Inspectable walkthrough
Why Four Workflows Needed Different AI Systems
Failure cost and review units differed even inside one company. These synthetic walkthroughs show how the responsibilities of data, code, models, and people changed across four workflows.
I Did Not Build One General-Purpose Agent
Interviews revealed one shared bottleneck: people had to prepare data and reconstruct context before they could apply judgment. The consequences of a wrong answer, however, were different in each workflow.
| Workflow | Costly failure | Smallest useful architecture |
|---|---|---|
| Performance marketing | Slow first analysis and inconsistent starting points | deterministic data preparation + one model call + human revision |
| Synthetic survey | implausible respondents and invalid survey state | probabilistic persona + knowledge limits + state rules |
| Proposal | intent and evidence disappear between stages | artifact chain + decision-changing questions |
| ASEO | observed AI answers never become executable content work | answer and citation observation + issue classification + actions |
Performance Marketing · From Prepared Data to Human Revision
The synthetic campaign prioritizes new-customer conversion without increasing total budget. Raw exports become normalized JSON and period-based partitions. The marketer selects the relevant scope, prior action, KPI, and client request.
Raw export
→ JSON normalization and deterministic classification
→ previous/current periods and relevant campaigns
→ Insights / Actions draft
→ marketer revises with client context| Segment | Previous → current | Interpretation boundary |
|---|---|---|
| Search | new-customer CAC 12.0 → 9.6 | efficiency improved; cause remains open |
| Creative A | Frequency 2.8 → 4.1, CTR 1.5% → 1.1% | consistent with fatigue, not causal proof |
| Retargeting | ROAS 10.0 → 10.5, new customers 20 / 209 | separate revenue efficiency from acquisition priority |
The first draft moved budget from Creative A to Search. The marketer changed the sequence: test a replacement creative first, split validation into two stages, and remove the immediate budget shift. The observations remained; the action changed with client context.
The system did not replace the marketer's decision. It created a common input and a fast draft that the person closest to the client could revise.
Synthetic Survey · Give Each Respondent Only What They Could Know
Independent random attributes create contradictory people. The system generates 22 attributes through four dependent layers, using conditional probabilities and correction rules from parent attributes.
The model never receives the entire product brief. It gets only information the persona could plausibly know. Deterministic state owns question IDs, option IDs, selection limits, prior answers, branches, and skips.
Generate a coherent persona
→ select knowable information
→ pass current question, options, and branch state
→ generate response
→ validate selection and branch consistency
→ advance or repairThe system is a pre-test for finding weaknesses in questions, options, branches, and early hypotheses. It is not a replacement for a real panel.
Proposal · One Question Changes the Final Proposal
The fictional D2C company Project Lumen asks for both first-purchase growth and better 90-day repeat purchase without choosing which governs trade-offs. Drafting immediately would force the model to decide silently.
The system asks first:
Which outcome must govern the next 12 weeks: first-time customers or the share that makes a second purchase within 90 days?
The reviewer chooses repeat purchase as the primary KPI and acquisition as supporting. That answer becomes downstream state, not a comment appended to the draft.
| Artifact | What the decision changes |
|---|---|
| Research plan | prioritize post-purchase drop-off and replenishment timing over competitor ads |
| Factbook | foreground repeat rate, product cycle, and CRM reach |
| Strategy | improve the first 50 days and replenishment prompts first |
| Proposal language | frame the gap between first and second purchase as the primary problem |
The question appears before research because the answer changes research scope, evidence, strategy, and proposal language.
ASEO · Turn AI-Answer Observation Into Content Work
Search ranking alone does not reveal how an AI answer explains a brand or which sources it trusts. ASEO starts with a realistic question set, captures answers and citations, diagnoses gaps, and turns them into work.
Question set
→ AI answers and citations
→ brand, keyword, and source analysis
→ gap and citation-pattern classification
→ content opportunities and writing actions
→ human reviewThe key decision was not to ask AI to write content immediately. The observed weakness in an actual answer becomes the input to the next content task.
Evidence Boundary
Campaign metrics, Project Lumen, survey personas, and ASEO questions are synthetic. The confirmed operating patterns are raw-input preparation, scope selection, human revision, intermediate artifacts, decision-changing questions, and the four responsibility structures. This page does not reproduce client outcomes or isolate an AI-only effect.