Skip to main content

​
Mine automation opportunities into draft agents and flows

Choosing what to automate next is usually guesswork: you skim a few conversations, trust your gut, and hope you picked the highest-leverage workflow. The Automation Opportunity Miner replaces the gut call with a ranked list. It reads your recent support transcripts, groups them into recurring customer intents, and ranks each cluster by the deflection and money an automation could save each month. This guide covers AI → Automation Opportunity Miner (/agents/automation-opportunities) and the two API endpoints behind it. The goal is the same workflow the page assumes: read what the miner found, judge whether a cluster is worth automating, scaffold it into a draft, and review the result before anything goes live.

​
What the miner is

For the look-back window you choose, the miner clusters your support transcripts by intent — “reset password”, “track my shipment”, “update my billing details” — and computes two projections per cluster:
  • Projected deflection — the estimated fraction of those conversations an automation would resolve without a human.
  • Projected monthly cost — the human-agent time that deflection would free up, derived from the cluster’s volume and your assumed per- conversation handling cost.
Ranked by projected monthly savings, the list answers “what should we automate next?” with the expensive candidate on top, not the loudest one. The projections are estimates to prioritise with — verify them against the cluster’s example transcripts before you build.

​
Reading an opportunity row

Each row in the ranked list carries:
  • Volume and share — the number of conversations in the cluster and its percentage of all analyzed conversations in the window. A thin cluster with a high deflection rate is still a thin cluster.
  • Projected deflection — the integer percentage of the cluster an automation could auto-resolve.
  • Projected monthly savings — the human-agent cost avoided, shown in your account currency.
  • Projected ROI — projected monthly savings divided by the estimated build-and-run cost. Shown when it is computable; a high-savings cluster with no ROI figure is still scaffoldable.
  • Recommended build — one of AI agent, flow, KB article, or macro. Agents and flows scaffold in one click; articles and macros are guidance you build on their own surfaces.
Expand View example transcripts on a row to read short, redacted excerpts from the cluster before you commit. A cluster that mixes two intents (“cancel order” and “cancel subscription”) is a sign to build a narrower automation than the label suggests. A row already scaffolded is marked Scaffolded, and its button is disabled so you don’t build the same cluster twice.

​
Scaffold a winner into a draft

On a row whose recommended build is an AI agent or a flow, click Scaffold draft. Orbit turns the cluster into a DRAFT agent or flow and drops you straight into the builder to ground it, review it, and publish it on your own schedule. Nothing activates — the scaffold never wires an outbound channel or flips any live routing. What each build type looks like after scaffolding:
  • AI agent — you land on the agent editor with a starting prompt assembled from the cluster’s intent. Review it against the agent-from-prompt workflow: tighten the prompt, attach a knowledge base, and set handoff targets before you activate.
  • Flow — you land in the flow builder with a starting graph. Review it against the flows recipes: confirm the branch conditions and terminal steps, then publish from the builder.
Treat the scaffold as a first draft, not a finished automation. Sequence your rollout like any other agent launch — evals, guardrails, canary — before routing real traffic to it.

​
Data and API surface

Everything the page shows is available over the API if you want to export the ranking or scaffold from your own tooling:
The response is a report holding the window bounds, the analyzed conversation count, and the ranked opportunities. Each opportunity carries its id, title and summary, conversation volume and share, the projected deflection and savings figures, the recommended build, representative excerpts, and a scaffolded flag. To scaffold from the API, post the opportunity id plus the build type:
Pass ai_agent or flow in automation_type, and pass the same days window you listed the opportunity under — the opportunity id is only valid within the window that minted it, because the scaffold re-derives the cluster from the id. The response returns the new draft’s agent or flow id plus draft: true.

​
Tune what you automate

Ranking by projected savings is the starting sort, not the final word. Adjust what you keep human and what you pass to drafts with two habits:
  • Keep humans on judgement-heavy intents. Anything involving exceptions, goodwill, or account-risk decisions (refund overrides, contract disputes, security incidents) tends to show high projected deflection but deserves human review. Scaffold it into a flow with an explicit human-approval step rather than an autonomous agent.
  • Prefer drafts over rejections. A cluster you decide not to automate still cost your agents time this month. When the recommended build is a KB article or macro, build it on that surface instead of scrolling past — a thin deflection upgrade beats none.
Adjust the look-back window (7 to 90 days, default 30) and the top N list size from the page header. Widening the window catches slower recurring intents that a 7-day slice misses; narrowing it surfaces the current week’s spike.

​
Pitfalls

  • Recent window only. The miner reads transcripts inside the selected window. Seasonal intents that only spike in November never rank if you always mine a 7-day window — widen the window around known peaks.
  • Scaffold is always a draft. The one-click action creates a review artifact and deep-links you into its editor. It never activates an agent, never publishes a flow, and never connects an outbound channel.
  • Projections are estimates. Deflection and ROI figures are derived from the cluster’s traffic and your handling-cost assumption. Check the example transcripts before committing build time to a cluster that only looks coherent from the ranking.
  • Opportunity ids are window-scoped. The same intent under a different window carries a different id. Like-for-like comparisons or API-driven scaffolding must reuse the same days value end to end.