AI inside Karbon Analytics

AI for ecommerce,
grounded in your unified model.

Detectors run across Shopify, Meta, Google Ads, GA4, and Klaviyo. AI writes each brief from your unified model, never from a guess.

Start free trial Talk to us
Available on Shopify App Store
Karbon Analytics
Live

Signal detected

Meta ROAS dropped 38% in 5 days

Trigger: ROAS · 5-day rolling · breakeven threshold

−$5.8K
AI explanationfrom model

Three campaigns account for $4,200 of the $5,800 spend swing. C-23 creative is at 2.8x frequency in the 25-34 audience. CPA up +64%.

High prioritySave ~$1.4K/wk

Pause C-23, rotate hero creative on C-17

Reasoning

Frequency above 2.5x triples CPA in this audience. C-17 creative rotated 14d ago and still converts at 3.1x.

01 / 04

How AI works in Karbon Analytics

Detect. Explain. Recommend.

Detection is deterministic. Explanations are written. Recommendations are reasoned. The AI never invents a number, only translates one.

[01] Detect

Patterns surfaced from the unified model

40+ rule-based detectors run every night against the model. ROAS drops, stockout risk, refund spikes, cohort retention shifts, channel mix changes, wasteful ad spend, scaling opportunities.

  • Deterministic, rule-based
  • Ranked by impact
  • 40+ patterns covered
[02] Explain

Plain-language explanations with context

AI reads the detected signal and decomposes the move against the model. Which channel contributed. Which cohort. Which SKU. Which campaign. The explanation cites the numbers, not the gut.

  • Numbers from the model
  • Cited contributors
  • Operator-readable
[03] Recommend

The next action, named and reasoned

AI proposes the next move. Pause this campaign. Replenish this SKU. Investigate this funnel. Scale this audience. Every recommendation ships with the reasoning so you can trust the call.

  • Action named
  • Reasoning shown
  • Priority ranked

What the AI produces

Three surfaces, one job.

A brief in your inbox. An explanation under every metric. An action proposed for every signal. Different surfaces, same engine.

Surface 01 · The brief

Wake up to an actionable brief.

Every morning, the brief lands with what changed in the last 24 hours. Ranked by impact. Each signal ships with the explanation and the next action, so you start with decisions, not dashboards.

  • Sent before market open, every weekday
  • Ranked by business impact, not chronology
  • Multi-recipient: founder, marketing, ops
Karbon Analytics

Subject

Yesterday in numbers: 1 risk to fix, 1 lever to pull, 1 SKU to reorder

01

Meta ROAS dropped 38% in 5 days

−$5.8K revenue

Three campaigns (C-23, C-17, C-09) account for $4,200 of the $5,800 spend swing. Creative on C-23 is at 2.8x frequency in the 25-34 audience.

Next step Pause C-23, rotate hero creative on C-17
02

Klaviyo Flow A re-engagement scaling

+$9.4K revenue

Re-engagement flow delivered 38% of last week’s lift on its own. Attributed revenue $9,400, ROAS 6.2x, up from 4.1x baseline.

Next step Increase flow audience window to 90 days
03

3 SKUs trending below 14-day cover

$11.2K at risk

TOWEL-SET-04, MUG-AMBER-12, and CANDLE-IRIS hit 9, 11, and 13 days of cover at current velocity. Lead time to replenish averages 12 days.

Next step Trigger replenishment from supplier

Generated by Karbon Analytics · reads from your unified model

Open full brief →
Karbon Analytics
AI Explain

Revenue · last 7 days

$84,329

+12.4% vs prior 7 days
AI explanation
2.4s · from model

Revenue rose +12.4% driven primarily by the Klaviyo re-engagement flow (+$9.4K, ROAS up to 6.2x from 4.1x baseline). Meta campaign C-17 contributed +$4.2K after a creative rotation on May 14. Direct returning customers added +$3.8K, with no change in acquisition spend.

Contributors

  • Klaviyo Re-engagement flow +$9.4K
  • Meta Ads Campaign C-17 +$4.2K
  • Shopify Direct returning +$3.8K
  • Google Brand search +$1.6K
  • GA4 Organic landing −$1.2K

Surface 02 · The explanation

Why a number moved.
Not just that it did.

Click any metric and AI decomposes the move. Which channel contributed. Which cohort. Which SKU. Which campaign. Plain language, sourced from the unified model.

  • Decomposed by channel, cohort, SKU, campaign
  • Contributing values cited inline
  • Available on every KPI, not just headline metrics

Surface 03 · The action

Decisions,
not just descriptions.

Every explanation ends with what to do next. Pause this campaign. Replenish this SKU. Scale this audience. Investigate this funnel. Reasoning shown, so you can trust the call.

  • Action named, not implied
  • Reasoning visible, every recommendation
  • You stay in control: dismiss feedback weights future signals
Karbon Analytics
01
High Save ~$1.4K/week

Pause Meta campaign C-23

Meta Ads · Campaign C-23

Reasoning

Frequency at 2.8x in 25-34 audience, CPA up 64% in 5 days, ROAS at 0.7x against 2.1x breakeven.

Apply Dismiss Snooze 7d
02
High Protect $11.2K revenue

Replenish 3 SKUs before May 27

Shopify · Inventory

Reasoning

TOWEL-SET-04, MUG-AMBER-12, CANDLE-IRIS at 9-13 days cover. Lead time averages 12 days from current supplier.

Apply Dismiss Snooze 7d
03
Medium Lift ~$3.4K/week

Scale Klaviyo re-engagement flow

Klaviyo · Flow A

Reasoning

ROAS lifted to 6.2x from 4.1x baseline. Current audience window is 60 days. Expanding to 90 days adds ~$3.4K incremental.

Apply Dismiss Snooze 7d

Recommendations regenerate as your data evolves

Last refresh: 4m ago

What the AI covers

Eight categories of insight.
Read by AI, every night.

[01] Revenue

Revenue trends

Explanations for revenue moves, week-over-week and against your baseline.

  • Revenue lift attributed to flow
  • Weekly cohort drag
  • AOV shift by channel
[02] Ad

Ad spend efficiency

Wasteful spend called out, scaling opportunities ranked.

  • $1.2k spend, 0 purchases
  • ROAS below breakeven
  • Scale candidate identified
[03] Customer

Customer behavior

Repeat rate, churn signals, segment shifts, in plain language.

  • Repeat rate down 8pts
  • New-vs-returning shift
  • Segment churn lift
[04] Inventory

Inventory risk

Stockout, slow-mover, overstock callouts with replenishment urgency.

  • 3 SKUs near stockout
  • Slow-mover capital tied
  • Overstock margin drag
[05] Refund

Refund patterns

Refund spike explanations drilled to SKU, reason, and channel.

  • Refund rate +3.4pts
  • SKU-level driver named
  • Refund-by-channel split
[06] Channel

Channel mix

Which channels are gaining or losing share, and which signals to read it from.

  • Direct share dropping
  • Klaviyo flow lift
  • Paid social compression
[07] Cohort

Cohort retention

How recent cohorts compare to baseline curves. Lift, drag, and reasons.

  • M2 retention drag
  • Cohort LTV outpacing
  • Promo-acquired cohort risk
[08] Margin

Margin & pricing

Contribution-margin compression, AOV moves, discount impact.

  • Margin compression flagged
  • Discount erosion
  • AOV step-up by SKU

Coverage grows as new ecommerce patterns surface. Detectors and explanations are tuned by Karbon Analytics. No configuration on your side.

Trust the explanation

Why this AI doesn’t make things up.

Operators get burned by AI that confidently invents metrics. Karbon Analytics is architected so that can’t happen: the signal is computed, the numbers are queried, the AI writes around facts it cannot edit.

[01]

Every number comes from the model

AI can only reference values queried from the unified data model. It cannot invent a metric, extrapolate a trend, or estimate a number that the model did not produce.

[02]

Detectors are deterministic, not generative

The signal itself (a ROAS drop, a stockout risk, a refund spike) is computed by rule-based detectors. AI only writes the explanation around a fact already established by code.

[03]

Contributors are cited, not implied

Explanations decompose movement by channel, cohort, SKU, campaign, and audience. The contributing values are inline. If the AI says a Klaviyo flow drove the lift, the flow and the number are named.

[04]

PII and business identifiers never reach the model

Customer PII (emails, names, addresses) and your business identifiers (brand name, store name) are removed before any AI processing. The language model sees aggregates and detector outputs, not your customers or your brand.

Out-of-the-box AI

All this,
without the prompt engineering.

Karbon Analytics ships with detectors, explanation templates, and recommendation logic already tuned for ecommerce. You skip the prompt iteration, the eval harness, and the in-house ML team that an analytics-grade AI normally takes to ship.

[01]

Prompts to write

  • System messages
  • Few-shot examples
  • Output schemas
  • Jailbreak guards

No prompt engineering

[02]

Datasets to prepare

  • Training data labels
  • Fine-tuning examples
  • RAG embeddings
  • Eval harnesses

No dataset prep

[03]

Schema to teach

  • Column descriptions
  • Metric definitions
  • Business glossary
  • Sample queries

Karbon Analytics knows the model

[04]

ML team to staff

  • Prompt engineer
  • ML engineer
  • Eval analyst
  • On-call rotation

Expertise included.

Tuned, hosted, and maintained by Karbon Analytics

By the numbers

Karbon Analytics already analyzes serious data. Built to handle the volume your store produces.

0
Shopify orders analyzed
0
GA4 sessions analyzed
0
Data points across the unified model
0
Meta and Google ads analyzed
0
SKUs across inventory data

Tomorrow morning

Wake up to answers. Not a blank dashboard.

Connect Shopify and your ad accounts. Tomorrow's brief lands in your inbox with the first signals, explanations, and actions from your data.

Start free trial Talk to us

FAQ

Frequently asked questions.

About accuracy, privacy, the models behind the explanations, and what you have to configure to make it work.

About accuracy

How does the AI avoid making up numbers?
Every number in an AI explanation comes from a query against the unified data model. The AI can only reference values the model produces; it cannot invent a metric or extrapolate a trend. Detectors are deterministic and rule-based, so the signal itself (a ROAS drop, a stockout risk, a refund spike) is computed before AI is involved. The AI writes the explanation around that fixed fact, not the other way around.
Can the AI explain why a metric moved?
Yes. When a metric moves, Karbon Analytics decomposes the change against the unified model (by channel, cohort, SKU, campaign, audience, geography, and other dimensions) before the AI writes. The explanation cites the specific contributors. For example, a 12% revenue lift attributed to two Klaviyo flows and one Meta ad set, with the contributing values inline.
What if the recommendation is wrong for my business?
Recommendations are suggestions, not commands. Every action card shows the reasoning, the contributing data, and the priority. You stay in control of what gets paused, scaled, or reordered. You can also dismiss recommendations, and the AI will weight that feedback for future signals on the same surface.

About privacy

Is my data sent to a third-party model?
Detection runs entirely on the Karbon Analytics infrastructure on top of the unified data model. Only the structured signal (the detector output, the contributing aggregates) and a sanitized prompt template reach the language model that writes the explanation. Customer PII (emails, names, addresses) and your business identifiers (brand name, store name) are anonymized or removed before any AI processing. The model never sees your customers or your brand.
What language model do you use?
Karbon Analytics routes explanations and recommendations through an enterprise-tier language model under a data processing agreement that prohibits training on customer inputs. Specific providers and models are selected per task for accuracy and cost, and may change over time. Customer data is never used to train any underlying model.

About setup

Do I have to configure the AI?
No. Karbon Analytics ships with detectors, explanation templates, and recommendation logic already tuned for ecommerce. There is no prompt engineering, no schema teaching, no fine-tuning, and no eval harness on your side. The AI runs the moment your sources are connected and the model has data to read.
Can I get explanations on dashboards, not just signals?
Yes. Any KPI or chart in the dashboards surface can be expanded for an AI explanation. The AI reads the same unified model that produced the chart, decomposes the movement, and writes the explanation inline. The morning brief is the prioritized version; on-demand explanations are the deep-dive version.
Does the AI work for non-Shopify stores?
The AI works against whatever data is in your unified data model. Today Karbon Analytics supports Shopify, Meta Ads, Google Ads, GA4, and Klaviyo. As new sources join the model, AI explanations cover them automatically with no additional prompting or configuration.

Still have questions?

Talk to the team

Continue exploring the platform

What the AI reads from

Layer 01

Unified Data Model

The model AI reads from. Sources mapped, customers linked, metrics defined once. Every explanation is grounded in the model, not in raw exports.

Layer 02

Daily Signals

The detection layer AI explains. 40+ rule-based detectors run nightly against the model and surface what changed, ranked by impact.