Make Flowcart identify the next move

Flowcart Opportunity Radar is an AI-powered growth layer that analyzes shopper conversations, commerce events, and Flowcart performance; identifies a specific missed opportunity; estimates its impact; and recommends the next flow or action a merchant should activate.

This is deliberately not another shopper-facing AI feature.

The problem

Flowcart can keep shipping powerful flows—abandoned cart, browse recovery, winback, reorder, upsell, back-in-stock, referrals, post-purchase, reviews, search, and checkout. But every new capability creates a second problem:

A merchant should not need to understand Flowcart's feature architecture to get value from Flowcart.

A second card might explain that 186 conversations in the last 30 days contained stock-related intent, 63 of those customers asked about products that have since restocked, and Back in Stock is the recommended action.

AI is moving from doing work to finding the work

The market pattern is not merely AI executing a workflow. It is AI observing behavior and helping a business decide which workflow should exist.

The merchant wants a growth decision, not a flow builder

Primary user: ecommerce growth or CRM manager.
Secondary users: founder, ecommerce manager, retention manager.

Current behavior — a hypothesis to validate

The second and third groups are the opportunity.

That is materially different from: “Help me configure a WhatsApp flow.”

Connect what happened, why, and what to do

Flowcart can potentially observe two unusually valuable forms of evidence:

Most ecommerce analytics products answer what happened. Conversation data can help explain why it may have happened. Flowcart can connect the insight to what should we do about it.

1. Observe
2. Understand
3. Recommend
4. Activate
5. Measure
6. Learn

Start where an action already exists

Do not launch with 50 opportunities. Begin with signals for which Flowcart already has a credible action.

Do not start with AI

The first version does not require an LLM to discover most opportunities. Use deterministic logic where structured data already contains the truth.

# Reorder opportunity
customers_eligible_for_reorder > X
AND repeat_purchase_interval_confidence > Y
AND predictive_reorder_active = false
→ surface opportunity

# Cart recovery opportunity
abandoned_carts > threshold
AND abandoned_cart_flow_active = false
→ recommend Abandoned Cart

AI becomes useful for unstructured signals: classifying conversations into stock requests, price objections, delivery concerns, payment failures, size questions, recommendation requests, comparisons, and discount requests—then aggregating those patterns.

Your opportunities, on Home

The third card should not force “Activate feature X.” Sometimes the correct product behavior is: Something is wrong. Investigate. That restraint builds trust.

Clicking an opportunity

Do not immediately drop the merchant into Flow Builder. First show evidence: 312 previous customers approaching the reorder window; 221 reachable on WhatsApp; a 47-day median reorder interval; an 18.4% current repeat purchase rate; and a 221-customer potential audience.

AI products need inspectability. The merchant should understand: Why am I seeing this? What evidence produced it? What happens if I approve it?

Compress setup into a decision

After an opportunity is detected and reviewed, Flowcart can eventually pre-generate:

• Audience
• Trigger
• Timing
• WhatsApp message
• Recommended products
• Success metric

The merchant edits if needed, then activates. A 20-minute setup becomes a 60-second decision. That is a meaningful activation hypothesis—not merely a nicer interface.

Onboarding becomes a business diagnosis

Instead of asking “Which flows would you like to activate?”, Flowcart connects Shopify and analyzes the previous 90 days.

Setup is not value

Many SaaS products define activation as “User completed setup.” That is weak.

An even better definition is: Merchant receives measurable incremental value from an Opportunity Radar recommendation.

Prove opportunity identification is the bottleneck

Telemetry to inspect

• Flows available versus activated per merchant
• Time between integration and first active flow
• Time to first attributable order
• Features activated after 30, 60, and 90 days
• Relationship between feature adoption and retention
• Flow-configuration abandonment and feature-discovery paths
• Dormant merchant behavior

If not, do not build this. The real bottleneck may be setup complexity, WhatsApp approvals, lack of trust, poor ROI, or unclear operational ownership.

Five assumptions can kill the idea

Trust beats cleverness

Do not repeatedly cry opportunity

The dangerous failure is not a slightly inaccurate recommendation. It is training merchants to ignore every recommendation.

• Every login claims a giant revenue opportunity and creates dashboard fatigue.
• Flowcart optimizes merchant activity rather than merchant value.
• The same merchant receives contradictory recommendations.
• A recommended flow cannot run because prerequisites are missing.
• Opportunity estimates are mistaken for incremental revenue.

Three deterministic opportunities

On Flowcart Home, each card follows: signal → evidence → recommended flow → preview → activate.

Deliberately out of V1

• Autonomous flow activation or multi-step agents
• Generative campaign strategy or AI-generated experimentation
• Conversational analytics and cross-merchant benchmarks
• Opportunity revenue prediction
• Automatic budget allocation, discount optimization

Does Radar change behavior and time to value?

Eligible cohort

Merchants with Shopify connected, at least 90 days of order history, a sufficient eligible audience, and at least one recommended flow inactive. Randomize at merchant level.

Primary experiment metric: recommended-flow activation rate within 14 days. This is only a leading metric.

Real success metric: incremental revenue generated by Radar-activated flows per eligible merchant. Activating useless flows is not success.

Follow value beyond the click

Opportunity Value Adoption Rate

merchants with ≥1 value-producing Radar recommendation
────────────────────────────────────────────────────────────
eligible merchants

Version the recommendation, not only the event

At minimum, instrument the entire path from detection to value:

# Lifecycle events
opportunity_detected
opportunity_impression
opportunity_opened
opportunity_evidence_viewed
opportunity_dismissed
opportunity_dismiss_reason
opportunity_flow_previewed
opportunity_accepted
recommended_flow_created
recommended_flow_activated
recommended_flow_paused
opportunity_order_attributed
opportunity_revenue_attributed

Every opportunity should carry:

# Required properties
merchant_id
opportunity_id
opportunity_type
recommendation_version
eligibility_rule_version
eligible_audience
evidence
estimated_value_type
flow_id
created_at
acted_at

The learning system may matter more than another flow

A competitor can copy Back in Stock, Winback, or Abandoned Cart. It is harder to copy a system that learns:

Over time Flowcart could develop an opportunity → intervention → outcome dataset. That is potentially valuable proprietary learning.

High leverage, medium-low confidence

Use directional scoring rather than pretend the team knows precise numbers.

The confidence score is the reason not to jump directly into a large AI build.

Manage the product system, not the AI feature

If opportunity discovery is the bottleneck, the elite PM builds the smallest system capable of changing behavior. They do not stop at recommendations generated or even flows activated. They follow the chain:

That is the difference between shipping an AI feature and managing a product system.

Do not confuse activation lift with product success

How does an agent earn the right to act?

Today’s opportunity is not another shopper-facing AI feature. It is a merchant trust and deployment problem: How can a merchant know Flowcart’s AI agent is safe and commercially useful before allowing it to talk to thousands of real customers?

This matters more as Flowcart agents move beyond answering questions into product recommendations, cart changes, discounts, checkout, payments, and post-purchase actions.

Primary user: the ecommerce, CRM, or support owner responsible for deploying a Flowcart agent. Their current mental model is likely: configure instructions → manually test a few conversations → enable → discover edge cases from real customers.

That workflow creates a nasty asymmetry. The merchant has perhaps 10 test conversations, while the agent may encounter thousands of permutations after launch.

The usual AI onboarding model is wrong

Traditional SaaS onboarding asks: Have you configured everything?

Agentic-product onboarding should ask:

This is especially relevant to WhatsApp because a bad web experience can be closed. A bad proactive or conversational WhatsApp interaction lands directly inside a customer relationship.

Choose realism deliberately

The interesting Flowcart product is ultimately Historical Replay + Shadow Mode, not merely a chatbot playground. The smallest credible starting point is Historical Replay.

Flowcart Agent Readiness Lab

A merchant configures an AI shopping or support agent. Instead of immediately showing Activate Agent, Flowcart shows:

The merchant selects a historical window—for example, the last 30 days. Flowcart samples real conversations across product discovery, inventory, discounts, checkout, payment failures, delivery, cancellations, refunds, and unknown or long-tail queries. PII and unnecessary customer details are stripped where appropriate.

The proposed agent processes each conversation offline. Crucially, action tools operate in simulation mode. Instead of actually executing a write action, evaluation records what the agent attempted and whether it was allowed:

# Simulated action
agent_attempt: cancel_order(48291)
preconditions: valid
expected_behavior: escalate
result: FAIL — unsafe action

The merchant receives an Agent Readiness report that diagnoses specific blockers and connects directly to a staged deployment decision.

A generic accuracy score is almost useless

Do not show only AI accuracy: 91%. Show capability-level readiness and consequence-aware risk.

This creates something far more useful than a generic quality score: an autonomy map.

Agent configuration is increasingly easy. Competitors can all provide prompts, knowledge bases, tools, models, and actions. The harder merchant question is: Can I trust this thing with my customers and my business?

1. Answer
2. Recommend
3. Modify cart
4. Discount
5. Checkout
6. Modify order
7. Refund

Flowcart should not treat this as one global “AI enabled” toggle. Each capability should earn autonomy independently.

Judge ambiguity with AI; enforce commerce truth with code

The readiness score should combine deterministic assertions + model graders + eventually human-labelled samples.

Turn failures into launch decisions

Rather than merely exposing the score, Flowcart says: Three things are blocking a safer launch.

The merchant fixes the configuration, re-runs failed tests, and the score becomes 94. Flowcart then recommends:

Start 10% rollout. Simulation must connect directly to deployment.

The shopper sees nothing during simulation

That is the point. When the agent graduates to live traffic, WhatsApp feels exactly like the intended Flowcart experience. For uncertain or high-risk cases, the agent says: “I want to make sure we handle this correctly. Let me bring in a support specialist.”

The customer should never see AI confidence = 63%. Internal uncertainty should translate into sensible product behavior.

Shadow Mode — V1.5

After offline validation, put the agent on live traffic without letting it respond. A customer messages: “My payment went through but the order isn’t showing.” The human or current workflow handles it normally. Meanwhile, Flowcart’s agent independently produces its intended response, tools it would call, action it would take, and escalation decision.

Flowcart later compares that with the actual outcome:

This is much more persuasive than a demo conversation.

Prevent a readiness score from becoming false confidence

Failure modes

The largest is false confidence: a shiny “97% Ready” label can cause merchants to trust the system too much. Other failures include unrealistic simulations, graders rewarding plausible language instead of correct commerce outcomes, historical datasets missing new edge cases, configuration changing without regression tests, testing only successful tool responses, and merchants optimizing the score rather than customer outcomes.

Start with Historical Replay, not live Shadow Mode

Choose merchants with meaningful AI-agent traffic or a significant agent rollout in preparation.

Keep initial scope to three capabilities with materially different risk profiles: Product discovery, Order support, and Checkout or payment support.

Deliberately deferred V2

Do not initially build automated prompt repair, autonomous policy rewriting, synthetic customer personas, continuous production shadowing, automatic rollback, merchant-to-merchant benchmarking, adversarial red teaming, or self-improving agents. Those extensions only become logical after merchants demonstrably use simulation results to change deployment decisions.

Find out whether trust is actually the bottleneck

Inspect Flowcart’s own operational evidence: time from agent setup to production; number of internal test conversations; support requests asking whether actions are safe; percentage of tools merchants enable; agent disablement after launch; corrections and escalations; merchant-requested QA cycles; and Flowcart team time spent manually testing.

If that is not true, this is probably not the highest-priority problem.

Test whether evidence changes deployment behavior

Measure safe autonomy, not simulations run

The Readiness Lab is valuable only if it helps Flowcart safely increase that number.

Version the agent and the evaluator

# Readiness lifecycle
simulation_run_started
simulation_case_executed
simulation_assertion_failed
failure_reviewed
configuration_changed_from_failure
regression_run_started
capability_approved_for_launch
capability_restricted
rollout_percentage_changed
agent_action_attempted
agent_action_blocked
agent_escalated
production_policy_violation

Every result must preserve:

# Version tuple
agent_version
prompt_config_version
tool_version
evaluator_version

This is not primarily a QA feature

If it works, the Readiness Lab becomes infrastructure that lets Flowcart safely ship more powerful agents faster.

Testing is not the product outcome

The elite PM measures the entire chain.

A high pass rate does not erase consequence

Sources reviewed

• Intercom Help — Deploy and test FinHistorical questions, testing, response inspection, and deployment practices.
• Intercom Help — Simulation testing for FinAI-customer multi-turn simulations, routing, attributes, actions, and explicit Pass/Fail criteria.
• Intercom Help — Building Fin ProceduresProcedure testing and mocked connector behavior.
• Gorgias Help Center — Set up and go live with AI AgentTest conversations, draft configuration, and action-safety guidance.
• Klaviyo — AI agents in marketingComposer workflow and human review before campaign publication.

Competitive product behavior is evidence of an emerging pattern, not proof that merchant trust is Flowcart’s highest-priority deployment bottleneck. The concierge experiment is designed to establish that.