Agents
Nine specialists working your account in the background. Three Optimisers act on it, three Watchers alert you to changes, two Analysts make sense of the data, plus a Generalist you can talk to in plain English. Powered by Claude where it helps. Deterministic math everywhere else.
30-day trial · no card · flat pricing, never % of spend
Pick an agent. Each one says what it does, when it runs and what powers it, because “AI” should never be a hand-wave.
Reallocates daily spend inside each portfolio every night. It fits a saturation curve per campaign from your last 30+ days, then solves for the mix that maximises ROAS within your velocity caps and budget locks. The only agent that moves money, and it is pure math.
optimiser · nightly · runs on math · staged budget changesScans your Google search terms and active keywords every Sunday. Deterministic spend and close-rate filters pick the candidates, then Claude ranks each one against your portfolio goal: “second-hand fitting request” gets flagged as off-intent because the agent knows what the portfolio is actually for, a judgment a regex can’t touch.
optimiser · weekly, sundays · math + claude · review queueSpots fatigued ads across Google RSA, PMax and Meta weekly: Low-rated assets, Below-average ads with rising frequency. Fatigue signals find the candidates, Claude judges each against your goal rather than a hard threshold, and sibling-count guards make sure a campaign is never left without a working ad. Everything lives in one creative library, where platform and type are filters, not pages.
optimiser · weekly, sundays · math + claude · review queueFlags portfolios drifting off their monthly spend pace: a daily digest when month-to-date spend runs more than 15% off the prorated target, worst offenders first.
watcher · daily · runs on math · digest + alertsA ROAS drop, a revenue gap between a platform and your shop or CRM, campaigns choked by low search volume: when a metric moves unexpectedly, it ranks the likely causes from your own account data, then Claude writes the explanation in plain English.
watcher · on anomaly · math + claude · notificationsContinuously grades measurement health: tracking gaps, broken UTMs, revenue divergence between ad platforms and your shop or CRM. When something breaks, it proposes the specific fix.
watcher · continuous · runs on math · audit tabRe-fits every campaign’s response curve nightly, then grades the engine’s own predictions against what actually happened. The receipts live in the Calibration tab.
analyst · nightly · runs on math · calibration tabWrites your weekly brief and daily mini-brief: plain-English prose that narrates what the team did, agent by agent. A real ops report, not a chart dump.
analyst · weekly + daily mini · runs on claude · briefsChat with full team access. Ask anything in plain English and Flume reads your data, proposes actions, and hands structured plans back. Your own AI tools can direct this team too; the full story lands on the Flume stop.
generalist · on demand · runs on claude · chat + mcpNegative keywords, pause candidates and fatigued creative are the highest-ROI tactical work in PPC, and the thing most accounts do worst. On Meta, creative is the only daily lever there is. The two hunters scan every week and stack their findings in one client-wide review queue: approve, snooze or dismiss in one place.
A standalone tool flags “this term had 0 Google conversions.” ROASt flags “this term drove 8 leads that never closed in your CRM.” When HubSpot or Shopify close-rate context exists, the suggester routes through it.
When the fatigue scanner and the engine’s anomaly detector agree on the same campaign, the Audit tab raises a critical finding. Two systems agreeing is a much stronger signal than either alone.
Pause candidates and wasted-spend terms apply together: one button, parallel API calls, consolidated result. Every apply uses the smallest reversible unit per platform, and sibling-count floors are re-checked at apply time so an ad group or campaign is never left stranded.
Search-term scanning is Google-only at launch; Microsoft and Meta sync are on the roadmap. Creative is pause-only by design: ROASt won’t generate copy, duplicate ad sets or edit audiences. Suggestions expire and reactivate on real signals: three strikes, two-week absence, or a 2× spend resurgence.
Five of the nine run on deterministic math. The rest use Claude for judgment and prose, never for arithmetic. And no agent autonomously changes budgets or targets: that is the Allocator’s job, and the Allocator is pure math operating inside your own caps.
