“The fastest way to lose 40 basis points of TACoS is to let an algorithm decide what counts as a converting search term.”
The amazon ppc ai workflow that runs the ads operation across the ClearSight client book is structured around one specific decision. We keep it out of AI’s hands. Which search terms count as winners.
Every other piece of the ad work uses AI somewhere in the pipeline. The rule-engine flagging. The daily campaign-hygiene checks. The bid-action compilation. The campaign-restructuring proposals. Search-term classification is the one place we have decided AI is structurally wrong for the job. It is also where the algorithmic bid optimizers our clients used to run before joining us consistently failed.
This post walks through the amazon ppc ai workflow end-to-end. Where AI is. Where it is not. And why.
The shape, what the daily ads operating cadence looks like
The ads team runs on a daily cycle for every active account. Each cycle is seven steps. Two are deterministic rule-engine work. Three are human work. Two are AI compile, both downstream of human decisions.
The total operator time per account per day, with the amazon ppc ai workflow running properly, is roughly 12 to 25 minutes. Concentrated on the human decision steps. Not on the data collection or the compilation. Without the rule engine and the AI compile layer, the same workflow takes 60 to 90 minutes per account per day. The compression is real.
The ai-only Amazon agency model attempts to compress further by removing the human decision steps. The amazon ppc ai workflow compresses the surrounding work but keeps the human as the binding constraint on what ships.
Step 1 (rule engine), daily campaign-hygiene flags
The bid_rules and campaign_rules modules in the ClearSight portal run on every account every morning at 06:00 Eastern. The data sources are the prior-day Intentwise account snapshot, the Adtomic search-term performance pull, and the native Amazon Sponsored Products / Sponsored Brands / Sponsored Display reports.
The rules flag every account against a fixed list of conditions:
- Search terms that have spent more than $X with zero conversions in the last 7 days
- Campaigns whose impression share dropped more than 30% week-over-week
- Placement modifiers that produced TACoS more than 100 bps above target
- Bid floors that have not been touched in 30+ days but where the search-term ACOS has shifted by 40%+
- Branded defensive campaigns where competitor share-of-voice has increased on the brand’s own terms
This is deterministic. The strategist does not have to remember to look for these conditions. The rule engine surfaces them every morning. Same time. Same place. Same severity-ranked output. This is the layer the offshore VA shop and the ai-only Amazon agency model do not have. That is why their accounts drift week after week without anyone noticing.
Step 2 (human), search-term harvest from STR and SmartScout
The ads strategist opens the Search Term Report for every campaign with rule-flagged anomalies. Plus a rotating sample of campaigns that are not flagged. The strategist also opens SmartScout’s search-term spy on the top three competitors in the category. Looking for terms the competitors are bidding on that the brand is not.
The strategist’s job here is to classify each search term into one of four buckets:
- Winner, converting, profitable, worth bidding harder on or breaking out into its own ad group
- Maintain, converting, marginal economics, keep at current bid
- Negative, not converting, kill it via negative keyword at the appropriate match type
- Watch, not enough data yet, set a reminder for 14 days from now
This classification is human work. AI cannot make it. The reason: “converting” is not just a ratio. It is a judgment about whether the conversions that did happen represent the customer type the brand wants to acquire.
A search term might convert at 8% ACOS but pull in a customer cohort that does not reorder, leaves bad reviews, or returns at three times the category baseline. An algorithm sees the 8% ACOS and labels it a winner. The strategist sees the cohort and labels it negative.
This is the failure mode of every algorithmic bid optimizer the brands we onboard used to run. The optimizer optimized for the metric it could see (ACOS). It ignored the metric that mattered (lifetime value of the cohort the keyword acquired). The amazon ppc ai workflow keeps the human in this loop specifically because the LTV-versus-ACOS tradeoff requires brand context the algorithm does not have. We wrote more about that failure shape in the offshore VA + ChatGPT shops post.
Step 3 (AI compile), search-term clustering and intent classification
Once the strategist has classified the search terms, the cleaned list goes to Claude. Typically 40 to 200 newly-acted-on terms per account per week. With two instructions. Cluster the terms by customer-intent type. Problem-solving. Comparison-shopping. Brand-search. Occasion-driven. And flag any clusters where the strategist’s classification looks inconsistent.
The cluster output goes back to the strategist for review. The intent clustering is the part AI is genuinely good at. Recognizing that “best xylitol gum for kids” and “kid-safe xylitol gum” are the same intent expressed differently. And that both belong in a single ad group rather than splitting bid budget across two near-duplicate campaigns. The strategist accepts the clustering about 70% of the time. Edits the rest.
The “flag any inconsistencies” instruction catches strategist errors. If the strategist labeled two near-identical search terms differently, one a winner, one a negative, Claude flags that for review. Most of the time the strategist had a reason. Sometimes the strategist was rushed and missed the inconsistency. Either way the flag is useful.
Step 4 (human), bid action decisions
The strategist takes the rule-engine flags from Step 1. The human classifications from Step 2. The AI-clustered intent groups from Step 3. And decides which actions to ship today. Bid increases. Bid decreases. Negative-keyword additions. New ad group breakouts. Paused campaigns.
This is the load-bearing decision step of the amazon ppc ai workflow. AI cannot make this call. The strategist is weighing the rule engine’s alerts against the brand’s monthly ad-spend cap, the inventory readiness on the affected SKUs, the timing of an upcoming new-product launch, and a dozen other constraints that live in the strategist’s head and in the account’s history. The rule engine surfaces the situation. The AI compiles the options. The strategist decides.
Step 5 (AI compile), campaign-restructuring proposals
On the weekly cadence, AI runs a second compile pass. Given the trailing four weeks of search-term classifications, the rule engine’s hygiene flags, and the strategist’s bid action history, propose three to five campaign-restructure options. Each proposal includes the expected impact on TACoS, the implementation effort, and the dependencies (inventory, creative, brand approval).
Most weeks, the strategist picks zero or one of the proposals. Some weeks, the strategist picks two. The AI compile here is doing the “given what just happened, what are the structural moves worth considering” synthesis. A strategist could do it. It would take 90 minutes. Doing it in five minutes, with the strategist editing the proposals down to what is actually shippable this week, is where the AI compile layer earns its keep.
Step 6 (human), strategist edit and ship
The strategist takes the day’s bid actions and ships them. This means logging into Seller Central (or into Adtomic, which writes back to Seller Central). Making the bid changes. Adding the negative keywords. Breaking out the new ad groups. Pausing the dead campaigns. The action is logged in ClickUp with the rationale, the data context, and the success criteria for the 14-day post-action review.
This is the step the ai-only Amazon agency model entirely lacks. There is no human shipping the action. The algorithmic optimizer is making the bid changes directly. The cost of that compression shows up in the metrics. TACoS that drifts upward. Branded share-of-voice that erodes. Search-term harvests that never happen.
Step 7 (rule engine), post-action monitoring
Fourteen days after each bid action, the rule engine flags the action for review against its predicted impact. Actions that hit the predicted TACoS lift are logged as wins. Actions that missed are flagged for the strategist to review.
Usually the issue is that an upstream condition changed. Inventory shifted. Competitor entered. Category seasonality flipped. And the action needs a follow-up. This is the closed-loop part of the amazon ppc ai workflow. The same rule engine that surfaces the daily flags also tracks the strategist’s prior actions. It surfaces the ones that need a second pass. Nothing falls through.
The most common failure mode in the ai-only model, the bid change that worked initially and then quietly underperformed two weeks later, is caught by Step 7 before it compounds.
What AI does and does not do in the amazon ppc ai workflow
AI is used in exactly two of the seven steps. Clustering search-term intent (Step 3). And proposing campaign restructures (Step 5). Both are downstream of human judgment. AI does zero of the search-term classification. Zero of the bid action decisions. Zero of the Seller Central shipping work.
The reason is structural. The parts of ads work that are deterministic and pattern-recognition heavy are good for AI. The parts that require brand-context judgment are not.
The stack-reference post earlier in this series describes the architecture in the abstract. The catalog AI/human workflow post walks through the upstream catalog discipline. This post is the ads-specific instance of the same shape.
The next post in the series walks through the creative and A+ content workflow. The human/AI division is even sharper there because brand voice is the binding constraint.
Reviewed by the Amazon Growth Team.
The Hybrid Stack, 10-post series. You are reading post 7 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
- Catalog AI/human workflow
- Ads AI/human workflow (you are here)
- Creative AI/human workflow
- Inventory AI/human workflow
- 12 months of the hybrid stack, results recap
← Previous: Catalog AI/human workflow | Next: Creative AI/human workflow →