A haulier rarely loses money on the driving. It leaks in the gaps around it. The phone rings while the dispatcher is already on another call. A quote request sits in a shared inbox until the customer books with somebody faster. A delivery note is photographed at a loading bay and typed into the ledger four days later by a person who was not there.
Search this sector and you land on enterprise dispatch platforms: load boards, solvers and fleet-wide optimisation built for operations running hundreds of trucks. Those products are real, and they answer a question most operators are not asking. If you run a handful of vans and two dispatchers, the useful question is narrower. Which jobs can software take over safely, which should you buy rather than build, which should nobody automate at all, and what do European and Estonian rules require before any of it touches a customer's freight.
Where a transport business leaks money
Every operator watches the fuel bill and the service schedule. Those are visible and budgeted and argued about. The expensive leaks are the ones nobody counts.
- Calls that arrive on a busy line. Many are a customer asking where a load is. Some are a new job that goes to whoever picked up first.
- Quote requests that age. A tender arrives as an email with a spreadsheet attached and the answer goes out the next morning, because the dispatcher spent the day on the road side of the business.
- Paper that has to become data. Consignment notes, proof of delivery, fuel receipts, weighbridge tickets. Somebody retypes them, and the invoice waits until they do.
- The same status question, asked again. The customer rings because nobody told them the truck was running late.
None of that is a routing problem. Routing is what vendors sell hardest, and it is the part a small fleet already handles with a map and a dispatcher who knows the roads. The general form of this argument is in business process automation for companies: automate the handling, keep the judgement.
Before you buy anything, count one of those leaks for a normal month. Our operator scorecard is a short version you can run yourself, and what a missed call costs your business shows how to put a number on the phone line specifically.
The jobs software takes over well
The phone line
A voice agent answers on the first ring, at 3am and during the Monday rush. It does three things a busy dispatcher does badly under load: identify the caller, look up the job, and either answer or take a clean message into the system. We build these on Vapi, with ElevenLabs for speech and Twilio for the line itself.
In Estonia the language mix matters more than the model choice. Freight gets arranged in Estonian, driven in Russian, and quoted to a customer in Finland or Latvia in English. Speech models handle small languages unevenly, so listen to a live call in each language you actually take before you sign anything.
The honest limit: a voice agent is good at answering and poor at deciding. It should tell a customer where a load is. It should not tell a driver to take a different route. The wiring detail sits in our voice agent CRM integration guide.
Quote and tender requests arriving as email
Most freight enquiries land as unstructured text with an attachment. A workflow in Make.com or n8n reads the mail, pulls out origin, destination, dimensions, weight and dates, writes a structured row into your CRM or transport system, and alerts the person who prices it.
That does not price the job. It removes the twenty minutes of retyping between the enquiry and the person who can price it, which is often where the work is lost.
Paperwork that has to become an invoice
The CMR consignment note is still, across much of this market, a piece of paper that gets photographed. Document extraction reads the photograph, pulls the reference numbers, matches them to the job and pushes the result into the ledger. The pattern is the same one we set out for accounting firms, where the money also sits in handling and not in the professional work.
Keep a person on exceptions. A blurred photo, a hand-written amendment or a mismatch against the order should stop and wait for a human instead of guessing.
Status updates sent before anyone asks
The cheapest automation in this sector is often outbound. If the telematics system knows an ETA has moved, a message can go to the customer before they pick up the phone. Every one of those messages is a call that never happens, and calls that never happen are the only kind that never sit in a queue.
What to buy instead of building
We build and run systems for a living, and there is still a short list we tell operators not to have built.
- Route optimisation. A crowded, mature software market. A custom solver costs more and performs worse than products already on sale.
- A transport management system. If you need a TMS, buy a TMS. The value we add is the glue between it and everything else.
- Telematics and tachograph handling. Regulated, standardised and already served by specialists.
What is worth commissioning is the connective tissue: the voice line, the intake, the document flow, the notifications, and the integrations between systems that were never designed to speak to each other. That is what our automation and voice agent work is in practice.
What nobody should automate
Three things, and the reason is the same each time. The cost of being wrong lands on somebody who did not choose the software.
Customs classification. Tariff codes and declarations to the Estonian Tax and Customs Board carry legal and financial consequences when they are wrong. A model may draft. A person signs.
Driving and rest time decisions. Regulation (EC) No 561/2006 sets driving and rest limits recorded by tachograph. Software can show a dispatcher what the tachograph says. It should not be the thing that decides whether a tired driver takes one more load.
Damage and claims judgement. When freight arrives broken the question is liability, and that is a conversation with a customer and an insurer rather than a workflow.
What is different about running this in Estonia
Language, and then paperwork
The three-language reality above is the first thing that breaks demos built and tested in English. The second is the local plumbing: permits and operator licensing through Transpordiamet, customs through the tax board, and accounting software that expects Estonian invoice formats. Anyone who has wired those before saves weeks over somebody meeting them for the first time.
Driver data is personal data
Telematics and tachograph records identify a named person and describe where they were and when. That makes them personal data, with the duties that follow. The Estonian Data Protection Inspectorate publishes guidance on processor obligations, and any vendor touching that data should be able to tell you where it is stored and under which agreement before the first call is recorded.
Being findable when the buyer asks a machine
Freight buyers increasingly ask an assistant instead of typing a query. On a transport client in Tartu the work was concrete and it was not dispatch: structured data coverage went from 4 schema types to 75, the hero image dropped from 60KB to 14KB, and an llms.txt file plus an explicit AI crawler allowlist told the assistants what they were allowed to read. That is a different job from automating operations, and we wrote it up in how to get your business cited by AI.
The rules you inherit either way
Two frameworks shape this work in Europe. The EU AI Act sorts systems into 4 risk tiers: unacceptable, high, limited or transparency risk, and minimal. Dispatch and document handling normally sit in the bottom two. The duty that bites daily is transparency: when a system speaks to a person, that person has to know it is not a human. Say it in the first sentence of the call.
The NIST AI Risk Management Framework is voluntary and American and remains the cleanest checklist published. Its four functions, govern, map, measure and manage, work as an honesty test: if you cannot say who owns a system and how you measure it, you have a pilot rather than an operation.
What it costs and where to start
Published market rates for AI consulting run 150 to 350 dollars an hour, with scoped builds landing between 20,000 and 150,000 dollars. We broke that down in how much an AI consultant costs. Our own model is deliberately different: the audit is free, runs about 30 minutes, and ends with your leaks ranked by monthly cost in a document you keep whether or not you hire us. Build-and-run starts from 600 euros per month plus VAT, quoted after the audit, because a price quoted before anyone has seen your call volume is a guess.
The sequence that works is short. Measure one normal month. Pick the single leak with the largest number beside it. Ship one system next to the existing process, not in place of it. Compare against the month you measured. The reasoning behind starting with a measurement is in the AI audit for business walkthrough, and the one-page version of the plan is in how to write an AI strategy.
How you know it worked
One test, monthly. Take the number you counted before you built anything and ask whether it moved. Not whether the team likes the tool, not whether the demo impressed the board. Whether the number moved.
A second test, quarterly. Can somebody other than the person who built it explain how it works and switch it off. If not, you bought a dependency rather than a system, and dependencies get expensive at exactly the moment you need to change something.
Questions operators ask about AI in logistics
What can AI actually do for a logistics company?
It handles the work around the freight rather than the freight itself. The four jobs that pay back first are answering the phone, reading quote requests out of email into a structured form, turning delivery paperwork into ledger entries, and sending status updates before the customer calls to ask. Routing is not on that list, because a solver you buy will beat one you commission.
Should a small transport company build its own dispatch software?
No. Route optimisation, transport management systems and tachograph handling are mature markets, and a custom build costs more and does less. What is worth commissioning is the connective tissue: the phone line, the intake of enquiries, the document flow and the integrations between systems that were never designed to speak to each other.
Can an AI voice agent handle calls in Estonian and Russian?
It can, and the language mix is the part to test first. Estonian freight gets arranged in Estonian, driven in Russian and quoted to Finnish or Latvian customers in English. Speech models handle small languages unevenly, so ask to hear a live call in each language you actually take, not a recorded reel in English.
What should a logistics company never automate?
Three things. Customs classification, because a wrong tariff code carries legal and financial consequences and a person has to sign it. Driving and rest time decisions, which sit under Regulation 561/2006 and belong to a human. And damage or claims judgement, which is a conversation with a customer and an insurer rather than a workflow.
What does it cost to start with AI in a transport business?
Our audit is free and takes about 30 minutes, and it ends with your leaks ranked by monthly cost in writing. Build-and-run starts from 600 euros per month plus VAT, quoted after the audit rather than before it. Market rates for the wider consulting sector run 150 to 350 dollars an hour with scoped builds between 20,000 and 150,000 dollars.
The trucks are the part of this business that already works. The phone, the inbox and the paper pile are the parts that quietly cost you jobs, and those are the ones worth pointing software at first. If you want the ranked version for your own operation, the shape of the engagement is on the consultancy page and you can book the audit from there.