A sales rep gets an RFQ from a long-standing account. They open the ERP, cross-reference pricing tiers, check stock across two warehouses, pull up the customer's contract rates, and manually build a quote in a spreadsheet. Ninety minutes later, the quote goes out. By then, the buyer has already requested a competing bid.
That cycle — slow, manual, and error-prone — is the problem. An AI assistant connected to your ERP can compress it to five minutes. But the gap between "AI can do this" and "AI is doing this reliably for us" is where most projects fail.
Why manual quoting breaks down
The core issue is that B2B quoting in manufacturing and distribution is a multi-system process that doesn't scale. Every RFQ requires a rep to:
- Look up customer-specific contract rates in the ERP
- Check real-time inventory across warehouse locations
- Validate product configurations against the catalog
- Apply volume-tier discounts and regional pricing
- Build the quote in a format the buyer expects
Each step touches a different system. Adobe Commerce holds the product catalog. Epicor P21, NetSuite, or SAP Business One holds pricing rules and inventory truth. The rep bridges the gap manually, and every bridge is a place where errors enter.
The friction compounds when systems don't sync. Pricing pulled from an overnight batch may be stale by morning. Customer-specific discounts that live in a senior rep's head don't transfer to a junior rep handling the same account. Out-of-stock items get quoted because nobody checked inventory before building the line items.
What AI changes — and what it doesn't
An AI assistant connected to your ERP does the data retrieval and assembly automatically. The buyer submits a request through your portal, and the system:
- Parses the line items and matches them to SKUs
- Pulls real-time pricing from the ERP (not a nightly sync)
- Checks inventory across all warehouse locations
- Applies customer-specific contract rates and volume discounts
- Generates a formatted quote ready for review
For standard requests — established products, known customers, in-stock items — this happens without rep involvement. The rep's job shifts from data entry to exception handling: orders that need credit review, custom configurations, new customers without pricing history.
This is where the 90-minute-to-5-minute claim comes from. It's not that AI replaces the rep. It's that AI eliminates the 80 minutes of system-hopping that the rep currently does before they can even think about the quote.
The data problem nobody talks about
The hardest part of AI RFQ automation isn't the AI. It's the data preparation.
Your product identifiers must match exactly between Adobe Commerce and your ERP. If SKU-123 in the catalog maps to PART-456 in Epicor P21, the AI will generate quotes with conflicting information. Run a reconciliation report and fix mismatches before you start.
Customer records need the same treatment. Duplicate accounts, orphaned records, and mismatched pricing tier assignments cause the AI to pull incorrect contract rates. If a customer has two accounts in the ERP — one from a merger, one current — the AI doesn't know which one to use unless you tell it.
Pricing logic must be machine-readable. Contract rates, volume breaks, promotional pricing, and regional adjustments need to exist in structured form in your ERP. If pricing logic lives in a spreadsheet or a rep's memory, formalize it first. The AI can't apply rules that don't exist in a system it can read.
Real-time inventory feeds are non-negotiable. The AI needs current stock levels to generate accurate lead times. If your ERP pushes inventory updates once per night, you'll need to upgrade to near-real-time for the RFQ workflow. Fifteen-minute intervals work for high-velocity SKUs.
The approval workflow you need to design
Not every quote should be fully automated. The split between "AI handles it" and "AI drafts, human approves" depends on your risk tolerance and deal complexity.
Fully automated path — Standard products, established customers, orders under a dollar threshold. The AI generates the quote and sends it directly to the buyer. Target: under five minutes from submission to delivery.
Assisted path — Complex configurations, large orders, customers near their credit limit, or requests involving items with long lead times. The AI assembles all the data and drafts the quote. The rep reviews, adjusts, and sends. This is where you cut 90 minutes to 10-15 minutes of focused review.
Escalation triggers — Orders above a certain dollar amount, new customers, products with configuration dependencies, or requests involving items not in the standard catalog. These route to a manager or product specialist before anything goes to the buyer.
Map these paths before you build. Define who approves what, what thresholds trigger escalation, and what data the AI needs to surface for human review. The approval workflow is the difference between AI that accelerates your process and AI that creates new problems.
Implementation: three workstreams
Deploying AI RFQ automation on Adobe Commerce with an ERP typically involves three parallel workstreams:
AI configuration — Connect the assistant to your product database, pricing engine, and inventory feeds. The ERP is the single source of truth: the AI always defers to ERP data for pricing, credit limits, and stock levels rather than maintaining its own copy.
Integration layer — API connections between Adobe Commerce and your ERP for real-time data exchange. Webhook triggers that fire when a new RFQ arrives. Response templates that match your brand's quoting format.
Workflow configuration — Approval chains, exception rules, and escalation paths. Who can approve AI-generated quotes, who can override pricing, and who can only view. Map these to your existing role structure.
What the VP of Sales owns: approval thresholds, pilot product categories, and team communication. What IT owns: API configuration, data mapping, testing, and security. What both own jointly: user acceptance testing with real RFQs.
Measure before and after
Without baseline metrics, you can't prove ROI. Before deploying anything, capture:
- Current quote turnaround time per RFQ
- Pricing error rate and rework frequency
- Rep hours spent on quoting per day
- Quote-to-order conversion rate
- Volume of RFQs by complexity (simple vs. complex)
After deployment, track the same metrics. A realistic benchmark: 75% faster quote workflows, with rep time on manual quoting reduced by 6+ hours per day per quote. Pricing errors should drop by 80% or more when the AI enforces contract rates directly from the ERP.
The common failure modes
Skipping process mapping. Teams that jump straight to tool selection automate a broken workflow. Spend the first month understanding what actually happens today.
Treating AI as a standalone tool. An AI assistant that isn't connected to your ERP is a fancy search bar. Without real-time pricing, inventory, and customer data, it generates quotes nobody will trust.
Trying to automate 100% of RFQs on day one. Complex configurations, custom fabrication, and new-customer quotes need human judgment. Design for a realistic split, not a fantasy.
Confusing UI with automation. Adding a quote request form to your storefront is a surface-level change. Real automation means the system generates the quote, not just collects the request.
Ignoring the sales team. Reps who feel replaced will resist the tool. Position AI as eliminating their least favorite tasks — data entry, pricing lookups — so they spend more time selling.
How this gets solved
The checks above — pricing accuracy, inventory validation, customer-specific terms — are all data retrieval problems. The AI solves them by pulling from the ERP in real time rather than relying on a rep to bridge the gap manually.
But the approval workflow and exception handling require a system that tracks what happened, not just what the AI generated. RFQTrace timestamps when the request arrived, when the quote went out, and who approved it — so turnaround is measured rather than remembered, and pricing decisions have an audit trail.
The old spreadsheet stays exactly where it is, readable, for as long as you want it. Nothing about starting requires you to have finished with it.