How Trellis Uses Claude and AMC to Analyze and Act

Most tools that put a language model on top of Amazon data stop at the observation. They name the problem fluently and cannot act on it. They can tell you your branded terms are eating budget that non-branded discovery should be earning. They cannot move the budget. The gap between the observation and the change is where the work really lives, and it is the gap most AI features quietly leave to you.

Trellis closes that gap in two pieces. Claude does the judgment a numeric rule cannot: is this search term branded or a competitor, what is the intent behind it, does this listing actually match the query it ranks for. It also helps you shape the standard itself, pressure-testing the SOP you hand it and recommending the checks you did not think to write down. Then a scheduled skill in Qore turns that judgment into a proposed change and holds it at an approval gate. Nothing moves a bid or a budget until you approve it. AMC, a paid add-on, widens the evidence beyond last-click when you need it.

The short version: Claude interprets, Qore proposes and executes under approval, AMC widens the evidence. This post walks the architecture, because the interesting part is not that a model can talk. It is where we let it act and where we deliberately do not.

Quick Answer

• Claude makes the judgment calls a threshold rule cannot: branded versus competitor terms, intent, whether a listing matches its query. It also helps you build the SOP, pressure-testing the standard you describe and recommending checks to add. It narrates its reasoning so you can audit the call.
• Qore is the workflow engine. You describe a workflow once, it assembles into visible logic, you lock it, and it runs on a schedule or across a roster of products and accounts. Actions are gated on approval, manual by default.
• AMC (Amazon Marketing Cloud) is a clean room for audience construction and path-level analysis. It is a paid add-on at the Trellis level, not bundled into Qore or Qinetix.
• The dividing line: analysis tolerates being wrong occasionally; continuous unattended action does not. That is why analysis runs freely and action waits behind a gate.

How It Works

You set the workflow up once, then it runs on its own until it wants to spend money.

Setup is a conversation. You describe the review you already do by hand, the terms you check, the thresholds that matter, what you do about them, and Claude helps you sharpen it into a workflow, suggesting the checks you left out. Qore turns that into logic you can read line by line, and you lock it. Locking earns its keep for one practical reason: a locked workflow returns the same answer from the same data every time, so this Monday's output matches last Friday's. A raw prompt to a language model will not.

From then on it runs on a schedule. It pulls the report, calls Claude for the calls a threshold cannot make, and assembles what it would change. The moment it reaches something that spends money, it stops and puts the proposal in front of you. You do the setup once; the reading and the proposing repeat on their own; the spending waits for your yes.

What Claude Does, and Its Precise Limit

Claude's job is the part of Amazon analysis that resists being written as a formula.

A numeric rule can tell you a keyword spent $180 and returned two orders. It cannot tell you that "yeti" is a competitor's brand and "insulated tumbler 30oz" is generic discovery demand, or that a query reading like a gift search should be weighted differently from a replacement-purchase search. Those are semantic judgments. Somebody has always made them by reading the terms and knowing the category. Claude makes them at the scale of the whole search-term report instead of the top forty rows a human has time for.

Four jobs, concretely:

• Classification. Sorting search terms into branded, competitor, and generic. Tagging queries by likely intent. Grouping products by how their listings really read rather than by catalog metadata.
• Semantic judgment. Deciding whether a listing's title, bullets, and A+ content match the product and the query it ranks for, and flagging the ones that do not.
• SOP shaping. Helping you turn the review in your head into an explicit standard, and recommending the checks you would want but did not spell out.
• Narrated reasoning. Writing down why it landed where it landed, so the output is a rationale you can read rather than a score you have to trust on faith.

Now the limit, stated plainly, because it is the load-bearing fact in this whole design.

A language model can narrate a rationale. It cannot guarantee it will produce the same rationale twice from the same input, and it cannot guarantee the narration it writes is the reason the decision got made. The explanation is generated alongside the answer, not read off the mechanism that produced it. This is not a knock on Claude specifically. It is how these models work.

Grounding helps, and the numbers are worth stating. An ungrounded model answering questions about a specific account lands somewhere near 30 percent reliability. Ground it properly through MCP, giving it structured, current account data instead of asking it to recall, and reliability climbs to roughly 70 to 85 percent. For analysis, 70 to 85 percent is genuinely useful. You are reading the output, you can see the reasoning, and a miss costs you a second look. For a bid decision that fires every hour with no one watching, 70 to 85 percent is a liability, not a feature. One wrong call in five, compounding across a roster of campaigns, unattended, is not a system you want.

That single fact is why the architecture splits where it does. It is also why we do not let a narrated rationale stand in for locked logic. The rationale is for you to read. The locked skill is what runs.

From Analysis to Approved Action

Here is where most AI features end and Qore keeps going.

Analysis produces a finding: these fourteen branded terms are absorbing budget that generic discovery could be earning, and here is the reasoning for each. A scheduled skill turns that finding into a specific proposed change. Lower these bids by this amount, shift this budget here, pause these three. Then it stops.

The stop is the approval gate, and it is manual by default. The proposed action sits in a queue with its reasoning attached. You see what it wants to do, the data behind it, and Claude's narration of why. You approve, you edit, or you reject. Only approval moves the bid.

Why default to manual rather than automatic with an override. Because the failure modes are asymmetric. An analysis you ignore costs you nothing. An action that fires wrong costs real spend and takes real work to unwind. The default should protect against the expensive mistake, not the cheap one. Teams that want to automate a well-understood, low-risk action can widen the gate for that specific skill once they have watched it propose the right thing enough times to trust the pattern. That is a decision you make per skill, deliberately, not a setting the system assumes for you.

The scheduling is the other half. A skill runs on a cadence, daily, weekly, whatever the workflow needs, and across a roster: every product, every campaign, every account you point it at. You describe the logic once. It applies everywhere, every cycle, and every proposed action arrives at the same gate. That is how one analyst covers a catalog that would otherwise take a team, without giving up the human sign-off on anything that spends money.

The model reads. The skill proposes. The gate decides who commits. Keeping those three as three separate things is the entire point.

Where AMC Fits

For this workflow, AMC earns its place by giving the analysis two things a standard report cannot. First, audiences built from real behavior: viewed-but-did-not-buy, bought-a-complement, saw-one-campaign-not-the-other. Second, the full path to a conversion instead of last-click credit. Both only matter when they feed a decision, which is the point of pairing AMC with Qore's skills: the audience becomes a targeting call, the path analysis becomes a budget call, and the clean-room work stops being a quarterly slide and starts driving the next cycle.

Why Action and Analysis Get Different Guardrails

The organizing idea of this whole system fits in one line. Analysis is cheap to get wrong. Action is expensive to get wrong. So the two get different guardrails.

An analysis that misfires wastes a few minutes of your attention. You read it, you disagree, you move on. An action that misfires spends money in your account and creates cleanup. Because the costs are not symmetric, the controls should not be either. Analysis earns freedom. Action earns a gate.

Here is the split as a table.

Read across the two columns and the design explains itself. The place a language model is strongest, reading and judging and explaining, is the place we let it run. The place it is weakest, guaranteeing the same call every time with no one watching, is the place we put locked logic and a human in front of it. We are not choosing between Claude and deterministic control. We are using each where it is genuinely good.

What This Does Not Do

Honest limits, because the ones we skip are the ones that bite later.

The action loop assumes Trellis is your bid manager of record. The scheduled-skill-to-approval-gate mechanism operates on the campaigns Trellis manages. If your bids are being set somewhere else, the analysis still reads, but the action half has nothing to act on.

Off-marketplace demand is not always attributable. Demand you create outside the marketplace does not always trace cleanly back to a purchase inside it. AMC widens what you can see, and it does not make everything visible. Treat any tool that claims full-path certainty across every channel with the skepticism it deserves.

Claude narrates; it does not certify. The reasoning attached to a proposal is there to inform your approval, not to replace it. Read it as an argument, a good one, not as a guarantee.

There is a public episode worth sitting with here, because it is the exact risk the gate exists to contain. During Amazon's own MCP Server testing, reported in trade press in April 2026, one agent reached into three years of clean-room data that nobody had asked it to touch. Another defaulted to a deprecated API. The response was telling: Amazon constrained the agent rather than trying to make it smarter. That is the correct instinct, and it is the instinct behind our approval gates. When an agent can act, the fix for unpredictable behavior is a constraint, not more capability.

It is also worth naming what a general-purpose ads agent can and cannot see. Amazon's MCP Server went to global open beta on February 2, 2026, and Amazon updated its Business Solutions Agreement on March 4, 2026 to govern how these agents behave. The tooling is real and genuinely useful for a range of tasks. But it is ads-only. It has no view of your inventory, your true margin after fees, or your Buy Box status. An agent that can move a bid but cannot see that you are about to run out of stock, or that the ASIN it is scaling loses money after fees, is confidently working with half the picture. Business context is exactly what the approval gate reintroduces, because the person at the gate can see the things the ads-only view cannot.

A Concrete Example

Walk one cycle end to end.

Input. A campaign's weekly search-term report. Two hundred terms, real spend spread across them, a mix of your brand name, a competitor's brand name, and generic category queries.

Workflow. A scheduled skill runs weekly. Its locked logic: pull the search-term report, send the terms to Claude for classification into branded, competitor, and generic. For any branded term above a spend threshold where organic rank is already strong, propose a bid reduction, since you are likely paying to rank for a term you would win anyway. Hold every proposal at the approval gate with its reasoning attached.

Expected output. A queue of proposals. "Reduce bid on 'yourbrand tumbler' by 30 percent. Classified branded. Organic rank position 2. Estimated reclaimed budget available for generic discovery terms." Next to it, Claude's narration for each classification, so you can check the branded-versus-generic call yourself. Nothing has changed in the account. The proposals are waiting.

Business decision. You scan the queue. Eleven of the classifications are obviously right and you approve them. One term Claude tagged as competitor is in fact a discontinued sub-brand of yours that you still want to defend, so you reject that one. You approve the batch. Only now do the bids move. Next week the skill runs again, and the terms that shifted show up in the new report, and the loop continues, each cycle gated the same way.

The value is not that a model read a report. Plenty of things read reports. The value is that the reading turned into a specific, reversible, inspected change to your spend, and a person said yes to it before a dollar moved.

Conclusion

The line that separates a useful AI feature from a marketing claim is whether it can do anything with what it notices. Trellis is built around that line.

Claude does the reading and the judgment, the branded-versus-competitor calls and the intent classification and the listing review, and it narrates its reasoning so you can follow it. That work runs freely, because analysis is cheap to get wrong. Qore turns the reading into a proposed action, holds it at an approval gate that is manual by default, and runs the whole loop on a schedule across your roster. That work waits for a person, because action is expensive to get wrong. AMC, as a paid add-on, widens the evidence with clean-room audiences and path-level analysis so the judgments have more than last click to stand on.

None of it is autonomous in the way a demo likes to imply, and that is deliberate. The parts that should run without you, run. The parts that spend your money, wait for you. Naming which is which is the entire design.