Case study · AI-First SaaS · Devsphinx
Vertical AI · Transportation

AI Dispatch
Platform.

An AI-native Transportation Management System for small and medium US trucking carriers. Seven cooperating agents automate dispatching, load sourcing, rate negotiation, driver communication, tracking, and back-office operations — replacing $240,000/year in dispatch costs for a 10-truck carrier.

MVP v0.1 · In active development
550K+
US carriers — target market
$240K
Saved per 10-truck carrier/yr
$1.65B
Total addressable market
7
Cooperating AI agents
85%
Dispatch cost reduction
IndustryUS Trucking · Transportation
Product typeAI-native SaaS platform
Core modelClaude API (Haiku + Sonnet)
First customer targetJune 2026

The market opportunity

A $900B industry running
on 30-year-old processes.

The US trucking industry generates approximately $900 billion in annual revenue. The operational backbone — dispatching — has not meaningfully changed in 30 years. Carriers are not just open to automation; with -2.3% average operating margins in 2024, they need it to survive.

550K+
US carriers
(1–100 trucks)
99.6%
Of carriers under
100 trucks
-2.3%
Average operating
margin 2024
$1.65B
TAM (550K × $3K
ACV)

"The value pool isn't the software price. It's the $240K/year a 10-truck carrier is spending on dispatch services that can be reduced by 85%."


Competitive landscape

Why this problem hasn't
been solved yet.

The TMS market is crowded — DAT, McLeod, Tailwind, Samsara, Motive, AscendTMS, Rose Rocket, and dozens more. So why is the dispatch workflow still manual for 99% of small carriers? Every incumbent is structurally disincentivised to solve it.

IncumbentWhy they haven't solved it
McLeod / Tailwind (legacy TMS)Per-seat pricing model. Automating dispatcher tasks destroys their revenue. Structurally disincentivised to build this.
Samsara / Motive (telematics)Monetise hardware and safety compliance, not back-office labour. Different cost centre entirely.
DAT / Truckstop (load boards)Monetise load discovery, not dispatch workflow. No incentive to replace the dispatcher.
Uber Freight / Convoy modelDigital brokerage — takes the margin, doesn't reduce the carrier's cost. Convoy shut in 2023.
New AI entrants (HappyRobot, FleetWorks)Broker-side or dual-sided focus. Carrier workflow is under-served by all current entrants.

The structural gap: no product owned the carrier's financial system-of-record while also automating their daily dispatch workflow. That's the wedge.


System design

The dispatch decision tree.
Mapped, then automated.

A human dispatcher makes five sequential decisions every time a load comes in. We mapped every branch, every data input, and every possible outcome — then designed a specialist agent for each decision phase.

Phase 1 · Screening
Is this load worth evaluating?
Intake Agent
High automation
Broker reputation, equipment type, lane, timeline — pure rules
Phase 2 · Rate
Accept, negotiate, or reject?
Rate Analysis Agent
High automation
RPM, carrier cost basis, lane history, diesel price — deterministic
Phase 3 · Negotiate
What's the counter?
Negotiation Agent
Supervised
Lane rate history, broker flexibility, carrier utilisation
Phase 4 · Match
Which driver gets it?
Dispatch Agent
High automation
HOS, equipment, deadhead, home time, fairness — optimisation
Phase 5 · Confirm
Did the driver accept?
Driver Comms Agent
High automation
Driver reply, 30-min escalation logic — fully automatable

Agent architecture

Seven specialist agents.
One event-driven platform.

Each agent has a single, narrow responsibility. No agent calls another directly — they emit events and subscribe to events. This means any agent can be replaced, upgraded, or disabled without touching the others.

Agent 01
Intake Agent
High
Agent 02
Rate Analysis
High
Agent 03
Negotiation
Supervised
Agent 04
Dispatch
High
Agent 05
Driver Comms
High
Agent 06
Tracking
High
Agent 07
Back Office
Supervised
Click any agent above to see its role, inputs, outputs, and tools.

Agent design philosophy

Three principles govern every agent in the system — chosen to eliminate hallucination risk and make outputs auditable in a regulated industry.

Principle 01
Tool calls, not prose
Every agent returns a structured JSON tool call — never free text. Outputs are parseable, database-ready, and auditable. No regex. No hallucinations reaching brokers or drivers.
Principle 02
Temperature 0 for decisions, 0.3 for drafts
Load scoring, driver ranking, and reply classification are fully deterministic (temperature 0). Message drafts and counteroffer language use slight creativity (temperature 0.3).
Principle 03
Human approval gates at v1.0
Every agent action that communicates with a broker or commits a load requires a dispatcher click. Autonomy is earned with outcome data over 6–12 months, not assumed on day one.

"The goal is not to replace the dispatcher. The goal is to make one dispatcher as effective as three."


Technical decisions

Every decision chosen
for a specific reason.

Each technical choice was made to eliminate unnecessary cost and complexity at MVP scale while keeping the path clear to 1,000+ customers. Zero-overhead infrastructure. No Redis dependency. No separate auth provider.

LayerTechnologyWhy this choice
FrontendNext.js 14Server Components for fast initial load; Supabase realtime for instant board updates without polling
BackendNestJS 10Module-per-agent architecture; class-validator for DTO safety; scales to microservices if needed
DatabaseSupabase (PostgreSQL)Multi-tenant RLS, realtime subscriptions, and Auth in one service — no separate auth provider needed
JobsInngest50K runs/month free; step-level retry; dev server for local debugging; no Redis dependency at MVP
Inbound emailCloudflare Email RoutingFree, unlimited, unblockable; Worker forwards to NestJS webhook with shared secret
Outbound emailResend3K/month free; developer-first API; clean TypeScript SDK
AI agentsClaude API (Anthropic)Tool-calling for structured outputs; claude-haiku-4-5 for high-volume tasks (scoring, ranking); Sonnet for complex parsing
MessagingTwilio (WhatsApp + SMS)WhatsApp Business API with SMS fallback; TCPA-compliant consent and disclosure built in from day one
CachingSupabase cache tableTTL caching via expires_at field; nightly Inngest cron cleans stale rows; zero extra infrastructure

Competitive moat

The AI model is not the moat.
The data is.

The underlying models commoditize any thin wrapper. The durable competitive advantage is built at two layers — both of which compound over time and are impossible to replicate without operating history.

Moat 01 · Lock-in
System-of-record lock-in
Once a carrier runs its loads, settlements, documents, and customer portal on the platform, switching means re-platforming the entire back office. The TMS becomes too painful to leave — not because of contracts, but because of data gravity.
Moat 02 · Data
Proprietary outcome data
Every load processed is a labelled training example: offered rate, negotiated rate, accepted or rejected, lane performance, broker reliability. At 200,000 loads/month, this rate and negotiation dataset has no equivalent anywhere in the market.

Risk transparency

Risks we watch closely
on every build like this.

We believe clients deserve an honest view of the challenges in any AI build. Here is what we monitor closely in this project — and in any similar engagement.

RiskHow we manage it
AI gross margin compressionVoice and inference costs are metered usage, not bundled into flat pricing. Monitored per-customer monthly with alerting thresholds.
Regulatory exposure (FMCSA, TCPA)Legal review of architecture before launch. Consent capture and AI disclosure built into the comms layer from day one — not retrofitted.
Thin-wrapper commoditisationProprietary data capture and system-of-record lock-in are first-class product goals, not afterthoughts. Architecture enforces this from the first line of code.
Trust and adoptionHuman-in-the-loop by default. Autonomy is earned with outcome data over 6–12 months, not assumed. Dispatcher remains in control throughout v1.0.

What this means for you

This case study is a
demonstration of how we think.

The AI Dispatch Platform is not just a product story. It shows how Devsphinx approaches a build — from market analysis and competitive mapping through architectural decisions to production engineering. The same thinking applies to your engagement.

🏗
SaaS founders adding AI to an existing product
The pattern here — identify high-frequency low-judgment workflows, build focused agents, gate autonomy with human approval, expand as trust accumulates — applies directly to your product. We can run the same discovery process on your workflow and identify the 20% of tasks representing 80% of the labour.
8–12 weeks to first AI workflow in production
🏢
Enterprises building internal AI automation
The AiTask table in this system logs every model call: input, output, tokens, latency, and whether a human overrode the AI decision. That's the audit trail your compliance team needs. The event-driven architecture means any agent can be isolated, replaced, or disabled independently.
Architecture-first · Compliance-ready from day one
🚀
Startups seeking technical partnership
Every decision in this build — Inngest over BullMQ, Cloudflare over SendGrid, cache table over Redis — was made to eliminate unnecessary cost and complexity at the MVP stage while keeping the path clear to scale. This is CTO-level thinking for every stage of your build.
Architecture · Vendor selection · Hands-on engineering

Want to build something like this?

Book a discovery call → See AI Revenue Engine →