A restaurant can be open, its kitchen busy, its ingredients ready, and still disappear. On a delivery app, the little switch says closed. No customer queues outside to complain. The orders simply go somewhere else.
That was among the blind spots UrbanPiper faced across its restaurant technology network. Stores appeared offline while physically operating; available dishes appeared unavailable. Problems reached the team through complaints or end-of-day checks. Bicycle AI connected to existing Kafka streams and webhook APIs to watch availability and order journeys. Its case study reports resolution moving from days to hours.
- Bicycle watches revenue KPIs across retail, payments, and travel.
- It investigates business and technical causes, then recommends a controlled next step.
- The useful lesson: give an agent a specific job, and give a person responsibility for its answer.
The intriguing thing about this failure is how ordinary it looks. The kitchen does not need a new stove. The business needs to know that its digital front door has shut.
A number needs an owner
Bicycle sells software for that kind of mismatch. A retailer might assign an agent to checkout conversion; a payments team, approval rates; a travel operator, search-to-book conversion. The agent watches the number across relevant segments, investigates movement, and brings back evidence and a recommended response. Alerts handle the immediate problem. Chat supports follow-up questions. Dashboards and data stories organize the findings.
Its position sits between business intelligence, anomaly detection, and operational troubleshooting. Bicycle works above an existing data stack. Its distinction is the proposed chain of responsibility: notice the movement, investigate across systems, find the owner, check the result. A warehouse may contain the failed transactions while a gateway log contains the event that explains them.
The company’s October article about building this with Claude makes a useful division of labor. People approve what a metric means. Statistical methods detect unusual movement. A language model helps investigate explanations. Rules constrain the action, and comparison checks the outcome. That is a more demanding product than putting a chat box beside a chart. It also gives a buyer separate parts to inspect.
That makes the alternatives surprisingly broad: another analyst investigation, a Tableau or Looker workflow, warehouse-native AI, or an internal monitoring system. Bicycle argues that it packages the connective work. The comparison depends on what a customer already has and how much maintenance its team can absorb.
Two days of detective work, compressed
Consider Billtrust, the accounts-receivable and payments software company. Settlement rates vary with card networks, transaction attributes, and data completeness. A small change can matter across many transactions. Bicycle’s deployment analyzes settlement behavior, identifies affected merchants, and sends prioritized alerts with supporting dashboards.
The published case reports investigation time falling from two or three days to under four hours, with a two-person team monitoring more than 1,000 merchants daily. Those are company-reported results. They describe a particular workflow, rather than a promise that every customer gets the same improvement.
Groceries offer a different version of the problem. Bicycle’s customer materials put bigbasket’s availability for its top 1,000 products at roughly 80% before and 95% after. The detailed retail case lists causes including skipped replenishment requests, incorrect base quantities, and delayed goods-receipt records. An empty shelf has several possible explanations. Picking the right one determines who can help.
The disappearing hour
The demo won. The deal stalled.
There is a useful admission behind Bicycle’s current pitch. In a July 2026 interview on Experiencing Data, CEO and co-founder Bhaskar Sunkara described early enthusiasm that failed to reach a deal. The company had concentrated on business users while giving analytics teams too little reason to trust and administer the product.
Setup was another obstacle. Customers had to supply KPI definitions and mappings before the automated analyst became useful. Bicycle responded with more industry context and narrower agents tied to specific outcomes. The buyer wanted revenue impact; the data team needed defensible definitions and evidence. Both had to come along.
“Wherever you can be deterministic, be deterministic.”
Today, that lesson appears in reviewable drafts, controlled publishing, and customer-defined action rules. An alert and a change to payment routing carry different consequences. The person accountable for the second needs a clear scope, approval process, and way back.
Fifteen minutes is the beginning
Bicycle’s self-service entry, Vibe Analytics, proposes one KPI and one agent over two weeks. Its Snowflake and BigQuery pages advertise initial setup in about 10 to 15 minutes, with scoped, read-only queries running in the customer’s warehouse. Teams can inspect and edit the inferred mappings and thresholds.
The economics include warehouse queries. Snowflake costs depend on warehouse uptime and configuration; BigQuery’s page flags scan costs. Narrower table scope and sensible refresh schedules matter. The free trial requires no credit card. For a buyer, the sensible calculation includes software, infrastructure, setup effort, and time spent reviewing findings.
The trial’s published milestones make adoption part of the experiment. Data teams validate definitions and evidence; business users confirm the language. By the end, someone should have subscribed to a useful alert or story. A working connector alone does not establish that the software has improved a decision.
Freshness is a practical constraint. Bicycle’s own setup guidance warns against analyzing a half-loaded table. Missing gateway or inventory evidence also limits an explanation. Teams outside its supported transaction-heavy sectors have less reason to assume the defaults will fit.
The approach is easy to borrow: select one consequential metric, define it, connect the evidence, name the owner, and measure the response. Count false alarms and unsuccessful fixes too. The test is whether someone caught a meaningful problem sooner and made a supported decision. Somewhere, a perfectly good kitchen is waiting for its front door to reopen.