Right now, a logistics operator is probably starting their day by opening fifteen or so browser tabs, one for each carrier portal. They check container statuses by hand, copy the details into a spreadsheet, flag anything unusual, and move on to the next tab.

This organization is aware of Agentic AI advancements and their potential. The real issue is that every carrier uses a different portal, many without APIs, and the data is only available if someone logs in to check. The problem isn’t awareness, it’s access.

This is the reality for Agentic AI in supply chain companies. Nearly every practical challenge in deploying Agentic AI comes down to this one issue.

The real data problem is about accessing systems.

People often say to fix your data first: consolidate master data, build a unified platform, and connect your ERP, WMS, and TMS. That’s the right long-term goal, but it’s also why many teams never get started.

Where the data actually lives

Supply chains have plenty of data, but much of it is stuck in systems that were never meant to share it:

• Carrier portals with no public API
• ERP systems that export to Excel but don’t support integrations
• Third-party warehouse platforms holding critical inventory data behind a login screen
• Regional carrier sites that vary by country, login flow, and data format

The data is there. The real challenge is accessing it programmatically.

How RPA made things harder

What makes this especially painful in logistics is scale. A mid-sized freight forwarder might be dealing with a dozen major ocean carriers plus a long tail of regional carriers in every country they operate in. When you rely on RPA scripts to interact with those portals:

• Every UI change breaks something.
• Someone has to identify the break, find the engineer, wait for a fix, and test it again.
• The cost of maintaining those scripts often outweighs the time saved by running them.

How agents solve this differently

An agent doesn’t use fragile UI scripts to interact with a portal. Instead, it works more like a person, adapting to what it finds. If it can figure out the API calls a portal uses as someone navigates, you no longer have to rely on the UI at all:

• A new carrier gets added in 24 hours.
• A UI change doesn’t break anything.
• No vendor dependency, no integration project

That’s the difference between building on a human workaround and building directly on the real data layer.

What Agentic AI is actually doing in production

In real deployments, the most common use isn’t ML models making forecasts. It’s agents taking over manual work that would overwhelm any human team.

1. Freight quote handling

A global shipping aggregator was receiving between 1,000 and 2,500 freight emails per day, with rates arriving in every format imaginable.

The problem

• Carriers sent rates in email bodies, PDFs, or Excel attachments - no standard format.
• A team of 100 to 150 people manually read and transcribed these into a spreadsheet.
• Salespeople had to wait for the spreadsheet to update before quoting customers.
• The company ran on 0.75% margins, so any transcription mistake directly affected profits.

How the agent works

Layer 1: Email ingestion

• A dedicated forwarding address collects all incoming rate emails.
• The agent reads each one, extracts rate data regardless of format, normalizes it, and populates a live dashboard.

Layer 2: Carrier portal network

• 15+ carrier websites carry rate data, but have no public APIs
• Agents navigate each portal, query rates for specific routes and container types
• The results appear in the same dashboard as the email-sourced rates, ranked by how competitive they are.

Layer 3: Final Output

• When a customer requests a quote, an agent activates immediately.
• Collects shipment details in natural language, queries all sources simultaneously
• It returns a structured quote, so a salesperson only steps in when the customer is ready to commit.

Results

‍

At 0.75% margins, faster quote responsiveness directly impacted revenue, not just efficiency.

2. Warehouse reconciliation

Reconciling data in multi-warehouse operations is a less obvious but significant source of operational slowdowns.

The problem

• 27 third-party warehouses, each sending its own file in its own format at its own cadence
• A 44,000-row ERP export that was supposed to match against those files
• Matching done via VLOOKUP, warehouse by warehouse, by hand
• Every discrepancy had to be found, traced, and resolved manually — a multi-day process

How the agent works

1. Reads the ERP export
2. Pulls the warehouse files across all 27 locations
3. Runs the match, flags discrepancies with context for investigation
4. Produces a reconciliation report on a schedule

The team reviews flags, not rows. What took days runs overnight.

3. Carrier portal discovery at scale

For a freight forwarder working in 50 or more countries with both global and regional carriers, integrating all those carriers is a huge challenge.

The problem

• Large carriers warrant proper integrations - the volume justifies it.
• For everything else: manual logins, missing data, or nothing at all
• Traditional integrations take 6 to 12 months per carrier.
• A company operating in 50 countries might have hundreds of relevant carriers, and the smaller ones rarely get integrated.

How it works with agents

• Any carrier portal gets mapped in 24 hours: the agent crawls the site, captures underlying network calls, and builds API documentation where none existed.
• Rate, booking, and tracking functions become queryable endpoints.
• Fifty carriers can be handled at the same time, so you get full coverage in weeks instead of years.

What still needs a human

Where the model falls short

Demand forecasting AI is good with numbers, but planners are still needed because a model doesn’t know about:

• The trade show next month will spike regional demand
• The competitor that just went out of stock
• The promotion that wasn’t flagged in the system
• Seasonal anomalies not captured in historical data.

People hold that kind of information in their heads. The model will just keep following past patterns unless someone steps in.

The right split between agents and humans

‍

The travel cancellation example shows why this division is intentional. The agent handled everything except the final refund step. That part stayed with a person, not because the agent couldn’t do it, but because it involved moving money and the team wanted an extra check. This was a deliberate choice, not a technical limitation.

The trap to avoid

Just because an agent can navigate a portal or match a row doesn’t mean it should decide to reroute a shipment or drop a supplier. Some agents can do this within strict limits. The best organizations set those limits carefully and make sure people handle decisions with real consequences.

The infrastructure point most teams miss

In the long run, having a unified data platform is important. But it’s also the reason many Agentic AI projects get stuck in planning for years.

What actually changes with agents

When you can navigate systems programmatically, even if there’s no published API:

• An integration that used to take 6 to 12 months has compressed to 24 hours.
• The agent explores the system like a user, captures the network calls, and creates API documentation where there wasn’t any before.
• The same approach that works for a carrier portal also works for a TMS, a WMS, or any web-based system.
• Integrations that used to take months can now happen in parallel within a week.

For a company with dozens of carrier portals, dozens of regional warehouse platforms, and an ERP that only exports to Excel, this is what makes it possible to start now rather than after a multi-year infrastructure project.

Where to start

Here are three areas where a first deployment usually pays off the fastest:

1. Exception management before forecasting

Find one workflow where someone is manually checking something every day:

• Container status across carrier portals
• Invoice mismatches between carriers and internal records
• Purchase order delays against expected arrival dates

Start by building the agent for that workflow. The task is clear, the results are measurable, and you’ll see time savings right away.

2. Reconciliation before optimization

If two systems are supposed to match and someone checks them by hand, you can automate that. Make sure your inventory numbers are correct before trying to optimize inventory levels. Optimization only works if the data is accurate.

3. Carrier portal coverage before TMS integration

If your biggest data gaps are in systems you don’t control, agents that can navigate those portals will close the gap faster than waiting for vendor APIs or long integration projects. Start where the data is stuck, not where the infrastructure is perfect.

The thing that actually makes decisions better

Companies that do regular scenario planning recover from supply chain disruptions 30 percent faster. The reason isn’t better algorithms—it’s having faster information.

If a supply chain team knows about a delayed container before a customer complains, they have more options. If they only find out from the customer, they have just one.

When agents pull data from carrier portals, warehouse systems, and ERP exports in near real time, instead of someone checking manually at the start of their shift, you get a much more up-to-date view.

At its core, practical supply chain AI is about replacing manual tasks:

• Replace the morning manual check.
• Replace the VLOOKUP reconciliation.
• Replace the freight email triage.

When you replace enough of these manual tasks, the planner’s job changes—not because AI replaced anyone, but because it took over the work that was getting in the way.

The routine of opening fifteen tabs every morning can be solved. That’s the best place to begin.

Adopt AI builds agentic automation solutions for enterprise operations teams. This post is based on active customer deployments in logistics, freight forwarding, and supply chain operations.