← Back to selected work

PRODUCT CASE

OnThePlanet Internal AI Systems

Interviews revealed the same contradiction across four workflows: people with the most judgment to contribute were preparing data, while their best reasoning remained individual rather than reproducible across the team.

  • Role: problem discovery, workflow redesign, system design, and implementation lead
  • Scope: performance marketing, synthetic survey pre-testing, proposal co-building, and AI-search operations
  • Status: deployed internal systems

Inspect now

The public walkthrough reconstructs the workflows with synthetic inputs to protect internal data. Adoption ranges and observed outcomes are bounded separately in the case.

Why These Systems Needed to Exist

Interviews with performance marketers and proposal teams revealed the same contradiction. The people expected to produce better insights and actions were spending that time cleaning source data and moving information. Strong decisions also stayed with experienced individuals, forcing the next person to solve the same problem again.

This was not simply a prompt-quality problem. The model needed a common data shape, the workflow needed durable context, people needed explicit decision points, and each output needed a state the next step could understand.

One Company, Four Different Architectures

WorkOperational bottleneckChosen structure
Performance marketingRepeated data preparation and analyst-to-analyst varianceStructured input, one initial model call, human revision
Survey pre-testQuestion, option, and branch failures found only after launchProbabilistic personas, information limits, deterministic survey state
Proposal creationRFP interpretation, research, evidence, and strategy collapsed into one generationIntermediate artifacts, decision-changing questions, human approval
AI-search operationsSearch rank could not explain a brand's place inside AI answersObserve questions, answers, mentions, and citations, then connect them to content action

The goal was not orchestration complexity. Each workflow used the smallest architecture that matched its failure cost and review unit.

What Users Actually Do

Performance Marketing Partner

Raw performance data is normalized into JSON. The user chooses a scope, with time periods as the default unit because that already matches client communication. Selected data is assembled with prior actions, KPI, client requests, and optional creative material. The system produces two stable sections, Insights and Actions, which the marketer revises in conversation.

Synthetic Survey

Conditionally related attributes generate plausible personas. Each respondent receives only information they could know. Question options, selection limits, branches, and skips run through explicit state and rules rather than free-form dialogue.

Proposal Co-building System

An RFP does not jump straight to proposal copy. The process moves through RFP analysis, research questions, a research plan, factbook, strategy, and a text proposal. The agent asks only when an answer would materially change the next artifact.

ASEO · An Additional System

I also built a system that observes what AI answers to brand-related questions, which sources it cites, and where the answer leaves gaps. A fixed question set, answer and citation capture, issue classification, and action generation turn those observations into content work.

Responsibility Split

OwnerResponsibility
Data structuresPreserve prior decisions and readable state for the next step
Deterministic codeNormalize inputs, partition scope, move state, branch, and validate format
ModelsExplore open-ended possibilities, interpret evidence, and draft insights
PeopleOwn priorities, client context, exceptions, approval, and revision

Problems Solved While Building

Data Readiness: Create a Common Starting Point Before the Model

  • Observed problem: Every marketer prepared source data differently, so prompt consistency could not remove output variance.
  • Reasoning: If classification and scope selection remained inside the model, failures would be harder to locate and compare.
  • Implementation: Raw data was normalized into JSON and deterministically classified. Time period became the default scope while users retained control.
  • What changed: Work started from a common input rather than repeated manual cleanup.

Proportional Architecture: Match System Size to Workflow Risk

  • Observed problem: Turning every workflow into a multi-agent chain would increase latency, maintenance, and review burden.
  • Reasoning: Fast analysis needed a reviewable first draft; proposals needed rationale and evidence to survive across stages.
  • Implementation: Marketing analysis used one controlled initial call plus human revision. Proposal work used persistent artifacts and stage-level review.
  • What changed: Complexity followed operational risk, not the desire to demonstrate more agents.

Intent Preservation: Keep Human Judgment Across Artifacts

  • Observed problem: In long proposal work, decisions left only in chat were easily lost downstream.
  • Reasoning: The system should not ask everything. It should ask when the answer changes research scope, evidence, or strategy.
  • Implementation: Answers and intermediate decisions were written into artifacts consumed by the next stage, with human approval and revision points.
  • What changed: Newer team members could follow the reasoning sequence while experienced users could deepen it.

Adoption: Put AI Inside Existing Review Habits

  • Observed problem: A strong demo does not become a habit if it ignores how the team already reviews work and talks to clients.
  • Reasoning: Generated text should be a starting point inside familiar units, not a final answer outside the workflow.
  • Implementation: Period-based inputs, two stable output sections, and document-centered revision stayed aligned with existing work.
  • What changed: The Performance Marketing Partner became part of the weekly standard process for roughly 15 marketers and remains in use.

Observed Change

  • Performance Marketing Partner: used in weekly work by a team of roughly 15
  • Synthetic Survey: internal pre-test from about two weeks to roughly one business day
  • Proposal System: preparation from about two to three weeks to about one; factbook from days to roughly one to two hours
  • A proposal written by three employees outside marketing with the system was selected in an actual bid

These are before-and-after ranges recorded while teams repeated the same work. The walkthrough below uses synthetic inputs to show the structure and human intervention without exposing client information.

Inspect the Work

Next caseGetsit AI