Ecommerce AI agents · back office

Specialists that never sleep and always show their work.

Each agent owns a domain, scans on a schedule, writes up what it found, and proposes what to do.

7 on shift

each dot is a specialist’s scheduled scan

this morning’s findings · tap one for its work

  • Before six, the shop is quiet. The first scan is at 06:00.

A morning on a demo store, on the agents' real schedules. Tap a finding for its reasoning — figures are illustrative.

One generalist can't run a store. A team can. Specialists go deep where a single assistant goes shallow — and because they share what they learn, the whole team gets smarter together.

Live now

Your first hire: the Inventory Forecaster.

It watches stock across your whole catalog on a schedule, flags products as critical or low against thresholds you control, and opens priority-ranked reorder recommendations — every number explained. Velocity-aware stockout projections are in the works. Review its page, accept a reorder, or run a fresh scan.

Inventory Forecaster

live
Review reorders

thresholds you control · priorities · reasoning — every number explained

Also on shift

Seven specialists, live today.

The Inventory Forecaster works alongside six colleagues — each with its own domain, schedule, and page.

Profit Lens

Computes true per-product margins from your costs — profit KPIs and margin signals.

daily 07:00live

Liquidation Planner

Finds dead stock and proposes bounded clearance discounts you approve.

daily 08:00live

Bundle Discovery

Mines your orders for products that sell together — real pairs become bought-together suggestions.

Mondays 09:00live

Demand Radar

Reads storefront conversations for demand your catalog didn't capture.

daily 07:00live

Revenue Detective

Decomposes every revenue move into quantified drivers — volume, basket, refunds, mix.

Mondays 07:00live

Self-Learning FAQ

Finds the questions that stumped the assistant and drafts the missing FAQ entries for one-click approval.

daily 08:00live

How an agent works

Scan. Understand. Recommend. Then act, with your nod.

They collaborate

Findings the whole team can trust.

When one agent establishes a fact about your store, it’s written once and reused by the others — stamped with who found it and how confident they were.

On your schedule

On a cadence, or on demand.

Agents run on a schedule — nightly, weekly — and on demand when you hit “Run now.” You’re never waiting on a batch.

Autonomy & safety

Recommend first. Earn autonomy later.

Everything starts in recommend mode — the agent suggests, you approve. Independence is then dialled up by area of the store rather than by agent: discounts, triggers, returns, FAQ and eleven more, each with its own level and its own limits. Trusting Lumnya with your FAQ never means trusting it with your refunds.

Confirm is the default — a category you never touch stays there
Autopilot runs only inside caps you set, and honors quiet hours
Refunds, customer emails and return decisions never automate at all
Per-store isolation of everything an agent learns
Every action journalled — and reversible ones undo in one click

Autonomy, per area

6 of 15 areas shown

  • FAQ answers

    level 1 · confirm

  • Discount codes

    level 1 · confirm

  • Prices

    level 1 · confirm

  • Triggers

    level 1 · confirm

  • Refunds

    never unattended

  • Customer emails

    never unattended

every area starts at Confirm · caps: a ceiling per action and per month · a daily maximum · quiet hours · one switch pauses everything

Try it — this panel changes nothing.

Did it actually work?

Every action gets marked against reality.

A recommendation you accepted isn't the end of the story. Lumnya goes back afterwards and measures what the action did — from your real data, never from the model's own opinion of itself.

A discount code is measured against the orders that actually used it
A trigger change is measured as the firing-rate delta, equal windows either side
An FAQ edit is measured as the change in how often chats escalate to you
Measurement starts a day after the action and settles over a week
What can't be attributed is shown as “—”, never dressed up as a win

measured against your real data

action taken

day 0 · actionday 1 · measuringday 7 · settled
  • Discount code SPRING10orders that used the code31 orders · €1,840
  • Abandoned-cart trigger, earlier sendfiring rate, equal windows+18% firings
  • FAQ: shipping to Cypruschats escalated to you−3 a week
  • Bundle suggestion, Hoodie + Beanieno clean way to attribute—

never the model’s opinion of itself · unattributable stays “—”

Joining the roster

The next hires are already in training.

Specialists across finance, marketing, product, customer intelligence, and operations — starting with inventory. And the roster keeps growing.

Pricing Optimizer

Watches margins and elasticity, proposes price moves worth making.

next

Churn-Risk Scorer

Flags the customers most likely to slip away — early enough to act.

next

Sentiment Pulse

Reads reviews and tickets so you feel the mood without reading them all.

next

Anomaly Detector

Beyond the Store Watchdog's three alarms: the same treatment for every metric you track.

next

Executive Summary

Your whole store, briefed like a Monday-morning memo.

next

Your roster, your priorities

Tell us which specialist would change your week most.

We weigh every request when deciding what ships next.

Private beta

Build your back office, one specialist at a time.

Join the private beta and be first in line when Lumnya opens to your store.