Skip to main content

​
Automation Opportunity Miner

The Automation Opportunity Miner (dashboard → AI → Automation Opportunity Miner, /agents/automation-opportunities) answers one question: what should you automate next? It reads your recent support transcripts, clusters them into recurring customer intents, ranks each cluster by projected deflection and money saved per month, and offers a one-click Scaffold that turns a cluster into a draft AI agent or flow. This page is the reference for that surface: what each column means, how the mining works, who can act on it, and what a scaffold actually produces. For the concept model behind it see Automation opportunities model; for the end-to-end workflow see Mine automation opportunities into draft agents and flows.

​
What the surface shows

Each row in the ranked list is one recurring customer intent mined from your transcripts, carrying:
  • Title and intent summary — the cluster’s label (for example “Account access & password resets”) and a one-line description of what those customers are asking for.
  • Volume — the number of conversations in the cluster over the window, and its percentage share of all analyzed conversations. A cluster with a high projected deflection but three conversations is still a thin cluster.
  • Projected deflection — the estimated percentage of the cluster an automation could resolve without a human. High for deterministic, policy-driven intents (password resets, order tracking); low for judgement-heavy ones (complaints, refund overrides).
  • Projected monthly savings — the human-agent cost the deflection would avoid each month, shown in your account currency. This is the ranking figure: the list is sorted by it, most expensive candidate on top.
  • Projected ROI — projected monthly savings divided by the estimated build-and-run cost for the automation. Shown only when it is computable; a cluster without an 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 (the knowledge base or saved-replies surfaces respectively).
  • Example transcripts — expand View example transcripts to read short, redacted excerpts from the cluster before committing. A cluster that mixes two intents (“cancel order” and “cancel subscription”) is a sign to build a narrower automation than the label suggests.
Above the list, three aggregate cards sum the whole report: transcripts analyzed in the window, projected monthly deflections across all listed opportunities, and projected monthly savings.

​
How mining works

The miner reads the cross-channel conversations records in your workspace — not chat transcripts only — over the look-back window you pick in the page header (7 to 90 days, default 30; up to 365 via the API). For each conversation it uses the LLM-classified intent label when one exists, and falls back to a contact-reason lexicon over the conversation summary (or last message) when there is no label. Conversations in the same intent bucket form one cluster. Two outcome signals gate every projection:
  • Resolved without a human assignee counts as already deflected — no saving left there.
  • Resolved with a human assignee counts as still manual — the automatable backlog the miner surfaces.
  • Unresolved or still open counts toward volume but never toward savings.
A cluster only surfaces when it holds at least three still-manual conversations, so one-off utterances and long-tail noise never rank. Handle time is averaged from the still-manual conversations only, so self-served conversations do not drag the per-conversation cost estimate down. Projections are priced from your assumed human-agent cost; the default is $30/hour when no override is set, and the defendant’s monthly run-rate is scaled from the window length (30-day equivalent). All figures are estimates to prioritise with — verify them against the cluster’s example transcripts before building.

​
Reading the list

The list is sorted by projected monthly savings descending; ties fall back to conversation volume, then title. You control two knobs in the page header:
  • Look-back window (7 / 14 / 30 / 60 / 90 days) — widen it to catch slower recurring intents a 7-day slice misses; narrow it to surface the current week’s spike.
  • Top N (5 / 10 / 20 / 50) — how many clusters the list shows.
A deflection-versus-ROI tradeoff falls out of the columns directly: a high-deflection cluster on a low-volume intent can project less monthly savings than a moderate-deflection cluster on your highest-volume intent. Rank by money unless you have a specific reason to do otherwise — and treat judgement-heavy intents (complaints, disputes) with suspicion even when they rank highly, because their projected deflection is capped low for a reason. A row you already scaffolded is marked Scaffolded, and its button is disabled so you do not build the same cluster twice.

​
Scaffold to DRAFT

On a row whose recommended build is an AI agent or a flow, click Scaffold draft. The miner turns the cluster into a draft — never an active automation, never a wired outbound channel, never live routing. Scaffolding requires the agents:write scope and an owner, admin, or developer role; read-only roles can view the list but cannot scaffold. What the scaffold creates:
  • AI agent draft — a new draft agent named Draft: <cluster title>, pre-filled with the intent summary as its description and a starting system prompt assembled from the cluster: what the intent is, instruction to resolve end to end grounded in the knowledge base, and an explicit reminder to review and test before activating. You land on the agent’s page (/agents/[id]) to attach knowledge base documents, set handoff targets, and ground it via the normal authoring loop.
  • Flow draft — a new draft flow named Draft: <cluster title> with a manual-start trigger node as its starting graph. You land in the flow builder (/flows/builder) with the draft loaded to design the branches and terminal steps.
Both pre-fill the intent description and stamp the opportunity id onto the draft so the list marks the row as scaffolded. Suggested knowledge base documents are not attached automatically — attach them yourself from the agent editor as part of grounding (see Creating agents). KB article and macro recommendations are not scaffoldable from this surface; build them on the knowledge-base or saved-replies surfaces, guided by the cluster’s example transcripts.

​
After scaffolding

A scaffold is a first draft, not a finished automation. The next step depends on the kind:
  • AI agent draft — continue in Creating agents: tighten the prompt, attach the knowledge base, set handoff targets, run evals, then activate.
  • Flow draft — continue in the flow builder: confirm branch conditions and terminal steps, then publish.
For the full recipe — when to widen the window, how to sequence a rollout, and how this miner relates to the Deflection Gaps and Auto-draft help articles surfaces — see the workflow guide.

​
Tuning the list

The miner deliberately excludes what would waste your attention:
  • One-off utterances — a cluster needs at least three still-manual conversations in the window to surface. Beneath that floor it is indistinguishable from noise.
  • Low-headroom clusters — only the still-manual share of each intent is projected. An intent your existing automations already deflect well drops out of the ranking naturally as its manual share shrinks.
  • Empty-excerpt clusters — rows with no usable transcript text contribute volume but cannot show examples.
A previously-listed cluster can disappear between visits when its still-manual count falls below the floor (your agents handled it), when the window shifts and the cluster ages out, or when upstream intent labels change and the same conversations re-cluster under a different title. Widen the window before concluding the intent vanished. Opportunity ids are window-scoped: the same intent under a different look-back window carries a different id, because the scaffold re-derives the cluster from the id within that window. Like-for-like API comparisons must reuse the same days value end to end.

​
Troubleshooting

The list is empty. The empty state means one of two things:
  • Insufficient volume — no intent cleared the three-still-manual-conversations floor in this window. Widen the look-back window to analyze more history.
  • Nothing left to automate — your existing automations already cover the recurring intents in the window. The miner only surfaces still-manual clusters; a fully-deflected intent is by design invisible here.
A workspace with no transcripts in the window at all shows the same empty state — check Transcripts analyzed in the aggregate strip: zero means the corpus itself is empty, not just the ranking. Scaffold fails. Common causes:
  • 404 not found — the opportunity id no longer resolves in the selected window, because the window shifted and the cluster aged out or re-clustered. Re-list with the same days value and scaffold from a fresh row.
  • Role or scope error — scaffolding requires agents:write and an owner, admin, or developer role. View-only roles can read the list but the button will not create anything.
  • Duplicate build — a row marked Scaffolded was already turned into a draft. Its button is disabled; look for the Draft: <title> agent or flow in the respective list instead of scaffolding again.

​
API surface

Everything the page shows is available over the API, under the same tenant scoping:
days accepts 1–365 (default 30), limit 1–50 (default 10). The response carries the window bounds, the analyzed conversation count, and the ranked opportunities with every field listed above. Scaffold from your own tooling by posting the opportunity id plus the build type — days must match the window the opportunity was listed under:
Pass ai_agent or flow; the response returns the new draft’s agent_id or flow_id with draft: true.
Projected deflection, savings, and ROI are estimates computed from your own transcript traffic and cost assumptions — not a billing commitment. Treat the ranking as a triage aid, and validate a scaffolded draft on real traffic before activating it.

​
See also