“A tool stack is not an intelligence stack. The difference is whether the tools talk to each other or just produce dashboards in parallel.”
Most Amazon agencies have a tool stack. Helium 10. SmartScout. Jungle Scout. Intentwise. Sometimes Pacvue or Stackline at the higher end. Each tool produces its own dashboard. Each has its own login. Each covers a slice of the brand’s operating picture.
The agency’s strategists open three or four tabs every morning. They look at the dashboards. They try to hold the cross-references in their heads. That is a tool stack. It is not an agency intelligence stack.
The difference is whether the tools talk to each other. Whether the keyword movement Cerebro is showing this week is being cross-referenced against the search-term cost shift Adtomic is showing, the inventory-cover compression SmartScout’s catalog gap report is showing, and the brand-share trajectory Intentwise is pulling from the back-end Amazon feed.
A tool stack leaves that cross-referencing in the strategist’s head. An agency intelligence stack does the cross-referencing in code. It surfaces only the situations the strategist needs to make a call on. We built one. Here is what is inside the agency intelligence stack we run.
The shape, five layers from raw data to action
The ClearSight agency intelligence stack has five layers. Each downstream of the one before:
- Parsers, connectors that ingest each tool’s native export format and normalize it into a single schema
- Rule engine, deterministic logic that flags anomalies and threshold breaches across the normalized data
- AI compile layer, Claude-powered synthesis of the rule engine’s output into human-readable briefs
- Strategist edit layer, the human who reviews, edits, and ships the brief’s recommendations
- Action layer, the integration that pushes shipped recommendations to ClickUp tasks and Slack alerts for the operators who do the actual Seller Central work
Each layer is downstream of the one before. AI does not generate without rule engine input. Rule engine does not run without normalized data. Normalized data does not exist without parser output. The architecture forces the dependency chain. No recommendation reaches a brand without passing through a deterministic check and a human edit.
Layer 1, the parsers
The parser layer is the unglamorous foundation of the whole agency intelligence stack. Six parsers run in production today:
- helium10.py, Cerebro keyword pulls, Magnet expansions, search-term reports, Index Checker output, inventory protector levels
- smart_scout.py, Brand Score history, catalog-gap reports, ad-spy data, AMS reports, organic-rank trajectories
- jungle_scout.py, Keyword Scout cross-verification, niche-hunter data, opportunity finder for product-line expansion
- intentwise.py, campaign performance, search-term cost analytics, share-of-voice trajectories, the daily account snapshot
- advertising.py, Amazon’s native Sponsored Products / Sponsored Brands / Sponsored Display reports
- business_report.py, Amazon’s native Business Reports (session counts, conversion rates, units ordered per page-view)
Each parser writes into a single normalized schema. With stable column names. Regardless of which tool the raw export came from. A “keyword” is a keyword whether it came from Cerebro, Keyword Scout, or SmartScout’s organic rank report. An “ASIN” is an ASIN whether the source was Helium 10 or Intentwise.
This normalization is the precondition for any cross-tool analysis. Without it, the strategist is squinting at four different column names for the same concept across four different tabs.
The brand never sees the parser layer. They see the consequences of it. Recommendations that reference keyword movement and ad-spend shift and inventory cover and conversion rate. All in the same sentence. Because all four data sources have been joined upstream.
Layer 2, the rule engine
The rule engine is where the deterministic operating intelligence lives in the agency intelligence stack. It is split across five domain modules. Each owns a specific class of decision:
- bid_rules, threshold logic on placement modifiers, bid floors and ceilings, dayparting, and search-term-level ACOS-based bid actions
- campaign_rules, structural logic on campaign tree health, single-keyword-vs-thematic segmentation, branded-versus-non-branded defensive bidding
- inventory_rules, weeks-of-cover thresholds on top-revenue SKUs, IPI score tracking, replenishment-PO-in-flight reconciliation
- keyword_rules, index health checks, organic-rank-movement thresholds, share-of-voice drift detection, negative-keyword harvest triggers
- listing_rules, attribute-completeness thresholds, A+ content freshness, parent-child variation health, Cosmo-attribute alignment checks
Every rule is deterministic. Every rule has a threshold the strategist team has agreed on. Every rule emits a flag with a severity (critical, high, medium, info) and a structured payload describing the situation.
Rules are not opinions. They are the codified version of what the strategist team has already decided constitutes a situation worth acting on.
This is the layer the ai-only Amazon agency model does not have. Without a rule engine, the strategist has to remember the thresholds, apply them in their head, miss them when distracted, and reconcile across tools manually. The rule engine is the part of the agency intelligence stack that scales. Once the rule is written, it runs on every account, every day, without drift.
Layer 3, the AI compile layer
The AI compile layer is where Claude sits in the agency intelligence stack. It does exactly one job. Take the rule engine’s flag output for an account on a given day. Cross-reference it against the account’s historical context. Last quarter’s TACoS trajectory. Last month’s ranking movement. Last week’s ad-spend cadence. Produce a single synthesis brief in plain English.
The brief has a consistent shape:
- The two or three flags that actually matter this week, ranked by dollar exposure
- The root cause hypothesis, with the data points that support it
- The recommended action, with the expected outcome and the dollar value if it works
- The flags the rule engine surfaced that are not worth acting on this week, and why
The last bullet is the part that proves AI is doing useful synthesis. A rule engine surfaces every flag it generates. Most flags, on most days, are not the most important thing. Filtering the noise, deciding which flag deserves the strategist’s attention and which can wait, is a task LLMs handle well when given the historical context.
Without the AI layer, the strategist drowns in the rule engine’s output. With it, the strategist sees the two or three signals that matter and a one-paragraph defense of why those are the right two or three.
The AI compile layer never ships to the brand. Its only customer is the strategist editing the brief in the next layer.
Layer 4, the strategist edit layer
Every synthesis brief lands in front of a human strategist before any recommendation reaches a brand. The strategist’s job at this layer is to:
- Verify the AI’s root cause hypothesis against context the AI does not have (brand owner’s preferences, ongoing initiatives, recent conversations)
- Tighten the recommended action to the brand’s actual operating constraints (budget caps, vendor lead times, scheduled campaign launches)
- Reject recommendations that read clean to the AI but are wrong in context (the most common reason for rejection: timing, the action is correct but the wrong week to ship it)
- Add the relationship context, what to say to the brand, in what tone, with what framing
Our internal metric is that strategists edit approximately 70% of AI-generated synthesis briefs before they ship. That edit rate is the proof point. The AI is doing useful synthesis without owning the decision.
The synthesis is structurally correct enough to be worth keeping. The framing and the timing are wrong often enough that human judgment is the binding constraint on what ships in the agency intelligence stack.
Layer 5, the action layer
Shipped recommendations route to two destinations:
- ClickUp tasks for the operators (catalog, ads, inventory) who do the actual Seller Central work, with the recommendation, the data context, and the success criteria attached
- Slack alerts to the brand’s dedicated channel for time-sensitive situations (Buy Box loss, suppression event, stockout signal, suspicious campaign spike) where the brand owner needs to know in real time
The brand sees the Slack alerts and the weekly summary the strategist composes. The brand does not see the layers underneath. That is the right shape. The intelligence layer is plumbing. Plumbing should be invisible when it is working.
What runs daily, weekly, monthly
The agency intelligence stack runs on three cadences:
- Daily, parser ingests, rule engine runs, critical and high-severity flags route to Slack within an hour of detection
- Weekly, full synthesis brief per account, strategist edit, brand-facing summary, ClickUp tasks for the operator team
- Monthly, trend analysis across the trailing 90 days, strategic call recommendations, quarterly planning input
This cadence is what the weekly Amazon agency report we wrote about earlier is designed to land in. The daily cadence catches the events that cannot wait. The weekly cadence is the operating report. The monthly cadence is where strategy gets reviewed against execution evidence.
Why this shape is the white-hat shape
Every black-hat archetype we have written about, the three shapes frame post and the deep dives on offshore VA shops, AI listing generators, and AI-first consultants, is missing one or more layers of this agency intelligence stack.
The offshore VA shop is missing the rule engine and the AI compile layer. The listing-generator SaaS is missing the strategist edit layer. The LLM ships to Seller Central directly. The AI-first consultant is missing the action layer. The deck never reaches Seller Central at all.
The agency intelligence stack is the white-hat answer because it forces every recommendation to pass through deterministic checks AND human judgment AND an execution path.
Removing any one of those three lets the agency cut cost and reach more brands faster. Each removal also produces a predictable failure mode that compounds on the brand quarter after quarter.
The next four posts in this series walk through how this stack runs in practice. One post each on catalog work, ads, creative and A+ content, and inventory forecasting. Each post shows where the rule engine fires, where AI compiles, and where the human still owns the decision.
Reviewed by the Catalog Team and the Amazon Growth Team.
The Hybrid Stack, 10-post series. You are reading post 5 of 10.
Black-hat track, three archetypes of the ai-only Amazon agency:
- The three black-hat shapes of the 2026 Amazon agency
- Why offshore VA + ChatGPT shops are the most expensive cheap option
- Why AI listing-builder SaaS can’t get a 7-figure brand to actually index
- Why AI-first consultants who never log into Seller Central miss the work that moves money
White-hat track, the ClearSight hybrid stack:
- The ClearSight intelligence layer, stack reference (you are here)
- Catalog AI/human workflow
- Ads AI/human workflow
- Creative AI/human workflow
- Inventory AI/human workflow
- 12 months of the hybrid stack, results recap
← Previous: Why AI-first consultants who never log into Seller Central miss the work that moves money | Next: Catalog AI/human workflow →