US Server Reseller

The company sells enterprise server infrastructure across the US market. Twelve account managers, three presales engineers, a catalog of 3,400+ SKUs, and hundreds of compatibility constraints between components.

The problem wasn't the sales team. The problem was the nature of the product. Customers don't come in with specifications — they come in with use cases. "We need to run ~30 VMs, NVMe storage, HA setup, budget's around $40K" Translating that into a validated server configuration requires knowing which CPU pairs with which memory, what the power budget allows, whether the chosen RAID controller fits the available slots, and a dozen other constraints.

Account managers couldn't do that without an engineer. Engineers were in back-to-back meetings. The average time from customer inquiry to first validated quote: 1 to 2 days. In a market where competitors quote overnight, that gap cost deals.

Eight months before working with us, the team built an internal RAG tool. They indexed product datasheets, compatibility PDFs, and spec tables. Managers could query it in natural language.

It helped — roughly 20% reduction in back-and-forth. But it didn't solve the core problem. RAG works by similarity matching: the system returns what looks like the right answer based on indexed content. For a 3,400-SKU catalog with evolving compatibility matrices, "looks like" wasn't good enough.

About one in four configurations contained an error that an engineer caught during review. The review step couldn't be removed. Engineers were still in the loop on every quote. The RAG tool became a reference assistant, not a configuration engine.

The root issue with RAG was architectural: it was guessing from documents instead of validating against data.

We replaced the RAG layer with Co Sales AI Configurator dual-agent architecture, connected directly to the company's PostgreSQL product catalog via a secure MCP bridge. No document retrieval, no similarity scoring — every component selection validated in real time against the actual database.

The Interviewer Agent handles the conversation side. When a manager pastes in a customer requirement or opens a new inquiry, the agent conducts a structured scoping session: workload type, virtualization platform, redundancy requirements, rack constraints, budget, delivery timeline. It adapts to context — a bare-metal request gets different questions than a virtualization cluster inquiry. No rigid forms, no decision trees.

The Engineer Agent handles the technical side. It runs multi-step function calls against the PostgreSQL catalog, checks component compatibility mathematically, validates power budgets, confirms slot availability, and assembles a complete Bill of Materials. Every constraint is explicit and traceable. No hallucinations — the output is either a validated configuration or a flagged edge case.

Integrations: PostgreSQL internal catalog (3,400+ SKUs, compatibility matrices, power and slot constraints), Salesforce CRM where completed configurations auto-create as Quote objects with full BOM attached, email parallel quote notification to the account manager, and Slack routing for edge cases and non-standard requests — sent to presales engineers with full conversation context already captured.

Implementation timeline: 5 weeks from kick-off to production — 2 weeks on data audit and catalog structuring, 1.5 weeks on validation logic, 1 week on Salesforce integration and testing.

Co Sales AI Configurator is deployed as an internal tool — account managers use it, end customers don't interact with it directly.

A manager receives a customer inquiry. They open Co Sales AI Configurator, describe the requirement in plain language, or paste the customer's message directly. The Interviewer Agent runs a focused scoping session — typically 6 to 10 questions, 4 to 8 minutes. The Engineer Agent queries the catalog, validates compatibility, and returns a complete, accurate configuration with component-level reasoning: why each part was selected, what constraint it satisfies, what alternatives exist within budget.

The configuration lands in Salesforce as a draft Quote with the BOM attached. The manager reviews, adjusts if needed, and sends to the client. If the request falls outside standard catalog logic — unusual form factor, custom power requirements, multi-site deployment — the system flags it and routes it to a presales engineer via Slack with the full conversation context already captured.

Presales engineers now engage on genuinely complex cases only.

Quote cycle time dropped from 1–2 days to 18 minutes average for standard configurations. Account managers handle the full quoting process independently without waiting for engineer availability.

Configuration accuracy on first pass went from 76% to 100%. The one-in-four error rate that required mandatory engineer review dropped to zero for standard catalog configurations. The engineer review step was eliminated from the standard workflow.

Presales engineering time freed: approximately 22 hours per week across the three-engineer team. That capacity now goes to complex deals, technical evaluations, and customer onboarding — not answering "will this RAM work with that CPU"

Quote volume capacity increased by 340%. Before launch, the team averaged roughly 8–10 validated quotes per week — the ceiling set by engineer availability. With the bottleneck removed, the team can respond to significantly more inquiries without adding headcount. In the first two months post-launch, they processed 3.4× that volume per week.

Salesforce data quality improved as a side effect. Previously, quote objects were created manually with inconsistent BOM detail. Auto-generated configurations standardized the data structure and reduced CRM entry errors.

The failure mode of RAG in technical sales isn't inaccuracy — it's undetectable inaccuracy. A RAG system that's wrong 25% of the time looks functional until you run the engineering review. You only see the problem when someone checks. That's why the review step couldn't be removed: it was the only quality gate.

Direct database validation changes the trust model. When account managers know the system is querying the actual catalog with actual constraints — not retrieving similar-looking documents — they stop treating the output as a suggestion and start treating it as a configuration. That shift in trust is what makes the engineer review step optional rather than mandatory.

The most valuable integration wasn't CRM — it was Slack routing for edge cases. The system needed a clear fallback: when it can't validate, it doesn't guess, it escalates with context. That design decision was what gave the sales team confidence to use it on real quotes from day one.