than The Comp Set

• Data gravity — the customer's training data, RAG corpus, and fine-tuning ground truth already live in Snowflake. The integration tax that suppresses Copilot and Einstein attach ("connect your data first") is pre-paid for Cortex.
• Consumption add, not seat add — Cortex sells against an existing credit balance instead of requiring a new per-seat PO. Procurement friction is the #1 attach blocker for Copilot and Einstein; Snowflake skips most of it.
• Native deployment — Cortex runs inside the customer's existing Snowflake account, governance model, and network perimeter. No new vendor security review, no new DPA, no new SSO integration.
• SQL-native LLM Functions — SNOWFLAKE.CORTEX.COMPLETE() lowered the build barrier from "hire an ML team" to "any analyst who can write SQL." Attach scales with the analyst headcount, not the ML headcount.
• Named customer references already exist — Bayer, Sigma Computing, CHG Healthcare, Sun Life, Cybersyn, and Siemens Energy have all been cited publicly using Cortex features. This is not a cold-start motion; the proof points are in market.

What Could Block 35-45% By FY27

• Cortex pricing not landing — if per-token credit pricing stays above what customers pay Bedrock or direct Anthropic/OpenAI APIs, the rational customer routes around Cortex even when the data lives in Snowflake.
• Partner-model passthrough margin compression — the Anthropic and Mistral integrations are partner-routed; if those partners raise wholesale pricing or open direct enterprise channels, Cortex's margin-and-attach flywheel weakens together.
• Anthropic and OpenAI direct-to-enterprise competition — both are now closing eight-figure deals directly with the same Fortune 500 buyers Snowflake sells to. "Just call Anthropic" is now a real procurement option in a way it wasn't in FY25.
• AI-skeptical regulated verticals — banking, insurance, healthcare, and federal customers throttle Cortex attach not on price but on model-risk-management review cycles. These cohorts may stall at 15-25% attach through FY27 regardless of pricing.
• "Cortex Lite" discount cannibalization — if Snowflake ships a free or near-free Cortex tier to chase logos attach, it inflates the headline number while crushing revenue attach and training the customer base that AI features are commodity bundle.
• Cortex Agents launch slippage — the agent layer is the multiplier on attach because it converts "used Cortex once" into "runs Cortex on a schedule." A six-month Agents slip is a five-point attach miss.

The Goal-Setting Math By Cohort

• Top-100 customers (the 8-figure accounts) — target 90%+ Cortex attach by FY27 close. These accounts have dedicated SEs, executive sponsorship, and the consumption headroom to absorb Cortex without a new PO. Reference points: Bayer, Siemens Energy, Sun Life, CHG Healthcare have all been cited as Cortex users; the Top-100 cohort should be near-saturated.
• Mid-Market (next ~1,500 accounts, $250K-$5M ACV) — target 40-50% attach. This cohort has the data volume to justify Cortex but lacks the dedicated AI team; attach depends on Cortex Analyst and Cortex Search lowering the build barrier enough for a single data engineer to ship.
• Commercial (the long tail of <$250K ACV) — target 20-30% attach. This cohort is price-sensitive, build-capacity-constrained, and most likely to substitute a free ChatGPT seat for a metered Cortex call. Attach here is a pricing-and-packaging problem, not a product problem.
• Public Sector / regulated — target 25-35%, gated by FedRAMP-High and HIPAA-aware Cortex variants shipping on schedule.
• Net effect — weighted-average lands in the 35-45% band if Mid-Market clears 40% and Commercial clears 20%. If either tier misses, the headline number drops to 25-30% and the Street narrative breaks.

Cortex Attach Targets By Cohort

How The Attach Goal Drives Outcomes

FAQ

What is the recommended Cortex attach goal for FY27? The target is 35-45% Cortex attach by end of FY27, defined as the percentage of paying Snowflake customers running at least one Cortex feature in production, not just trial. Below 30% means pricing, packaging, or partner economics are broken; above 50% means Snowflake either cannibalized partner-routing margin or is gaming the metric by counting trial seats.

Where does Snowflake's current attach stand? Snowflake exited FY26 disclosing roughly 4,000+ accounts using AI/ML features weekly against a base of ~11,000+ customers, which is about 35% "touch" attach. The FY27 bar is converting that touch into production attach.

What are the four different ways to measure Cortex attach? The four definitions are logos attach (% of paying customers using a Cortex feature, the easiest to inflate), revenue attach (Cortex credits as % of total product revenue, closer to truth but not broken out), query attach (% of queries touching a Cortex function, the cleanest engineering metric), and workload attach (distinct production workloads per customer, the depth metric).

The advice is to pick one and publish it.

How do comparable AI attach rates look at peers? Salesforce Agentforce implies low-single to low-double-digit paid attach against 150,000+ customers, ServiceNow Now Assist runs roughly 20-30% deal attach on large renewals, and Microsoft Copilot for M365 sits around 5-15% paid seat attach.

The pattern across all five is that bundled AI looks great on attach but opaque on revenue, while metered AI like Cortex looks lower but is honest.

Why should Snowflake's attach be higher than the comp set? Data gravity means the training data, RAG corpus, and fine-tuning ground truth already live in Snowflake, pre-paying the integration tax that suppresses Copilot and Einstein. Cortex also sells against an existing credit balance rather than a new per-seat PO, runs natively inside the customer's account, and is SQL-native via SNOWFLAKE.CORTEX.COMPLETE() so any analyst can use it.

Bottom Line

Set the FY27 number at 40% with a 35-45% public range, publish the definition once, and never restate it. The temptation will be to push 50%+ to win the earnings call; resist it. A clean, honestly-defined 40% logos-in-production attach with rising revenue attach underneath is a better five-year story than a 55% headline that gets unwound by the first analyst who asks how trial attach is being counted.

Snowflake's structural advantage — data already in the warehouse, consumption credit already on the PO, SQL-native LLM Functions — should clear the comp set, but only if Cortex Agents ships on schedule and Lite-tier discounting doesn't poison the revenue mix. *(see also: q1564, q1566, q1600)*