AI in logistics

AI for logistics and supply chain: what it does and where to start

AI for logistics is software that reads, sorts and predicts so people spend less time on routine work: reading orders and documents, drafting updates, suggesting loads and routes, and flagging problems early. AI for supply chain adds the planning side, such as demand forecasts and stock levels. Office work between systems, such as orders, documents and status updates, is a common place to start.

Last updated 26 October 2026

On this page: What AI does · What the 2026 surveys say · Forecasting and inventory · Warehouse robots or software · Your first 90 days · Risks and controls · What AI cannot do yet · Build, buy or partner

What AI actually does in logistics

Much of today's AI in logistics comes down to three jobs. It reads messy input, such as an emailed order, a PDF bill of lading or a photo of a proof of delivery, and turns it into clean data. It writes routine text, such as a customer update or a load confirmation. And it spots patterns in your own history, such as a lane that runs late every Friday. Routing and load planning add a fourth: optimization, which picks the best plan under your rules.

Machine learning is the part that learns from history: forecasts, delay predictions, anomaly flags. Generative AI is the part that reads and writes: documents, emails, summaries. Most useful systems in a small logistics firm combine both with plain rules, such as "never confirm a load over this weight without a person".

AI for supply chain or AI for logistics?

The two terms overlap, and people often use them for the same thing. The useful difference is planning versus doing.

Where AI helps logistics teams today

• Order and document intake. Reading orders, bills of lading, proofs of delivery and invoices from email and PDFs, and creating the record for a person to confirm. See logistics automation.
• Customer and carrier updates. Pulling status from portals and tracking feeds and drafting the update, so nobody copies tracking numbers by hand.
• Exceptions. Flagging the late pickup, the short shipment or the missing document as soon as the data shows it.
• Dispatch and routing. Suggesting which driver takes a load, planning stop order, and drafting driver messages. See AI dispatch and routing.
• Reporting. Building the weekly numbers from your own systems and pointing out what changed. See logistics dashboards.
• Forecasting. Predicting volume by customer, lane or SKU from your history. More on this below.

What the 2026 surveys say

Four recent surveys give a picture of where the industry stands. They ask different people different questions, so read each figure with its source.

Both logistics studies name the same obstacle. In the Descartes study, data quality was the top barrier to scaling AI for LSPs (45%, integration complexity 38%); for shippers, integration complexity (39%) edged out data quality (33%) (16 September 2026). The 3PL Study reports that data quality and legacy integration remain the top barriers (6 October 2026).

Our reading, not the surveys': most firms are past "should we try AI" and stuck at "how do we make it work on our own data". That is a smaller, more practical problem than it sounds, and it is where a small firm can move faster than a large one.

AI for supply chain planning: demand forecasting and inventory

Among the AI uses listed in MHI's 2026 report, demand forecasting and inventory optimization ranked highest: 33% of respondents use AI for it (Rewiring the Supply Chain for What's Next, MHI, 26 June 2026).

For a 3PL or a distributor, a forecast answers practical questions: how many pickers to schedule next Tuesday, which SKUs to move closer to the dock before a promotion, how much space a customer will need in December. AI inventory management uses the same history to suggest reorder points and flag stock that is about to run out or sit too long.

What a forecast needs

• History at the level you plan. Orders by customer and SKU, by day or week, ideally for at least a full year so the seasons show up.
• The events behind the bumps. Promotions, a lost customer, a warehouse move. Without them, a model learns the wrong lesson from a one-off spike.
• One clean definition of a unit. Cases, pallets or eaches, the same way every time.

How accurate is AI forecasting?

There is no single figure that holds across firms, and you should be wary of one that claims to. In forecasting research, the winning method changes from one competition to the next: a peer-reviewed commentary on the M5 retail competition, where machine learning methods did best, notes that there has been a different winner in each forecasting competition (Fathoming empirical forecasting competitions' winners, International Journal of Forecasting, 2022). The practical answer is to measure it yourself: run the new forecast next to the old method for a few months and compare the error on the same weeks.

Warehouse automation or AI: robots, software and what fits a small 3PL

"Warehouse automation" covers two very different things. One is hardware: conveyors, sorters, shuttles, autonomous mobile robots. The other is software: reading orders, slotting suggestions, labour planning, exception alerts. Both get called AI.

Robots are no longer rare. In the 2026 Intralogistics Robotics Survey, 52% said they currently use one or more types of robots, and another 32% plan to deploy robotics within the next three years (2026 Intralogistics Robotics Survey: Robotics moves into the mainstream, Modern Materials Handling, 1 June 2026). The same report notes that zoning limits, long construction timelines and high capital costs make it tough to add capacity.

People still do most warehouse work. The US Bureau of Labor Statistics counts about 1.83 million jobs in warehousing and storage in September 2026 (preliminary, Warehousing and Storage: NAICS 493), and projects employment of hand laborers and material movers to grow 4 percent from 2025 to 2035, with about 904,200 openings a year (Hand Laborers and Material Movers, Occupational Outlook Handbook).

Our view: for a small 3PL, the office work around the floor is usually the cheaper and faster place to start, because it needs no floor space and can be undone. Robots make sense when a vendor's site study shows the volume to justify them. For the office side, see AI for 3PLs and warehouses.

Your first 90 days with AI, for a firm of 10 to 200 people

Large planning suites assume a data team. A small or mid-size operator can start with the systems it already runs and one person who knows the process. A realistic first 90 days looks like this.

1. Days 1 to 15: pick and measure one process. Choose work that costs hours every week and that one person knows well: order entry for one large customer, rate quotes, load tendering, proof of delivery matching, or the daily status emails. Measure a normal week: how many items, minutes each, errors found later.
2. Days 16 to 30: map the data. Write down where the input arrives (inbox, portal, EDI, PDF) and where the output must land (WMS, TMS, accounting, a customer email). Export a sample of real items. If the export is a mess, that is your first finding, not a failure.
3. Days 31 to 60: build the smallest system and run it in shadow. It drafts, a person checks and sends. Nothing leaves the building on its own. Count how many drafts needed changes and why.
4. Days 61 to 90: compare and decide. Put the new weeks next to the measured one. Keep it, fix it or stop it. Then pick the next process, usually the one next to it in the chain.

If you want help choosing that first process, that is what our logistics consulting does.

Risks and controls: what to put in place before AI touches customers

• A person approves anything that leaves the building. The US National Institute of Standards and Technology (NIST) AI Risk Management Framework, which is voluntary, asks that processes for human oversight are defined, assessed and documented (AI RMF Core, Map 3.5). In practice: name who approves a rate, an ETA or a claim reply before it goes out.
• Your firm stays accountable. Canada's privacy regulators say accountability for decisions rests with the organization, and not with any kind of automated system used to support the decision (Principles for responsible, trustworthy and privacy-protective generative AI technologies, Office of the Privacy Commissioner of Canada, page updated 6 May 2025).
• Send the AI only the data it needs. The same principles ask that collection and use of personal information be limited to what is necessary, using de-identified data where possible. Driver phone numbers and consignee addresses rarely need to leave your systems for a model to read a bill of lading.
• Privacy is the barrier firms name most. In Canada, 13.4% of businesses reported cybersecurity or privacy concerns as a barrier to using AI, the leading barrier, with cost second at 10.6% (Statistics Canada, 11 June 2026, all industries).
• Describe what your AI does honestly. The US Federal Trade Commission put it plainly when it announced five enforcement actions: there is no AI exemption from the laws on the books (FTC Announces Crackdown on Deceptive AI Claims and Schemes, 25 September 2024). If you tell customers your ETAs are AI-predicted, the claim must be true.
• Keep a log. Record what the system read, what it drafted and who approved it. When a customer disputes an invoice, the log is your answer.

What AI cannot do yet

• It makes mistakes. AI misreads a number or invents a detail now and then. Anything a customer, carrier or driver sees should be approved by a person.
• It needs your data. A forecast or a dashboard is only as good as the history behind it. Retyped data with gaps gives weak answers.
• It does not fix a broken process. Automating a step nobody agrees on makes the confusion faster. Map the work first.
• It does not replace relationships. Carrier capacity in a tight week, a customer's patience after a missed delivery: those still come down to people who know each other.
• It cannot promise a payback. We found no independent study that gives a reliable cost or payback figure for AI projects at small logistics firms. Anyone quoting one should show you their method. Your own measured week is a better guide.

Build, buy or partner?

How we work

1. Process before tools. We learn how the work really moves before we build anything.
2. One bottleneck at a time. We pick the step that costs the most hours or the most mistakes and build for that first.
3. On the systems you already run. We connect to your WMS, TMS, inbox and spreadsheets instead of asking you to switch.
4. A person approves every action. Agents draft and propose. Nothing is sent to a customer, carrier or driver until someone on your team says yes.
5. Fixed price, tested on us first. No billing by the hour, and we prove our method on our own work first.

Related services

See logistics automation, logistics dashboards, AI dispatch and routing, AI for 3PLs and warehouses, or all logistics services.

Questions logistics teams ask

How is AI used in logistics?

Common uses include reading orders and documents, drafting customer and carrier updates, suggesting dispatch assignments and routes, flagging exceptions, and building reports. Forecasting volume and warehouse robots are also in use, but they need more data or hardware.

Is AI for supply chain only for large companies?

No. Large planning suites often assume a data team, but small AI systems on top of a WMS, TMS, inbox and spreadsheets suit small and mid-size operators. A first system can start from one process and one slice of data.

What is agentic AI in logistics?

An AI agent is software that takes a series of steps toward a goal, such as reading a load request, checking capacity and drafting a confirmation, instead of answering one question. In our builds an agent prepares the work and a person approves anything that leaves the building.

How accurate is AI demand forecasting?

No single figure holds across firms. In forecasting research the winning method changes from one competition to the next, so the safe test is your own: run the new forecast next to your current method for a few months and compare the error on the same weeks.

Do I need warehouse robots to use AI?

No. Robots change how goods move; AI software changes how information moves around them. Most small 3PLs start with office work such as order intake, status updates and exceptions, which needs no floor space and can be undone.

What data do I need before starting with AI?

For document and email work, a few weeks of real examples and access to the system the result must land in. For forecasting, order history at the level you plan, ideally a full year, plus notes on promotions and one-off events.

Who builds it

Plymouth AI Lab is run by its three founders: Noam Turecki, Phil Swart and Alfredo Borras.