Ask an operations leader how many automations touch their freight and you will usually get a number that is too low. The honest answer includes TMS add-ons, inbox rules, a macro someone wrote in 2021, a scraping bot, a couple of AI agents, and a handful of reps pasting customer emails into a personal chatbot.

The trouble is that nobody holds the full list, so nobody can say which of them can send a message to a customer, tender a load, or release a payment.

That gap is now getting attention well beyond logistics. By late 2026, enterprise software makers are shipping "agent management" products whose core job is to inventory AI agents across platforms and tier them by risk. An ops team can do the same with a spreadsheet and a clear set of tiers.

Why does an ops team need an agent inventory?

An ops team needs an agent inventory because it cannot govern, measure or safely switch off automations it cannot list. The inventory is the base layer for every other control: approval rules, audit trails, KPIs and incident response all assume you know what is running, who owns it, and what it is allowed to do.

Without it, three things happen in practice:

• Duplicate actions. An inbox rule and an AI agent both reply to the same "where is my truck" email, with different ETAs.
• Orphaned automations. The person who built the carrier-update script left. It still runs. Nobody knows what it writes to the TMS.
• Untraceable decisions. A customer disputes a quote. You cannot tell whether a person, a template or a model produced the number.

Each of these turns into a customer problem or an audit problem. Brokers already preparing for tighter oversight, as covered in our note on building an audit-ready brokerage, will find the inventory is the first thing a reviewer asks for.

What goes in the inventory?

Keep it to one row per automation and a few columns. Long forms never get filled in.

• Name — What to record: Plain-language label · Example: "Inbound quote intake agent"
• Owner — What to record: One named person, not a team · Example: Ops manager, LTL desk
• Trigger — What to record: What starts it · Example: New email to quotes@
• Systems touched — What to record: Read and write, separately · Example: Reads email, writes TMS quote
• Actions allowed — What to record: Verbs, not features · Example: Draft, send, tender, pay
• Channels — What to record: Where it speaks · Example: Email, SMS, WhatsApp, voice
• Human checkpoint — What to record: Where a person approves, if anywhere · Example: Rep approves quotes over a set value
• Kill switch — What to record: How to stop it in one step · Example: Disable rule in admin panel
• KPIs — What to record: How you know it is working · Example: Response time, error rate

The two columns that matter most are actions allowed and owner. Actions define risk. Ownership defines accountability. Everything else supports those two.

Include the uncomfortable entries. Personal chatbot use by reps belongs on the list, because customer data is leaving your systems whether or not anyone approved it.

How should automations be tiered by risk?

Tier by what the automation can do, not by how sophisticated it is. A simple inbox rule that sends a rate to a customer is higher risk than a capable AI model that only summarizes threads for a rep.

A four-tier model covers most logistics operations:

1. Tier 1: Read-only. Summarizes, classifies, searches, reports. Cannot change a record or contact anyone. Example: an agent that tags inbound emails by request type.
2. Tier 2: Drafts. Prepares messages, quotes or record updates that a person reviews before they take effect. Example: a drafted ETA reply waiting in a rep's queue.
3. Tier 3: Customer- or carrier-facing sends. Sends messages or updates records without a person in the loop. Example: automated ETA updates, document requests, quote replies within set parameters.
4. Tier 4: Commits money or capacity. Tenders loads, accepts rate confirmations, approves carrier payments, changes billing. Example: auto-tendering to a carrier from a routing guide.

What controls belong at each tier?

Controls should scale with the tier. Applying Tier 4 controls to a Tier 1 summarizer wastes review time; applying Tier 1 controls to a payment approval invites a loss.

• 1: Read-only — Minimum controls: Named owner, data access scope, quarterly review
• 2: Drafts — Minimum controls: Owner, sample review of drafts, edit-rate tracking
• 3: Sends — Minimum controls: Defined parameters, logged messages, escalation rules, weekly error review
• 4: Commits — Minimum controls: Approval thresholds, two-source verification, full audit trail, instant kill switch

The design of human checkpoints is where most of the judgement lives. Our guide to human-in-the-loop governance for freight covers where approvals add safety and where they only add delay. The broader policy layer, including data handling and vendor review, sits in a governance framework for logistics operations.

A practical rule: an automation moves up a tier only after it has run at the lower tier with measured results. An agent that drafts quotes with a low edit rate for a month has earned the right to send within limits. One that has never been measured has not.

Which KPIs should every agent have?

Every automation in Tiers 2 to 4 should carry at least three numbers, reviewed on a set cadence:

• Volume handled. How many requests, messages or records it processed.
• Accuracy or edit rate. Share of outputs a person corrected, or errors found in audit.
• Escalation rate. Share of cases it handed to a person, and whether those handoffs were correct.

Add one outcome metric tied to the workflow: quote response time for a quoting agent, on-time update rate for a tracking agent, days-to-close for a document chaser. If you cannot name the outcome, question why the automation exists.

These numbers only mean something if you can see them per agent rather than in aggregate. That is the argument for observability built around logistics metrics: an agent-level view that shows drift before a customer notices it.

A one-week way to build the first version

You do not need a program office. A workable first inventory takes about a week:

1. Day 1: Pull admin lists from your email platform, TMS, ERP and any integration tools. Every rule, bot, connector and app with write access goes on the sheet.
2. Day 2: Ask each desk lead a single question: "What runs without you touching it?" Add what they name.
3. Day 3: Assign one owner per row. Anything without an owner by end of day gets paused.
4. Day 4: Tier every row by its allowed actions.
5. Day 5: For every Tier 3 and Tier 4 entry, confirm a kill switch exists and write down how to use it.

Then set a monthly review. New automations go on the list before they go live, not after.

Frequently asked questions

What counts as an "agent" for inventory purposes? Anything that acts on your data or communications without a person doing each step: AI agents, inbox rules, RPA bots, TMS workflows, integrations, and scripts. Tier them by what they can do, not by what technology they use.

Should personal AI chatbot use be on the list? Yes. If reps paste shipment or customer data into a personal tool, that is a data flow you are responsible for. Listing it lets you replace it with an approved option rather than ignore it.

How often should the inventory be reviewed? Monthly for Tier 3 and Tier 4 automations, quarterly for the rest, and immediately whenever an automation changes owner or gains a new action.

The bottom line

Logistics teams already run far more automations than they have counted, and some of them can send, tender or pay without anyone watching.

List every automation, give each one owner, tier it by what it can do rather than how clever it is, match controls to the tier, and measure each agent on volume, accuracy and escalation. Build the inventory now, because the alternative is that it builds itself, one unlisted rule at a time.

Debales deploys AI agents for quoting, tendering, tracking and customer communication with clear action limits, named escalation rules and per-agent logs, so every agent in your inventory has an owner and a tier. Book a demo.