How NovaMart Deployed AI Agents Across Teams and Saved $449K/Year
NovaMart is a product marketplace that most people in e-commerce have heard of but few outside the industry would recognize. With roughly 300 employees, 200,000 active sellers, and over five million monthly buyers, they sit in that challenging middle ground: big enough that operational inefficiency costs real money, small enough that every headcount decision matters. When their VP of Engineering first reached out to us, the message was blunt: "We're spending more time compiling reports about our business than actually improving it."
That conversation turned into one of the most compelling Bonobot deployments we've seen — not because of a single dramatic transformation, but because it demonstrated something we've believed from the start: AI agents aren't a single-team tool. They're an organizational capability. NovaMart deployed Bonobot across two very different departments with very different problems, and the compounding effect of both deployments running simultaneously produced results neither team could have achieved alone.
The Ad Operations Problem Nobody Wanted to Talk About
NovaMart's Ad Operations team manages campaign performance across fifteen advertising channels. Google Ads, Meta, Amazon Sponsored Products, TikTok Shop, Pinterest, and ten other platforms that each have their own dashboard, their own metrics format, and their own reporting API. Every Monday, the team's four analysts would begin the ritual: pull data from each channel, normalize it into a common format, cross-reference against internal sales data, generate visualizations, write commentary, and compile everything into a weekly executive report.
The process took two full days. By the time leadership received the report on Wednesday afternoon, they were looking at data that was already a week old. Decisions about budget reallocation, channel optimization, and campaign strategy were consistently being made on stale information. The analysts knew it. Leadership knew it. But nobody had a better solution because the problem wasn't analytical skill — it was the sheer mechanical effort of pulling, normalizing, and synthesizing data from fifteen different sources.
"We had four brilliant analysts spending 40% of their week on data plumbing. They could tell you exactly which campaigns to kill and which to scale — but only after spending two days just getting the numbers into a spreadsheet."
“
The team had tried automation before. They'd built custom scripts, experimented with ETL pipelines, and even evaluated two dedicated ad analytics platforms. Each solution solved part of the problem but introduced its own maintenance burden. The scripts broke whenever a platform changed its API. The ETL pipeline required a dedicated engineer to maintain. The analytics platforms couldn't handle NovaMart's specific cross-channel attribution model. Every solution traded one form of manual work for another.
Bonobot's Fan-Out Architecture Changes the Game
The Bonobot deployment for Ad Operations used a pattern we call fan-out/fan-in coordination, and it's one of the most powerful multi-agent architectures available on the platform. Here's how it works in practice.
NovaMart configured a Coordinator agent as the primary interface for the Ad Operations team. When someone on the team says "Analyze all Q4 campaigns," the Coordinator doesn't try to do everything itself. Instead, it breaks the request into parallel subtasks and delegates each one to a specialized analyst agent using Bonobot's delegate_task capability. Five parallel delegate_task calls fire simultaneously: one agent analyzes Google Ads performance, another handles Meta campaigns, a third covers Amazon Sponsored Products, and so on.
Each analyst agent has access to the specific data it needs through Bonito's Resource Connectors. Campaign data lives in S3 buckets (read-only access, scoped to specific prefixes per channel), performance metrics are tracked in Google Sheets, and the agents can deliver results through the Slack connector. The tool policy for each agent is set to mode "selected" — only search_knowledge, query_data, generate_chart, and send_notification are enabled. No code execution. No arbitrary network access. No ability to modify campaign settings. The agents can analyze and report, but they cannot act on the advertising platforms themselves.
When all five analyst agents complete their work, the Coordinator agent uses collect_results to gather their outputs, synthesizes the findings into a unified executive report, and delivers it to the leadership Slack channel. The entire process — from request to delivered report — takes twelve minutes.
Twelve minutes. Not two days. The same depth of analysis, the same cross-channel comparisons, the same executive commentary — but delivered before the Monday standup ends instead of halfway through Wednesday. The Ad Operations team didn't lose their jobs. They stopped being data plumbers and started being strategists. The four analysts now spend their reclaimed time on work that actually requires human judgment: negotiating with channel partners, designing new campaign experiments, and building the attribution models that the agents use for their analysis.
The Seller Support Crisis
While Ad Operations was drowning in reports, NovaMart's Seller Support team was drowning in tickets. Two hundred thousand active sellers generate a staggering volume of questions, and the vast majority of them are repetitive: What are the listing policies for electronics? How does the fee structure work for international sellers? When is my next payout? What image specifications does the marketplace require? How do I set up promoted listings?
The support team of twelve agents (human agents, to be clear) was handling an average of 800 tickets per day. Average response time had crept up to four hours, and seller satisfaction scores were declining quarter over quarter. The team had built an FAQ section and a help center, but sellers either couldn't find the answers or didn't trust them enough to stop submitting tickets. The support team was caught in a vicious cycle: the more time they spent on repetitive questions, the less time they had for complex issues, which meant complex issues took even longer to resolve, which meant seller satisfaction dropped further.
"Our support team was answering the same twenty questions eight hundred times a day in slightly different words. Meanwhile, sellers with genuinely complex account issues were waiting days for help."
“
AI Context Turns Documentation Into a Living Resource
The Bonobot deployment for Seller Support centered on AI Context — Bonito's built-in RAG (Retrieval-Augmented Generation) engine. NovaMart's team uploaded forty-five documents into the AI Context knowledge base: the complete seller handbook, all fee schedule documentation, policy documents covering every product category, API integration guides, payout process documentation, and promotional tools guides. Bonito chunked these into 312 searchable segments, creating a knowledge layer that any agent on the platform could reference instantly.
The Frontline agent was configured as the first point of contact for all seller inquiries. When a seller asks "What are the fees for selling refurbished electronics?", the agent searches the AI Context knowledge base, finds the relevant chunks from the fee schedule and the refurbished goods policy, and synthesizes an accurate, sourced answer in about eight seconds. No hallucination, because the agent is grounding its responses in NovaMart's actual documentation. No outdated information, because when the policy team updates a document, the AI Context re-indexes automatically.
But the real sophistication comes from the multi-agent escalation pattern. Not every seller question can be answered from documentation alone. Account-specific issues — a missing payout, a listing that was incorrectly flagged, a dispute with a buyer — require access to NovaMart's internal systems. The Frontline agent is configured with a connection to a Specialist agent that has scoped read-only access to relevant database views. When the Frontline agent determines that a question requires account-specific data (through classification built into its system prompt), it escalates to the Specialist agent with full conversation context. The Specialist can look up the seller's account details, check transaction history, and provide a specific answer — or, if the issue requires human judgment, escalate to a human support agent with a complete summary of what's already been investigated.
The results speak for themselves. Seventy-eight percent of incoming tickets are now fully resolved by the Frontline agent without any human involvement. Average response time dropped from four hours to eight seconds. Seller satisfaction scores increased by thirty-four percent in the first quarter after deployment. And the twelve human support agents? They're now handling the twenty-two percent of tickets that actually require human judgment, with full context already assembled by the AI agents. Their resolution time for complex issues dropped by sixty percent because they're no longer context-switching between a complex account dispute and "how do I upload a product image."
The Cost Math That Made the CFO Smile
Let's talk numbers, because this is where the Bonito platform's smart routing really shines.
NovaMart processes approximately 25,000 agent interactions per day across both the Ad Operations and Seller Support deployments. In a typical enterprise AI deployment without intelligent routing, every one of those requests would go to a premium model — because that's the default, and because most platforms don't give you a meaningful way to route differently. At premium model pricing, NovaMart's monthly AI API cost would be approximately $45,000.
Bonito's cost-optimized routing changes that equation dramatically. Our production data shows that roughly sixty percent of NovaMart's interactions are straightforward queries that lightweight models handle perfectly — simple FAQ lookups, data retrieval, basic formatting. Twenty-five percent are medium-complexity tasks that mid-tier models handle well — multi-step analysis, document synthesis, nuanced classification. Only about fifteen percent are genuinely complex reasoning tasks that benefit from premium models — executive report synthesis, complex escalation decisions, multi-source analytical narratives.
With smart routing distributing traffic across model tiers based on actual task complexity, NovaMart's monthly API cost dropped to approximately $4,800. Add the Bonobot platform cost of $349 per month for each of their eight agents ($2,792 total), and their all-in monthly AI cost is $7,592.
Compare that to the $45,000 they'd be spending without Bonito's routing, and the net savings are $37,400 per month — $449,000 per year. That's a 12:1 ROI on the platform investment alone, before accounting for the operational savings.
And those operational savings are substantial. The Ad Operations team reclaimed the equivalent of two full-time employees' worth of time from report compilation — conservatively valued at $180,000 per year. The Seller Support team's ticket deflection reduced the need for additional support hiring that was already budgeted — a savings of approximately $120,000 per year. Combined with the API cost savings, NovaMart's total annual benefit from the Bonobot deployment exceeds $749,000.
Security That a Marketplace Demands
Running AI agents on a marketplace platform creates unique security requirements that go beyond typical enterprise concerns. NovaMart handles sensitive seller data including financial information, business identities, and transaction histories. Marketplace regulations require strict data isolation and comprehensive audit trails. A misconfigured agent that leaked one seller's data to another could be an extinction-level event for the business.
Bonobot's security architecture was designed for exactly this threat model. Every agent starts with zero permissions — the default-deny approach means that NovaMart's Seller Support Specialist agent, for example, can query specific database views related to the seller currently being helped, but cannot run arbitrary queries, cannot access other sellers' data outside of the active conversation scope, and cannot write or modify any records.
Each agent has a monthly budget cap of $500, enforced in real-time at the gateway level. If an agent approaches its budget — whether through legitimate heavy usage or through an attempted prompt injection attack designed to trigger expensive model calls — the system throttles or pauses the agent before the budget is exceeded. Rate limiting is set at 30 requests per minute per agent, providing an additional layer of protection against abuse.
Seller data isolation is enforced through S3 prefix scoping on Resource Connectors. The Ad Operations agents can access campaign data buckets but cannot see seller personal information. The Seller Support agents can access documentation and account views but cannot see raw campaign performance data. There is no lateral movement path between the two deployments, even though they run on the same Bonito control plane.
Every interaction generates a complete audit trail: which agent processed the request, which model was used, what data sources were accessed, what the response contained, and what it cost. NovaMart's compliance team can generate audit reports covering any time period, any agent, or any data source — a capability that proved invaluable during their most recent marketplace regulatory review.
"Our compliance team went from dreading AI audit requests to using them as proof points. Every agent interaction is traceable, every data access is logged, every dollar is accounted for. That's not just security — that's trust."
“
What NovaMart Learned (and What You Can Apply)
Six months into their Bonobot deployment, NovaMart's VP of Engineering shared several lessons that apply to any organization considering multi-agent AI deployments:
Start with the most painful manual process, not the most glamorous AI use case. NovaMart didn't start with a moonshot project. They started with the two workflows that were consuming the most human time for the least human-judgment-requiring work. The mundane starting point produced the most measurable ROI.
Multi-agent architectures compound in value. The fan-out pattern in Ad Operations and the escalation pattern in Seller Support are fundamentally different architectures solving fundamentally different problems. But running both on the same platform means shared knowledge, shared governance, and shared cost optimization. The AI Context knowledge base that powers Seller Support is now being referenced by Ad Operations agents when they need to understand fee structures that affect campaign profitability.
Budget caps are features, not restrictions. NovaMart's per-agent budget caps initially felt conservative. In practice, they've prevented three incidents where prompt injection attempts in seller support queries tried to trigger expensive recursive analysis loops. The caps caught what would have been costly runaway processes.
Measure time reclaimed, not just costs avoided. The $449K in API cost savings is the headline number. But the real transformation is 800+ hours per month of human time redirected from mechanical data work to strategic thinking. That's not a line item on a spreadsheet — it's a competitive advantage that compounds every quarter.
NovaMart's story isn't about replacing people with AI. It's about giving people their time back. The Ad Operations analysts are doing better work because they're not spending two days a week copying numbers between dashboards. The support agents are handling harder problems because they're not answering the same FAQ for the four hundredth time. The CFO is approving more ambitious AI projects because the first two delivered measurable ROI within sixty days.
If your teams are spending more time on data plumbing than data thinking, if your support team is answering the same questions hundreds of times a day, if your AI costs are growing faster than your AI value — that's exactly the problem Bonobot was built to solve.