Four expressions of the same command-center substrate. Same engine, four contexts: the bench, the field, the phone, and the web.
The CEO's request, restated precisely.
That request contains four separable problems. They have wildly different difficulty and they must not be sold as one thing. Here they are, in order.
The device is powered and connected. The bench captures everything it says on the bus. The system identifies the device, retrieves everything Ag Express knows about that device class and its failure modes, and walks the technician through a guided diagnostic derived from how this exact symptom was resolved the last forty times.
The verdict comes with the evidence and a confidence score.
Record the bus traffic a known-good machine produces around a specific device, once, in the real world. Replay that recording at the device on the bench. The device believes it is installed. Compare its responses against a known-good reference. Deviations localize the fault.
Honest name: stimulus replay, not simulation. Scales linearly with capture effort, not with engineering effort. One good machine and one good unit seeds each device family.
A model of the surrounding machine runs in real time and responds to what the device does. The loop is now closed. Analog and discrete IO drive sensor-level signals and read outputs beyond the CAN bus. Fault insertion confirms the device handles opens, shorts, and out-of-range conditions correctly.
Real engineering, per device family. The gate: fund Rung 3 for the top five device families by repair volume, prove it, then decide. Not before.
Not "this ECU is bad" but "the high-side driver on the section 3 output is bad." This is where board-level repair economics actually live. Combining behavioral modeling, signature analysis over a large labeled dataset of known-bad boards, and possibly thermal or electrical imaging.
Speculative today. Achievable over years, given the dataset that Rungs 1 through 3 generate. This is the destination, not a deliverable.
The interface layer (CAN bus hardware, programmable power supply, bench compute) is commodity. Any competitor can buy the same components tomorrow.
The connector adapters are different. Every competing bench tool fails at the connector. Ag Express is a harness manufacturer with an ISO 9001:2015 certified production division. You can build a validated adapter for every device you service, in house, at cost. That is a physical barrier to entry that no software company can replicate.
The adapter set is a product you can manufacture better than anyone else competing with you, and you already have the factory for it.
Feature BENCH-18: every verdict requires technician signature with a retained evidence trail. Calibrated confidence language, reviewed by counsel before the first customer sees a result. We raise liability before we raise the invoice.
When a diagnosis turns out to be wrong, what happens today? That question shapes the confidence design. We are building the answer to it into the tool from the first day it is used.
Same engine, different context. Fundamentally different constraint.
A loose component sits on the bench with no machine around it. Connection is direct to the component's bus pins. The simulation makes the component believe a machine is there.
An assembled machine sits in the field. Connection is through the machine's existing diagnostic port. Access is to the machine's CAN network, not to the component directly.
The field tool is not a port of the bench tool. It is closer to what Jaltest AGV already does, plus the corpus. We position it that way, honestly, because the truth is better than the alternative: a bench technician using a field tool in the wrong context and trusting it.
A tablet that needs connectivity is useless in a field in Nebraska. That means a local model, a cached corpus subset relevant to the technician's expected work, and reconciliation with the command center when connectivity returns. This is an extra-large feature and it is not optional.
The moment a dealer's technician uses the tool, dealer data isolation, seat licensing, and entitlement enforcement all become required. These are not features to add later. Do not build them until the internal tool is proven and the commercial model is settled. Premature multi-tenancy is how simple tools become complicated products that nobody adopts.
You named this sequence yourself in our conversation. It is exactly right, and it shapes the build order.
Your own field teams use it first. The corpus is proven, the bench is running, and your technicians know the tool before anyone outside the company touches it.
Dealer technicians access the tool under existing dealer relationships. Data isolation per dealer organization. This is where licensing and metering become real requirements.
After dealer success is demonstrated and the commercial model is settled. Not before. A farmer who contacts your call center with a field problem is a different engagement than a trained dealer technician.
Jaltest AGV does multi-brand agricultural diagnostics well. It talks to assembled machines through the diagnostic port. Fault codes, live data, guided troubleshooting. It does this for dozens of ag brands.
It does not have 34 years of Ag Express failure knowledge behind it. It does not know how the Case IH FM-750 display fails at 8,000 hours in a wet season in Iowa. It does not know which failures look like a bad module but are actually a pinout issue on the harness connector. Jaltest knows the machine. Ag Express knows the failure.
The field tool with the corpus behind it is a different product from Jaltest with a better interface. The corpus is why it is different, and the corpus is why nobody else can build it.
Three months of overwhelming volume. Nine months of trough. A fixed-cost call center sized somewhere unhappy in between.
The problem is not that the calls are hard to answer. The problem is that the staffing required to handle peak cannot be justified against what the trough actually needs. The goal is to absorb the peak without carrying the trough.
Real-time transcription. Repair intake, parts inquiry, status check, technical question; classified before a human hears it.
An intake agent captures device, symptom, machine, and customer. Opens a ticket. The same structured information a technician would need to pull the unit off the shelf and start.
"Where is my unit" answered directly from the ticket system without a human. No hold time. No transfer. Accurate to the current status of that specific unit.
Parts questions answered from the live store, not a static script. Pricing, availability, and shipping options current to the moment of the call.
Any caller can reach a person. The handoff carries everything the caller already said so they never repeat themselves. The person picks up with context, not a blank slate.
Concurrent call handling scales to meet harvest volume. Steps down in the off-season. The staffing model changes shape rather than size.
If the corpus is not sufficiently built by the time next harvest starts, a commodity voice vendor in front of our intake node is the right answer for one season. We are stating this out loud now because it is true.
Any vendor who does not say it is pretending the schedule is different than it is. The goal is a corpus-powered intake that improves every season. The path there goes through a season where it is listening, not leading.
Workstream 3 of Phase 2 includes voice transcription and intent classification in a listening-only mode. No caller speaks to an agent until the corpus has proven it can answer correctly.
Multi-state call recording consent and automated agent disclosure, applied at call open, with per-state policy covering Iowa, Nebraska, Indiana, and all states in the caller's footprint. This is not optional and it is not something to revisit after launch. It ships with Phase 3, before any caller speaks to an agent.
The extension story needs a first example that is concrete rather than visionary. We have one.
Your repair catalog on store.agexpress.com has a routing fault that returns a redirect loop. Your 2026 spring product catalog is still under construction. Both are visible today, from the outside, on the site where you sell repair services as e-commerce SKUs.
The catalog should be generated from the same knowledge graph that holds the parts, the repair services, the machines they fit, and the failure modes they address. Product copy stops being something somebody writes and becomes a projection of what the company already knows. The same graph feeds the store, the campaign pages on solutions.agexpress.com, and the dealer communications.
Fixing the catalog is the easiest first node. It is already visibly broken. It is the most concrete, non-speculative demonstration we can offer, and it requires almost no explanation.
Agent-assisted production of trade show, email, and campaign assets. Brand tokens produced against the graph, not against a creative brief written from scratch before every Commodity Classic.
A self-service portal where dealers query the corpus before calling. Every dealer interaction feeds the graph. Dealer count, territory, and the nature of the agreement directly sizes what this becomes.
Repair queue orchestration across four sites. Turnaround forecasting that flags units at risk of missing the published 3-5 day window. Parts demand signal from inbound repair mix, before the part runs out.
Harness quoting and change control extended from the same corpus. The ontology was designed so "ag" is a configuration, not an assumption, because your own stated expansion path is into construction and public works.
New technician onboarding against graded scenarios drawn from real cases in the corpus. The 250 combined years on your current team are the curriculum. This is what knowledge transfer actually looks like at the bench level.
Employee mailboxes as governed nodes contributing signal and receiving assistance. Last node, not first. Trust is earned by the bench and the call center before anyone's inbox is in scope. This is a list we hand you, not a thing we turn on.