Automation phases (Phase 1 to 4)
Picaro.AI is designed to deepen operational automation in four phases. Jumping straight to full automation carries both psychological and operational risk, so the path starts with one-shot manual execution and advances only as trust accumulates.
The four phases at a glance
Section titled “The four phases at a glance”| Phase | State | Typical use | Target timing |
|---|---|---|---|
| Phase 1 | One-shot | Send prompts on demand through your AI agent environment for immediate execution | Current |
| Phase 2 | Daily-executable | Register the same prompt as a saved prompt and run it every morning with one click | Q3 2026 |
| Phase 3 | Auto-execution with approval flow | Candidates are extracted automatically; a single approval click in Slack applies the change | Q4 2026 |
| Phase 4 | Fully automatic | Only proposals with sufficient ML confidence execute automatically within the guardrails | 2027 |
What changes in each phase, and what conditions need to be in place to move forward, are covered below.
Phase 1: One-shot (current)
Section titled “Phase 1: One-shot (current)”Every skill starts here. Sending a prompt through Claude or ChatGPT, Picaro responds immediately.
- Trigger: the operator sends a prompt manually in your AI agent environment
- Parameters: specified inside each prompt (some fields, like monthly KPI targets, are referenced automatically once registered)
- Application: bid, budget, and negative-keyword changes do not take effect until “execute” is approved
The point at this stage is to understand the output of each prompt and the meaning of the guardrails (caps on increase or decrease, scope, target count). The on-ramp to Phase 2 and beyond is built on the prompt-execution experience here.
Phase 2: Daily-executable
Section titled “Phase 2: Daily-executable”In this phase, the same prompt can be re-run with one click every morning. The prompt is registered as a saved prompt on the Picaro side and invoked from your AI agent environment.
Prerequisites to move from Phase 1 to 2
Section titled “Prerequisites to move from Phase 1 to 2”- Monthly KPI targets are registered (no need to retype the target ACoS each time)
- At least four weeks of recent data have accumulated (so period-over-period comparison is meaningful)
- Output quality has been validated through one-shot execution
- The prompt is registered as a saved prompt
Skills that become practical in Phase 2 (examples)
Section titled “Skills that become practical in Phase 2 (examples)”- Period-over-period KPI comparison on the dashboard (daily check routine)
- Drift alerts for favorite keywords
- Anomaly detection (sudden KPI shifts pushed to Slack)
- Drift monitoring on bid recommendations
- Mid-month KPI progress checks
The goal of Phase 2 is a five-minute morning routine that runs “dashboard KPI -> anomaly detection -> favorite keyword drift” in sequence.
Phase 3: Auto-execution with approval flow (3a)
Section titled “Phase 3: Auto-execution with approval flow (3a)”Candidate extraction is fully automated, and only the final approval needs a single click from a human. A weekly batch generates proposals, Slack delivers the notification, and one button press applies the change.
Prerequisites to move from Phase 2 to 3
Section titled “Prerequisites to move from Phase 2 to 3”- Automation rules are registered explicitly (scope, item cap, monetary cap)
- An approval flow (Slack notification plus button) is defined
- Guardrails are written down (for example, +20% / -30% caps on adjustment magnitude)
Skills that move to automatic execution in Phase 3 (examples)
Section titled “Skills that move to automatic execution in Phase 3 (examples)”- Bid reductions based on ACoS (weekly batch, cap of 100 items per week, one-click Slack approval)
- Adding negative keywords for high-ACoS / zero-conversion terms (weekly, up to 50 items)
- Daily budget adjustment based on spend rate and ACoS (daily, within the monetary cap)
- Auto-to-Manual promotion (weekly automatic candidate extraction plus manual approval)
- Monthly report auto-generation and Slack delivery
Because these actions move money directly, Phase 3 always routes through the approval flow. Nothing applies without approval.
Phase 3b to 4: Fully automatic
Section titled “Phase 3b to 4: Fully automatic”Once approval history has accumulated and ML confidence exceeds the threshold, proposals execute without human involvement.
Prerequisites to move from Phase 3a to 3b / 4
Section titled “Prerequisites to move from Phase 3a to 3b / 4”- A rejection rate below 10% across approval executions in the most recent four weeks
- Zero rollback incidents
- An ML confidence score of 0.85 or higher
- Approval from the responsible executive
- An automatic rollback mechanism for anomalies is in operation
Skills that become fully automatic in Phase 4 (examples)
Section titled “Skills that become fully automatic in Phase 4 (examples)”- ML-confidence-based automatic bidding (only proposals at 0.85 or higher)
- Automatic budget acceleration during sale periods
- Automatic campaign pause when inventory runs out
- Competitor ASIN monitoring with automatic response candidate generation
- Strategy adjustment by product lifecycle stage
Even at Phase 4, actions tied to strategic decisions — such as creating a new campaign — remain in human hands. Not “everything” becomes automatic; only the areas where accountability is clearly scoped become automatic.
Guardrails apply in every phase
Section titled “Guardrails apply in every phase”Automation deepens, but the guardrails are always active.
- Item cap: how many targets a single batch can touch
- Monetary cap: monthly budget caps and per-proposal monetary caps
- Magnitude clamp: no single adjustment exceeds +20% / -30%
- Scope: exclusion settings at the account / portfolio / campaign / keyword level
- Approval flag: each skill can be switched between automatic and approval-required
Even in Phase 4, anything outside the guardrails requires human judgment. The design is structured to prevent automation from running away.
Why the phased approach
Section titled “Why the phased approach”There are three reasons full automation is not offered from day one.
- Data accumulation is required: precision in automatic execution learns from past approval / rejection records. The training signal accumulated across Phases 1 to 3 is what supports the Phase 4 confidence model.
- Operational alignment: existing workflows, internal approval routines, and monthly reporting cycles do not fit full automation overnight. Organizational practices evolve alongside the phases.
- Clear accountability: “the AI decided” leaves accountability ambiguous. Going through propose -> approve -> execute records who decided what at which stage, providing the transparency enterprise governance requires.
Where to look to gauge the current phase
Section titled “Where to look to gauge the current phase”| Signal | Likely phase |
|---|---|
| Sending prompts on demand in your AI agent environment and reviewing results | Phase 1 |
| Morning routine is fixed and uses saved prompts | Phase 2 |
| Automation rules are registered and proposal approvals run through Slack | Phase 3a |
| Proposals above the confidence threshold execute automatically | Phase 3b / 4 |
Moving from Phase 1 to Phase 2 happens by registering monthly KPI targets and creating saved prompts. Moving from Phase 2 to Phase 3 requires explicit automation rules and Slack integration.
What to do next
Section titled “What to do next”- Set up automation rules — the on-ramp from Phase 2 to 3
- Set monthly KPI targets — the prerequisite for Phase 1 to 2
- View KPIs on the dashboard — the core of the Phase 2 daily routine