Build the tables once, wire them, let them run. Native Clay tables are the engine, tables writing into tables, conditional runs, scheduled refreshes, so the account list stays alive instead of dying the moment the market moves. Depth is the point: Sources → Accounts → Contacts → Message → Approval → Sync → Reply → Attribution → Worklist is the full chain of real wiring, and every hop is gated so credits only spend on rows that are in-window, in-ICP, and not already filled. Measurement before volume, ICP before sourcing, human before send.
Two plain-English one-pagers go alongside this for a first read: The Clay Engine Plan and The Golden List. Those explain the motion in exec terms and group the flow into eight stations; this page is the detailed, interactive version with all nineteen tables broken out. Same engine, same numbers.
Operating model
The most common question is "CSV every time, or the MCP?" Neither is the primary way of working. Here is how the three pieces divide up.
Native Clay tables
The living, auto-updating motion. Tables write into tables on conditional runs and scheduled refreshes. This is where the work happens day to day, and the only place the layered architecture actually runs.
MCP in Claude
Ad-hoc research, one-off pulls, feeding data into dashboards and briefs. It finds, enriches, and reads. It does not build or run the layered architecture. Use it to answer a question fast, not to run the motion.
CSV import
A one-time seed from an external file. After the seed, the native table owns the data and keeps it fresh. Re-exporting CSVs to work from produces snapshots that die when the market moves, the exact thing the engine exists to beat.
The nine-layer stack
Every layer is a real set of Clay tables wired to the next by a Send Table Data action, a Lookup column, or a conditional run. The hands-on version is at the bottom of the page. Here is the map.
CMS import · install-base import · signal feeds
Land the CMS Star Ratings file and the Salesforce install base. A conditional filter pushes only in-window, non-customer rows forward into Accounts.
The Golden List hub
Scored, enriched, graded, signal-aware. Firmographic and technographic waterfall, a fit score, an intent score, a grade, and the top signal. Enrichment runs only-if-empty; stale rows refresh on a schedule.
Star Ratings + back office, writing back into Accounts
Per-account intent monitoring. Tier-1 triggers fire outreach; Tier-2 raise priority. Signals write intent and the why-now line back onto the account.
Buying committee
People linked to accounts, ICP-filtered, persona-matched, carrying the read-only attribution tags. Email waterfall runs only-if-empty; invalid emails reject before they cost a send.
Message Gen, reading the reference tables
Personalized variant copy per contact, keyed to the account why-now and product angle. It looks up approved stats, the product-angle map, and the ICP rubric so claims stay inside what is verified.
Send Queue: critic gate + human approval
The single egress. Nothing sends from Clay directly. A critic checks email validity, in-scope motion, verified claims, and length; a human approves before anything leaves.
Outreach Sync + deliverability gate
Approved copy and verified email hand off to the sequencer, and source-tags write to Salesforce. The sender webhook stays off until domain health is green.
Reply sync + attribution
Replies and meetings write back onto contacts, then attribution computes the two-motion funnel and engine-sourced credit at the qualified-reply line.
AE territory + worklist
The human-facing end. Each rep opens a daily worklist: graded, signal-fresh accounts and their committee, with the why-now line and the suggested product angle, filtered to their territory.
Every table, and what it does
The nine layers are made of nineteen real Clay tables. This is each one: what it is, what feeds it, what it writes to, and when it runs. The five marked Core five are the first build and prove the whole motion on a slice; the reference tables are small and static; the rest layer on after. Filter by layer to focus.
Day-one build order
Do not skip. Prove each stage on a ten-row Tier A+B slice before widening. Status reflects what is designed and data-ready versus what still stands up inside the Intradiem Clay account. The head start is the design and the data; standing the tables up is the near-term work.
Confirm Salesforce attribution fieldsDesigned · data ready
gtm_engine_sourced, source_motion, sourced_date. If missing, flag to Naveen as a build so credit is provable from day one.
Seed the universe and the dedup baselineDesigned · data ready
Load the 307-row canonical universe (Tier A+B first) into CMS Import; load the 157 existing contacts into the dedup path. Intersect against the install base so current customers route to the back-office motion, not new-logo.
Accounts identity, then the enrichment waterfallDesigned · data ready
Add the refreshed data model (cs_star, clin_star, the addressable-forgone columns, the time-phased 2028/2029 columns, tier, new_faller). Enrichment runs only-if-empty on conditional runs so credits are protected.
Scoring: fit, exclusions, gradeDesigned · data ready
Rank on addr_2028_musd, the influenceable window. Add motion_exclude so clinical-only gaps never source. Run on a ten-row slice (Humana H5216, H5619, H0028) and confirm grade computes and excluded rows drop. Weights and bands are tunable in config and pending ratification.
Contacts: persona pull + attribution lockDesigned · data ready
Run the persona pull from the runbook: ICP filters, dedup against the 157, caps, and the persona_key column. Verify the three attribution tags lock and never overwrite.
Message Gen: paste v2.1, generate two rowsNext to stand up
Wire the v2.1 system prompt (window beat, payment-year clause, time-stamped UHC line, addr_2028 lead for Persona 2). Generate one Persona 1 and one Persona 2 row and eyeball against the worked examples before going wider.
Send Queue + three webhooksNext to stand up
Point webhook #1 at the approval queue. Leave webhook #3 (sender) off until deliverability is green.
Align with Naveen and ScottGated
Before lifting the test-slice cap we align on the business calls: the ranking basis (addr_2028 as the influenceable window), how we credit the engine, the back-office ICP with Scott, and the call-center rule-text check. The Clay weights are mine to tune and show. This gate is a decision, not a build.
Turn on Tier-1 signals, then first real runGated
Schedule the October CMS re-import and the earnings sweep. First live run is three variants on a Tier A+B segment, small, human-approved, deliverability watched. Humana leads, then UnitedHealth, CVS/Aetna, Centene, Elevance, HCSC.
Gates that stay closed until they pass
The engine is built to fail closed. A stage error stops the run; it never silently sends. Three gates stand between a built table and a live motion.
Attribution ratified
Scoring weights, the addressability ranking basis, and the credit definition agreed with Naveen and Scott before the test-slice cap lifts.
Baseline set
A measured starting point so surfaced opportunity and realized results never blend into one number.
Deliverability green
Warmup, inbox rotation, and bounce rate healthy. The sender webhook stays off until this reads green. This is the tripwire that stops spray-and-burn.
Human before send
Every message clears the critic checks and a human approval before it can leave Clay. No exceptions, no direct send.
Score an account, watch the engine decide
This is the part worth touching. Load a real account or move the inputs yourself, and the engine computes its fit score, checks the exclusion rules, picks the product angle, and drafts the why-now line, live, using the actual formula from the build. Then run it through the pipeline and watch it either flow to a rep or get stopped at the gate with zero credits spent.
The decision logic, in your hands
Every number here is computed from the inputs by the same logic these Clay tables are built to run. The weights and grade bands are mine to tune and prove on live accounts.
◇ Weights & bands are my dials · set on evidence, shown as they prove outEach layer is a live set of Clay tables
Select one to see its job and the exact Clay mechanic that carries the row to the next layer, or run the account above to watch the flow.
At segment scale: how the wired engine lands more deals
Illustrative model · not a forecastOne Tier A+B segment, roughly 98 accounts and their committees. The point is not the counts, it is that every drop-off becomes a gate you can see and tune, instead of a mystery. Run an account above to fill it.
*The $2.5B and the segment figures are Dallas's own analysis of the 2026 CMS Star Ratings universe (addressable forgone Quality Bonus in the influenceable 2028 window), used to size and rank the motion. They are not Intradiem customer results or company-verified ROI. The funnel is an illustrative throughput model for shaping the conversation, not a promised outcome or a target.
◇ Questions for tomorrow
The engine and the messaging are built and ready to stand up in our workspace. A handful of calls sit with you and Scott rather than with the build, and I'd rather shape them together than assume. Here is where your read moves this forward:
- What counts as a winnable account in Intradiem's terms. My recommendation is to rank on the influenceable 2028 window (addr_2028), and I will show you why on live accounts.
- How we credit the engine, tagging at the qualified-reply line, so surfaced and realized never blur.
- The back-office ICP, Scott's domain to define, mine to source against.
- The call-center rule-text check for the 2028 measures, a joint verification so the copy never inherits a claim that has shifted.