AI for Trucking Companies Still Running IBM i: What Actually Works

Four things. That is the honest count of what AI does today for a trucking company still running dispatch on IBM i, and only one of them has shipped as a feature inside a freight product you can buy. The other three you assemble yourself out of tools IBM shipped in 2026. Everything past those four is a roadmap slide.

I went looking because the demand is there and the answers are not. Carriers type "as400 trucking software" into Google and land on pages that never use the word trucking. Meanwhile the software is alive: TL2000, Innovative Access Plus, ShipSoft, the AR Traffic suite, TMT Fleet Maintenance. The AI story in this category is happening beside the dispatch system rather than inside it, and that turns out to be the useful part.

Each product named below has a profile on the parent site: TL2000 dispatch and operations, Innovative Access Plus truckload dispatch, ShipSoft Parcel and Freight, the AR Traffic logistics suite and TMT Fleet Maintenance, all filed under AS400 trucking, TMS and logistics software. This article is about the AI layer that goes next to them.

Play one: document extraction feeding the dispatch system

This is the only one of the four that has shipped as a named product feature, and it did not ship on IBM i. Carrier Logistics launched an AI BOL Extraction module for FACTS, its freight management system for LTL carriers, covered by Supply and Demand Chain Executive on July 10, 2025. The module reads a scanned bill of lading, pulls the shipment data off it, and creates or updates the shipment in FACTS in real time instead of a dispatcher keying it in.

Two caveats. Carrier Logistics does not publish the platform lineage of FACTS, so this is an adjacent industry example, not an IBM i product claim. And no accuracy or time-saving figure has been published. The announcement says measurable cost savings and stops there.

The reason it still matters is the shape. Document intake is the highest-volume manual keying job in a carrier's day, it is error prone, and it sits at the edge of the dispatch system rather than inside the rating engine. That is the correct place for a first AI project. Extraction runs off the box and hands the result to Db2 for i through the same interface an EDI 204 load tender already uses. Nothing in the dispatch core changes.

Play two: predictive maintenance on fleet maintenance data

A carrier running TMT Fleet Maintenance is sitting on the best training data in the building and almost nobody treats it that way. Work orders, technician time, preventive maintenance schedules, parts consumption and warranty claims, with dates and unit numbers attached. That is a labeled failure history, and it is the raw material for predicting a component failure before it becomes a road call: a shop visit on Tuesday instead of a driver on the shoulder of I-80 on Friday night.

Here is where a buyer has to be careful. Trimble's current TMT Fleet Maintenance page describes an end-to-end view of vehicle health, names its optional modules, refers to cloud-powered processes, and mentions AI-powered tools only at the category level with no detail. It does not say IBM i, AS/400 or iSeries anywhere. So if you run the IBM i edition and a salesperson demos a predictive feature, get it in writing which edition it ships in and when. Otherwise you are buying a roadmap for a product you do not run.

Nobody in this category has published a predictive maintenance benchmark for an IBM i edition. Until someone does, this is something you build against your own maintenance history, not something you purchase.

Play three: agent access to dispatch and load data

The IBM i MCP Server is IBM's open-source Model Context Protocol implementation, in beta since October 2025. It lets an AI agent query Db2 for i through YAML-defined SQL tools over Mapepire connectivity, which our MCP Server explainer covers in detail. For a carrier that means an operations manager asking which loads are still unassigned after four o'clock, or which shipments for a given consignee missed their delivery window this month, and getting the answer off the live dispatch tables without waiting for someone to write SQL.

IT Jungle's May 4, 2026 piece is the clearest read on where this sits in IBM's plan. The MCP server is the agent's door into IBM i. Db2 for i is not yet a vector store, in IBM's own words they do not support the storage on Db2 yet, and OpenSearch is being positioned as the hybrid engine that searches across vector spaces, text and structured data.

For a carrier that distinction is not academic. Rate confirmations, tariffs and claim packets are documents. Loads, stops and settlements are rows. A real operations question usually needs both, and today those live in two places with two different search stories.

Two things to keep in front of you: the project is still beta, and the tools an administrator exposes define how much of the box an agent can touch, which is a feature rather than a limitation. Expose read-only SQL tools over the dispatch and load tables first and nothing else.

Play four: Bob against a thirty-year-old dispatch codebase

The rating engine has been quietly correct since 1998. It is undocumented, it has been revised by four people, three of whom retired, and it is the real reason nobody wants to touch the dispatch system. That blocker shows up in almost every conversation in this category, and it is exactly what IBM Bob was built for.

IBM's Bob Premium Package for i announcement, published July 9, 2026, describes curated skills and agentic workflows for IBM i development work: explaining RPG and COBOL programs, converting fixed-format RPG to free-format, refactoring, generating code, documentation and unit tests, plus a database mode for queries and SQL and DDS knowledge. Note the order. Explanation comes before generation, and at a carrier it should stay that way.

The closest named evidence is Heartland Co-Op. IBM's July 15, 2026 Power announcement cites the co-op estimating 60 percent faster time for new-to-platform developers to understand complex applications using the Premium Package. Heartland is an agricultural cooperative, not a carrier: a grain and inputs operation running the same kind of scheduling, movement and settlement logic on the same kind of IBM i codebase. It is IBM's customer estimate, not an independent benchmark, and it measures how quickly a developer understands code, not defect rate. Directionally useful. Not proof.

What has not shipped, and how to test the claim

AI dispatch and AI route optimization inside an IBM i-native trucking product. Industry coverage says the sector is heading there; no vendor in this category has documented it as a shipping feature on the box. If one tells you otherwise, ask to watch it run against your own loads and your own accessorial rules. A demo on sample data proves only that the demo works.

Governance: what leaves the box, who approves, what gets written down

Every one of the four plays is a data governance decision before it is a technology decision. Freight data is customer data.

What leaves the box. A hosted extraction model reading your bills of lading is receiving your consignee list, your lane structure and, on a rate confirmation, your pricing. Get retention and training terms in writing before the first document is uploaded. If the answer on whether the vendor trains on customer documents is vague, that is the answer.

Who approves. Extraction writes shipment records, an agent with SQL access reads dispatch tables, Bob generates code. None of those should reach a production dispatch, rating or settlement table without a person signing off, at minimum until you know the defect rate. Human in the loop is not a compliance sentence here. It is what keeps a truck from being dispatched to the wrong door.

What gets written down. Every extraction, every agent query and every Bob-assisted change should leave a row: what was asked, what came back, who approved it, what changed. IBM i journaling already does this well and is already running, so point it at the AI workflow on day one. If a freight claim is argued two years from now, "the AI read the bill of lading" is not a defensible answer. The journal entry showing which person accepted the extracted values is.