Every software vendor in manufacturing is pitching AI right now. Most of the pitches sound the same: put a chatbot on your website, let it answer customer questions, watch the leads roll in.
Here’s the problem: your website isn’t where the bottleneck is.
Your dealers are. They’re the ones configuring complex products across your catalog, calling your inside sales team at 4pm on a Friday to ask if a 5.5-inch housing supports a 270-inch width, and waiting 24 hours for a quote that should take 24 seconds. That’s where AI should be working — not on a marketing page answering “what are your business hours?”
This post is a practical framework for where AI agents actually create value in the dealer channel, what to build first, and what the technology needs to look like under the hood.
The Dealer Channel Bottleneck Nobody Talks About
Manufacturers with dealer networks share a common pain point that rarely shows up in software vendor pitch decks: dealer wait time.
A dealer configuring a motorized retractable screen for a commercial project hits a constraint they don’t understand. They call inside sales. Inside sales is on another call. The dealer leaves a message. Three hours later, someone calls back. By then, the dealer has moved on to quoting a competitor’s product.
This pattern repeats hundreds of times a week across every manufacturer with a complex product catalog:
- Configuration questions — “Can I pair this motor with that track type?” “What’s the max width for a recessed install?” These have definitive answers buried in the rule engine, but dealers can’t access them without a human intermediary.
- Quote turnaround — A dealer sends specs for a 15-unit order. It sits in a shared inbox until someone has time to key it into the CPQ system, validate the selections, and send back a PDF. Hours or days pass.
- Order status — “Where’s my order?” is still answered by email or phone at most manufacturers. The dealer has a portal login, but the status page requires navigating five screens and interpreting internal codes.
- Spec validation — A dealer submits a configuration that violates a constraint they didn’t know existed. The order gets bounced back. Another round trip. Another day lost.
These are high-frequency, low-complexity interactions. Each one is small. In aggregate, they’re the single biggest drag on dealer productivity and satisfaction — and the primary reason dealers favor manufacturers who are easier to work with over those with better products.
This is where AI agents belong.
What an AI Agent Actually Looks Like Here
Let’s be specific, because “AI agent” has become one of those terms that means everything and nothing.
We’re not talking about a chatbot widget bolted onto your dealer portal. We’re talking about a software entity that lives inside your existing workflow — assigned to support tickets, responding to dealer messages, querying your product rules, and generating draft quotes. It operates in the same system your team uses, not in a separate window.
The distinction matters architecturally. A chatbot is reactive: it waits for someone to type into a box. An agent is event-driven: it wakes up when something happens — a new support ticket, a configuration question, a stale quote that hasn’t been followed up on — and takes action.
Here’s what that looks like in practice:
A dealer submits a ticket: “Can I use the 4-inch roller with a 180-inch drop on the recessed install?” The agent receives the event, queries the configuration rule engine with those parameters, and responds within seconds: “Yes — the 4-inch roller supports drops up to 192 inches on recessed installs. Note that this combination requires the heavy-duty bottom bar, which adds $45 per unit.”
The dealer gets their answer at 11pm on a Tuesday. No one on your team was involved. The ticket is resolved and logged.
What makes this work isn’t just the language model. It’s the system around it:
- Event-driven activation — The agent doesn’t poll or idle. It responds to signals: a new ticket, a new message, a status change. This keeps costs predictable and behavior auditable.
- Goal-oriented execution — The agent isn’t just generating text. It’s working toward a defined outcome: resolve this ticket, generate this quote, validate this configuration. Every action is a step toward completing that goal.
- Persistent memory — The agent remembers that this dealer ordered 40 units of the same product last quarter, that they always use recessed installs, and that their region requires specific code compliance. Context accumulates over time, making every interaction faster and more relevant.
- Graceful escalation — When the agent’s confidence is low or the question requires judgment a machine shouldn’t make — pricing exceptions, custom engineering, warranty disputes — it escalates to a human with full context. The human picks up where the agent left off, not from scratch.
This isn’t theoretical. We’ve built these systems. The architecture patterns are production-tested. The challenge isn’t whether it works — it’s knowing where to start.
The Three-Tier Framework: Where to Start
Most manufacturers who try AI in the dealer channel fail because they start with the hardest problem. They want a fully autonomous agent that handles everything from configuration to order placement to proactive outreach. That’s Tier 3. You need to earn your way there.
Tier 1: Configuration Q&A
What it does: Answers dealer questions about product rules, constraints, compatibility, and specifications by querying your configuration engine directly.
Why start here: These questions have definitive, computable answers. There’s no ambiguity, no judgment call, no risk. The agent either returns the correct answer from the rule engine or says “I’m not sure — let me get a human.” The failure mode is a non-answer, not a wrong answer.
Volume and impact: At most manufacturers we’ve worked with, configuration questions account for 30-50% of inbound dealer inquiries. Resolving even half of these automatically frees up significant inside sales capacity.
What you need: A configuration rule engine with an API (not a locked-down UI), a ticket/messaging system where agents can receive and respond, and a language model that can translate natural language questions into rule engine queries and translate the results back into plain English.
Tier 2: Quote Draft Generation
What it does: Takes a dealer’s product specifications — submitted via a form, a ticket, or even a plain-text description — runs them through your CPQ system, validates the configuration, and generates a draft quote for human review.
Why it’s Tier 2: Quotes involve money. Even though the agent is generating a draft that a human reviews before sending, the stakes are higher than Q&A. You need confidence in Tier 1 accuracy before you hand agents the quoting workflow.
The key insight: The agent doesn’t replace your quoting process. It front-loads it. Instead of a human spending 20 minutes keying specs into the CPQ system, the agent does that in seconds. The human reviews, adjusts if needed, and sends. You’ve cut quote turnaround from hours to minutes.
What you need: Everything from Tier 1, plus integration with your CPQ pricing engine and a review/approval workflow so quotes don’t go out without human sign-off.
Tier 3: Proactive Agent
What it does: Monitors dealer activity and takes initiative. Follows up on quotes that haven’t been accepted in 7 days. Notices a dealer configuring the same product repeatedly and suggests a volume discount. Alerts a dealer when pricing changes affect their pending orders. Identifies dealers whose order volume is declining and flags them for outreach.
Why it’s Tier 3: Proactive behavior requires judgment, timing, and deep context. Get it wrong and you’re the annoying AI that emails dealers about things they don’t care about. Get it right and you’re the manufacturer that somehow always knows what the dealer needs before they ask.
What you need: Everything from Tiers 1 and 2, plus historical dealer data, configurable trigger rules, and careful tuning of when to act and when to stay quiet.
The rule: Start with Tier 1. Prove accuracy. Build trust. Expand to Tier 2 once dealers are relying on the Q&A agent. Only move to Tier 3 when you have enough interaction history to make proactive behavior genuinely useful rather than noisy.
What the Tech Stack Needs to Support
AI agents in a manufacturing context have requirements that most off-the-shelf AI platforms don’t account for. Here’s what matters:
Real-time response over persistent connections. Dealers expect the same responsiveness from an agent that they get from a phone call. If the agent takes 30 seconds to respond to a configuration question, they’ll call a human instead. WebSocket-based architectures — where the dealer and agent share a persistent connection — eliminate the request/response overhead that makes traditional API-based chatbots feel sluggish.
Process supervision and fault tolerance. Agents crash. Models hallucinate. API calls timeout. In manufacturing, a crashed agent can’t mean a lost order or a corrupted quote. The runtime needs to detect failures, restart agents, and recover state automatically — without human intervention and without affecting other dealers’ sessions. This is where the BEAM virtual machine (Erlang/Elixir) earns its reputation. The same process supervision model that telecom systems use to achieve 99.999% uptime applies directly to AI agent reliability.
Multi-tenant isolation. A single agent infrastructure serves multiple manufacturers, each with their own product rules, pricing, and dealer network. Manufacturer A’s pricing logic must never leak into Manufacturer B’s responses. This requires tenant-scoped data access at every layer — from the rule engine to the language model’s context window.
Budget controls and observability. Language model API calls cost money. An agent stuck in a reasoning loop can burn through budget in minutes. Production agent systems need per-agent budget caps, execution timeouts, and detailed logging of every action taken. You should be able to answer “what did this agent do, why, and how much did it cost?” for every interaction.
Memory that persists across sessions. The difference between a useful agent and a frustrating one is whether it remembers context. When a dealer asks about a product they configured last month, the agent should know. When a dealer always orders with the same specifications, the agent should default to them. This requires structured memory — not just chat history, but facts, preferences, and relationships stored and retrieved intelligently.
If your first blog post about our CPQ architecture showed how we bridge runtimes with Erlport and Poolboy for sub-80ms configuration responses, the agent layer is the natural evolution: the same BEAM process model, the same fault tolerance guarantees, the same real-time LiveView transport — now orchestrating AI agents instead of Ruby workers.
The Real Question
The question isn’t whether AI will transform the dealer channel. It will. The question is whether you’ll be the manufacturer who deploys it thoughtfully — starting with the right use case, building trust incrementally, and creating genuine value for your dealers — or the one who bolts a chatbot onto a marketing page and calls it innovation.
Start with what your dealers are waiting on right now. That wait time is your AI opportunity.
If you’re a manufacturer thinking about where AI fits in your dealer network, we should talk. We’ve spent 10 years building software for this exact industry — and the last two building the agent infrastructure to make it intelligent.