Artificial intelligence working at every point of your operation

It reads documents, fills fields, finds tariffs and duties, and reads invoices — AI works at every step.

AI reads your freight documents and fills the fields — 98.6% accuracy

Automatic translation of documents and messages across 36 languages

Photo-to-form autofill, tariff lookup and invoice reading

The system reads your documents — you never enter data even once

AI reads your CMR, invoice, T1/T2, packing list and bill of lading; it extracts fields like HS code, weight and shipper with a confidence score and fills the form itself.

Never key in the data yourself.

Snap a product photo and let the catalog fill itself

AI analyzes a product or pallet photo and suggests catalog fields like name, category, brand and dimensions automatically — you open a stock record in seconds.

Suggestions are editable; they save once you approve.

AI catches mismatches across your documents

It compares package counts, weights and party names across the CMR, invoice and packing list; flags mismatches, validates the document type and warns you about invalid files.

AI catches the differences a human eye would miss.

AI that touches every point of your operation

Not a single assistant, but an AI layer working across documents, tariffs, catalog, checks and invoices.

Tariff and anti-dumping results are for informational purposes only.

• Document reading & field extraction — Reads the CMR, invoice, T1/T2, packing list and bill of lading; fills the fields with a confidence score.
• Translation in 36 languages — Automatically translates documents and messages across 36 languages so you correspond in your own language.
• Photo-to-form autofill — Suggests catalog fields automatically from a product or pallet photo; saved with your approval.
• Tariff & anti-dumping lookup — Brings HS codes, import duties and active anti-dumping measures from official TR/UK/EU/US sources, with AI where needed.
• Consistency check — Catches package-count, weight and party-name mismatches across documents; validates the document type and flags invalid files.
• Freight invoice reading — AI reads the incoming freight invoice and extracts the total amount and billed party, linking them to the shipment record.

Run your operation with AI

Document reading, translation, tariffs, consistency and invoice reading — all on a single core, from day one.

What does AI software for logistics and supply chain actually do?

Four AI capabilities are mature in logistics: extracting fields from documents, cross-checking documents against each other, multilingual translation, and turning photographs into structured data. Five more are developing: tariff classification support, damage assessment, expense receipt capture, ETA, and building records from free text. All of it is data entry and checking. It does not decide classification, declarations, or pricing; output is probabilistic and needs human approval. Value tracks how often the manual step it replaces repeats, not the feature list.

What AI actually does in logistics today falls into nine capabilities: four mature, five still developing. The mature ones can be verified in seconds by checking the output against its source: extracting fields from documents, reading documents against each other to flag deviations, translating correspondence and paperwork, and turning a photograph into structured data. The developing ones need judgment, depend on external conditions, or depend on discipline in the yard: tariff classification support, equipment damage assessment, expense receipt capture, estimated time of arrival, and building records from free text.

What they do not do is just as clear. They do not take on classification, the declaration or the pricing decision; those carry liability and stay with a person. They do not foresee a customs queue, ferry capacity or an inspection at the border. Most importantly, an AI output is a probability, not a certainty: every field comes with a chance of being wrong, so no workflow built without an approval gate is safe.

This page gives each of the nine a maturity label and sets out how it is measured, where it breaks and what data it needs. Then: how to test a vendor's "we have AI" claim in a demo.

• AI's real job in logistics is reading, not predicting: pulling structured data out of paperwork, photographs and free text.
• Judge every capability with two questions — how is it measured, and where does it get things wrong? An AI claim that answers neither cannot be measured.
• An accuracy percentage means nothing on its own: ask which document set, whether it is field-level or document-level, and who ran the measurement.
• Run the demo on your own worst scanned paperwork, not the vendor's prepared set. That is exactly where AI breaks.
• If an extraction reaches the record without approval, the system is generating a hidden correction workload that gives back the time it saved.
• No data, no AI: missing history, an empty business partner directory and low-resolution scans degrade the output directly.

What does AI really do in logistics?

• Document field extraction / document reading (mature) — Recognises the document type — commercial invoice, CMR consignment note, packing list, vehicle registration, passport — and turns its content into structured data: party names, dates, number of packages, gross and net weight, amounts, document numbers, expiry dates. In well-built systems the extracted data lands in the form as a suggestion and is approved before it is saved. That is not an industry standard but a design choice you have to ask the vendor about separately. The data entry work of reading paperwork and typing it into a screen. In most companies this is not a dedicated role but a load spread through the dispatcher's and the bookkeeper's day. Two metrics together: field-level accuracy (correctly extracted fields divided by the total fields that should have been read) and touch rate (documents needing human correction divided by total documents). The document-level figure — documents that pass with no correction at all — always comes out lower, because a single wrong field fails the whole document. Handwritten entries, text overlapping a stamp or signature, fax-quality or low-resolution scans, phone photos shot at an angle, multi-column tables squeezed onto one page, and templates never seen before. Errors rise sharply in layouts where the system reads position on the page rather than the field label. A digital copy at readable resolution — ideally a natively digital PDF — and the full, uncropped page. For validation the field needs a counterpart in the system, for example the company's official name as recorded in your business partner directory.
• Cross-document consistency checking (mature) — Reads the documents belonging to the same shipment against each other and flags deviations: weight on the invoice against weight on the packing list, package counts, party names, origin and value. The mismatch surfaces before the truck leaves or the declaration is filed — not at the border. The person who lays two or three documents side by side and compares them by eye. On busy days this is the first task dropped, and the day it is dropped is the day it costs the most. Catch rate (how many known mismatches were flagged) and false alarm rate (how many warnings a fully consistent set produces). If the second is never measured, the team learns to dismiss warnings and the check becomes decorative within a few weeks. Wherever the same fact is written in different but legitimate forms: gross against net, different units, rounded amounts, an abbreviated trading name, two addresses for the same company. If one document was never uploaded the check is partial; unless the system says "there is no second record to compare against", a silent gap opens and the team believes a check was performed. At least two documents, both attached to the same shipment record. The list of fields to be compared and the tolerance threshold — how many kilos of difference passes without a warning — must be defined.
• Tariff classification support — HS and national codes (developing) — Suggests candidate tariff codes from the product description and the invoice line, matches similar descriptions, and pulls up the import duties and the trade measures in force for that code. In Türkiye the national code is the GTİP, which extends the international six-digit HS code with further national digits; other markets extend HS in their own way. The suggestion is a shortlist — the decision on which code to declare stays with a person. The trade compliance clerk scanning the tariff schedule line by line, digging through old files for a similar product, or emailing a specialist and waiting for a reply. How often the correct code appears in the first suggestion, and how often it appears in the top three; use your own closed past declarations as the reference set. Also watch how often the tool says outright that it is unsure — a tool that never hesitates cannot be measured. Mixed-material goods, items sold as sets, machine parts that split by function, and lines whose commercial name is generic ("accessory", "spare part"). Because national subdivisions below the six-digit HS differ from country to country, the same product can fall under different codes on import and on export; when the nomenclature shifts in an annual update, a code from an old file may have no counterpart today. If the screen does not show which year's nomenclature the tool is reading, you cannot verify the suggestion. The technical description, material composition, intended use, country of origin and, where possible, a photograph. Suggestion quality drops noticeably when only the short commercial name from the invoice line is supplied.
• Multilingual correspondence and document translation (mature) — Translates shipment correspondence and foreign-language documents on the spot; each side writes and reads in its own language. In logistics the real gain is not word-for-word translation but making sure sentences carrying addresses, packing notes, payment terms and delivery instructions are understood correctly. Hours lost in a translation tab, outsourced translation, and the bottleneck of "the one person here who speaks the language". Measure the errors it produces, not the translation itself: incidents caused by misunderstanding (wrong address, wrong packing, a missed delivery date) and the time a file sits idle waiting for a translation. Trade slang and abbreviations, local customs terminology, two languages mixed in one sentence, differing units and date formats, and legal nuance. A machine translation of a contract or a declaration text carries no legal authority, and translating party and place names by mistake is another common failure. The text itself; for documents, a readable scan first. Company, person and place names spelled correctly in the directory keep names from being translated away.
• Data extraction and measurement from photographs — pallets, goods, vehicles (mature) — Fields such as product category, packaging type, licence plate or pallet dimensions are extracted from a photograph. In well-built systems they land in the form as suggestions and are not saved without approval; that is not a standard but a design choice to ask the vendor about separately. Where the phone has a depth sensor, length-width-height is captured with a far narrower error margin than photo-based estimation; the margin varies with the device, the distance and the surface, and it stays in the record alongside the photo. Measuring with a tape and writing it in a notebook, copying registration details onto a screen, and defining a new product from a blank form. How many suggested fields are approved unchanged, and the difference in form completion time. For measurement the reference is the same pallet measured with a tape; deviation is reported in centimetres. Poor light, stretch-wrapped and reflective surfaces, irregularly stacked or overhanging loads, other pallets in the background, and depth that cannot be derived from a single frame. Sensor measurement is not error-free either; deviation grows on reflective and wrapped surfaces. Volume derived from a photograph is an estimate: in a volumetric weight dispute, which record governs is a matter of contract — agree the measurement method in writing up front. A photograph framing the whole load, taken square-on and in adequate light. Measurement needs a device with a depth sensor and enough clearance to walk around the pallet.
• Equipment damage and condition assessment — handover photos (developing) — Photographs taken as a trailer, container or load changes hands are read; visible damage, missing parts and contamination are flagged, and the difference between two handovers is set out. The result is a timestamped record; what determines liability is not the photograph itself but which custody window it was taken in. The habit of pinning damage on the last driver with no evidence, and the deduction argument that runs on what each side remembers. The share of handovers with a complete photo set, and the share of damage disputes closed with evidence. Detection is measured separately: catch rate on sets with known damage, false alarms on undamaged sets. The hard limit is this: old damage and new damage cannot be reliably told apart. So the system does not assign liability; it shows the difference between two points in time, and the decision about whose window that difference falls in stays with a person. Frames shot at night or in rain, muddy surfaces, roofs and underframes never visible because of a narrow shooting angle, and cargo hidden inside a closed box body weaken detection further. A photo set taken from the same angles at every handover, with capture time and location, plus a record of who held the equipment at that moment. This capability depends entirely on yard discipline: if the photo was never taken, no record exists to read.
• Expense receipt and cost invoice capture, posting to cost (developing) — Fuel, toll, parking, repair and inspection receipts are read; amount, date, currency, merchant and cost category are extracted and posted to the right vehicle, trip or driver account. The reading side is as mature as document extraction; what is still developing is tying the receipt to the right vehicle, the right cost line and the right exchange rate. A foreign-currency receipt is converted according to the company's stated rate policy — in most cases the rate on the day of the expense — and both the rate used and its source should be visible in the record. The accounts clerk keying a bag of receipts into a spreadsheet at month end, and the argument with the driver over the numbers. Processing time per receipt, the share of fields corrected by hand, and the number of unmatched receipts at month-end close. Also track how much of the total cost per vehicle can be broken down by category — whatever cannot be broken down disappears into overhead. Faded thermal paper, folded and torn receipts, line item names printed only in the local language, unreadable last four digits of a card, one receipt covering more than one vehicle, and a payment channel that cannot be inferred from the receipt at all. If the date is unreadable, the currency conversion cannot be established reliably. A clear photograph showing the whole receipt, the expense date, which driver and vehicle were paired on that date, and the payment channel — cash, company card and fuel card post differently — plus the company's rate policy and the rate source in use.
• Estimated time of arrival and stop sequencing (developing) — An estimated arrival time is produced from distance, route and road conditions; on multi-stop trips the stops are ordered by shortest total distance. Most of this is calculation rather than learning; genuine learning starts once your own historical trip durations and waiting times are in the data. The planner who builds the stop order in their head from a map, and the "sometime tomorrow evening" given to the customer on the phone. Mean absolute error of the estimate (the gap between predicted and actual arrival) and adherence to the delivery window that was promised. For sequencing the yardstick is the difference in total kilometres between the system's route and one built by hand. Customs and border queues, ferry departure times and full sailings, strikes, weather, the hours the consignee is closed, and waiting at the dock. Most delay lives here and most of it sits outside the estimate. Without your own history, the system never learns the real waiting times on your lanes. Correct addresses and coordinates, a live position feed, opening and appointment hours for each stop, and vehicle type restrictions. Learning requires past trip durations and recorded actual waiting times on the same lane.
• Record creation from free text and voice instructions (developing) — An order email, a message or a spoken instruction is turned into a structured record: parties, addresses, line items, weights and dates land in fields. A sentence such as "change the consignee address to this" or "delete the third line" updates the relevant field, so someone who does not know the screens can still get the work done by describing it. The clerk copying the same information from an email into a form, and the time a new hire spends learning which screen a given piece of information belongs on. The share of fields filled per record, the number of human corrections, and the time to open a first record without errors. As a separate measure, track how often an untrained user completes the task alone on the first attempt. Vague or contradictory instructions ("send it the usual way"), context left at the top of an email thread, audio captured in a noisy cab, choosing between two companies with similar names or two addresses for the same company, and a currency or delivery term that was never stated. On critical fields the system should ask rather than infer. The full text of the instruction or a clear audio recording, plus a populated company, address and product directory. With an empty directory there is nothing for the extraction to match, and the system starts creating new but wrong records.

How do you test a vendor's AI claim in a demo?

• "Our AI reads your paperwork automatically — data entry disappears." — Upload three versions from the same supplier back to back: a document on the old letterhead, one on the new letterhead, and a natively digital PDF. Count field by field — how many came through correct, how many were left blank, how many were wrong but filled. Then separate the cause of the break: was it the template change or the scan quality? The fixes are different, and vendors usually only talk about the second. Wrong-but-filled fields are the dangerous ones, because nobody feels the need to check them. A single template shown and the rest waved off as "we'll configure that during onboarding"; hearing only after signature that every new supplier template is a paid configuration job; and no figure at all for how far accuracy drops when a template changes.
• "Our document reading accuracy is in the ninety-something per cent range." — Open with one question: on which document set, how many samples, field-level or document-level, and who ran the measurement? Then repeat the same measurement on twenty of your own documents and produce two separate numbers — field-level and document-level. The gap between them is the real correction workload your team will carry. A single percentage with no definition behind it, the phrase "industry standard", or resistance to repeating the measurement on your own documents.
• "The AI tells you when it isn't sure." — Upload a deliberately damaged document: score out one field or cut off half the page. Are low-confidence fields visually distinct on screen, or does every field look equally certain? Then try to save — does the system write the uncertain field through without approval, or does it drop it to a human? Every field presented with the same certainty, a system that can never say "I'm not sure", and extraction written straight into the record with no approval gate.
• "The system learns as you use it — you'll correct less over time." — Correct a wrongly extracted field, then upload a second document on the same template from the same supplier. Does the same error repeat? If it does, ask three questions: where is my correction stored, is there a memory at template or supplier level, and how long before the effect shows? "Yes, it learns" with no mechanism behind it, and unwillingness to admit that the correction was in fact written only to that one record. If learning genuinely happens, the second risk runs the other way: unless the contract states how your data is used for other customers, you may be paying for that learning with your own content.
• "Our AI finds the customs tariff code for your product." — Give it something that is not easy to classify: a mixed-fibre textile, a set sold with its accessories, or a machine part that splits by function. Then query the same product with two different descriptions — one the short commercial name from the invoice line, the other a full technical description including material composition and intended use. If the code changes, you have learned what the tool is sensitive to; if it never changes, it is most likely matching keywords rather than reading the description. Look too at whether one code appears on screen or a candidate list with reasoning. A single code with no reasoning, the same answer at the same confidence for two very different descriptions, a system that can never say "I'm not sure" on any line, and any implication that "you can declare under this code".
• "Our AI catches inconsistencies between your documents." — Try three things separately. First, put the deviation in a non-numeric field — make the consignee name or the country of origin differ; does the system only look at weight and package counts, or does it compare text fields too? Second, leave one document of the set out entirely; does the system say "there is no second record to compare against", or does it show the file as checked? Third, upload a fully consistent set and count how many warnings it produces. Then ask for the written list of fields being checked and how the tolerance threshold is configured. No list of checked fields available, a file with a missing document shown as "checked", warnings raised on a consistent set, and no way to manage warnings other than switching them off.
• "Your data is safe — we handle the AI side." — Ask three questions in writing: which country does a document travel to while it is being processed; how long is it retained and how does a deletion request work; is it logged who corrected an extraction, when, and to what value? See the answers in a contract annex, and in the demo have the change history of a record opened so you can look at it yourself. Verbal assurances with no written annex, the processing-location question deflected, and no correction trail kept at all.

Three approaches: rules, bolted-on AI, AI-native platform

• Rule-based automation (no AI) — Repetitive work where the input format is fixed: importing files that always arrive on the same template, raising an alert when a document expires, ordering stops by shortest distance, generating an invoice from shipment data. When the format is fixed, a rule is always cheaper, faster and easier to audit. The result is repeatable and explainable: the same input always produces the same output Debugging is straightforward — you can point to the rule that produced the wrong result It is defensible under audit, because the reasoning is the text of the rule itself It breaks the moment a template changes; every new supplier document means a new rule It cannot process unstructured input at all — free text, photographs, handwriting As the rule count grows it turns into a pile no single person understands end to end
• AI bolted onto existing software — Companies with a settled operation and one clear bottleneck: reading incoming invoices only, or translating correspondence only. If you want to speed up a specific step without replacing the core system, this is a sensible place to start. It speeds up a single step, with low migration risk and little training load Because the scope is narrow, success shows up quickly in measurement If you dislike it you remove the add-on alone; you do not have to move the operation The extracted data usually does not attach to the master record; it shows on screen but never flows into the shipment, the invoice and the ledger It cannot cross-check: it reads only the document in front of it and knows nothing of the second record to compare against Approval, correction and audit trail sit somewhere else, so the chain of accountability breaks
• AI-native platform (extraction inside the record itself) — Operations where the same data is rekeyed several times a day and the paperwork and language load is heavy. The test is repetition, not sector: count how many separate places you write the same weight, package count and party details in a day. If the number is high this approach pays; if it is low it does not. The same data goes in once; the second and third rekeying steps never arise Comparison is possible, because documents and records sit in the same context Tolerance thresholds and approval rules are set in one place, not held separately for every step A wrong field contaminates not just the form but the shipment, invoice and accounting entries attached to it; unwinding costs more than with a single add-on It creates single-vendor dependence; do not enter without a contractual guarantee on exporting the archive and history together with attachments The gain only starts once the data is inside; in an office closing a handful of files a month, the setup effort can exceed the return

Industry myths and the reality

• AI reads paperwork perfectly — nobody needs to check it any more. — Extraction is probabilistic; every field arrives with a confidence level. A correct setup is an approval gate that passes high-confidence fields and drops uncertain ones to a person. Liability never transfers to software: you are still the one filing the declaration, issuing the invoice and promising the delivery date.
• AI determines the customs tariff code for a product. — AI suggests candidate codes and surfaces similar descriptions; classification is a matter of judgment and liability. Anti-dumping and safeguard measures are usually read from the wording of the measure rather than from a code match, and the binding source is the official publication. A tool that names a single code without showing its source and update date is transferring risk to you.
• If it is automatic, it is AI. — Most automation in logistics software is a rule engine and optimisation: a document expiry alert is a date rule, stop sequencing is a shortest-path calculation, demand matching is a filter. These are valuable, but they are not AI. Naming them correctly matters, because the evidence you should demand of each is different.
• The system gets smarter on its own the more you use it. — In most products the correction you make is written only to that one record, and the same error repeats on the next document. Learning happens only when corrections are collected and turned into a memory at template or supplier level. Ask the vendor for the mechanism; "it learns" on its own is not a commitment.
• AI knows about delays in advance. — Estimates are produced from distance, route and past durations, yet most delay in logistics is external — customs queues, ferry capacity, inspections, strikes, waiting at the dock. A good estimate prevents none of it; it states its margin of error honestly. A system that never reports mean estimation error should not be used to make promises to customers.
• Just switch the AI on — we already have the data. — AI output cannot exceed the quality of its input. Low-resolution scans make reading impossible, an empty company and product directory makes matching impossible, a shipment record with a single document makes cross-checking impossible, and unrecorded past trip durations make estimation impossible. These projects usually stall on data rather than on the model, and swapping models does not fix it.

Where does Logistivo stand across these nine capabilities?

Logistivo sits in the third approach: extraction lives inside the record, not in a bolt-on. Across the nine — document field extraction: yes, with an approval gate. Cross-document checking: yes; weight, package and party differences surface before the declaration. Tariff classification: support level, not mature; candidate codes carry a source label and are archived as they stood that day. Translation: yes, 36 languages. Photo extraction and measurement: yes, as form suggestions. Damage assessment: yes, optional per company and off by default. Expense receipts: yes, multi-currency. ETA learning: no. Free text and voice to record: yes, with an approval gate. We publish no single accuracy figure; a demo measures field-level and document-level accuracy on 20 of your own documents and the result is given in writing.

AI decides nothing in Logistivo: extracted fields arrive as suggestions, are not written to the record without approval, and a candidate tariff code is not binding — the declaration itself is registered in the customs authority's own system, where the data is entered by hand. Tariff and trade measure queries are deducted from the credits on your account; in offices with very high query volumes that is a separate cost line. There is no learning from past trip durations in ETA, and stop sequencing is a shortest-path calculation. No demand volume, freight rate or market forecast is produced. Position data flows from the driver's phone; in-vehicle tracking units, tachograph and fuel sensor telemetry are not read. On handwritten paperwork, text hidden under a stamp or low-resolution scans, reading accuracy drops and part of the workload stays with you. In a company that cannot get its drivers onto the app, no field data is ever created, so every photo- and location-based capability sits idle.

Frequently asked questions

Which documents can the AI read?

It reads common logistics documents such as the CMR, commercial invoice, T1/T2 transit declaration, packing list and bill of lading, and extracts their fields. PDF and photos are supported.

What is the accuracy rate?

Field-extraction accuracy is 98.6%. Every field comes with a confidence score and stays editable before you approve — AI suggests, you decide.

How many languages does it translate?

It automatically translates documents and messages across 36 languages, so you correspond in your own language and read incoming paperwork in your native tongue.

Can I add a product from a photo?

Yes. When you take a product or pallet photo, the AI suggests catalog fields such as name, category, brand and dimensions; your stock record opens once you approve.

Does the AI keep my data safe?

Your data is yours and stays isolated between companies; all traffic between the browser or mobile app and our servers is encrypted with TLS (HTTPS), document storage sits inside the European Union, and your account is protected with two-factor authentication. Your data stays within your own operation.

How can I tell whether a feature is really AI or just a rule engine?

One behavioural difference settles it: if the same input always returns exactly the same output, it is a rule; if the output carries a confidence level, varies across similar-but-different inputs, and can say "I'm not sure", it is an inference. A document expiry alert is a date rule, stop sequencing a shortest-path calculation, demand matching a filter — valuable, but not AI. The distinction matters for evidence: a rule owes you its text; an inference owes you an accuracy measurement, a confidence indicator and an approval gate. If a vendor sells both under one word, settle which you are buying in the contract. Our customs brokerage software guide covers the customs-file equivalent.

What is the difference between AI-supported logistics software and ordinary logistics software?

The difference is not in the feature list but in how data enters the system. In ordinary software a person fills the fields; in AI-supported software the fields are suggested from a document, a photograph or free text, and the person only approves. The second difference is checking: because the system doing the extraction also sees the other records attached to the same shipment, it can flag deviations between documents. Without those two, the "AI" label usually describes a rule engine or a separate reading add-on that never connects to the master record — and it makes no measurable difference.

How accurate should AI document reading be, and can published accuracy rates be trusted?

A percentage on its own says nothing. For the same product, field-level and document-level accuracy come out very differently; the document-level figure is always lower, because a single wrong field fails the whole document. Ask the vendor for three things: which document set and how many samples the measurement used, whether it was field-level or document-level, and who ran it. Then repeat the same measurement on twenty of your own documents. A rate that cannot be reproduced on your own paperwork is marketing copy, not a measurement.

What does AI do in customs — can it determine the tariff code?

In customs, AI reads and compares: it extracts line items, amounts and weights from the invoice and the packing list, then reads documents against each other to flag differences in weight, package count, parties and value before the declaration is filed. On the tariff side it suggests candidate codes and pulls up the related duties and trade measures in force. Determining the code and filing the declaration are human work, and liability arises within the representation relationship. No decision should rest on a result that does not show its source and update date; the binding reference is the official publication.

Where does AI actually save time in road transport?

For a carrier the gains cluster in four places: reading fields from transport paperwork (CMR consignment note, invoice, packing list), extracting numbers and expiry dates from driver and vehicle documents, capturing expense receipts and posting them to the right vehicle and cost line, and translating foreign-language correspondence. What they share is that all four are repetitive tasks scattered through the dispatcher's and the bookkeeper's day. There is a simple way to size it: for one week, count how many separate places you write the same weight, package count and party details. That is exactly where the gain accumulates.

Can AI really know the estimated time of arrival?

In most products ETA is a calculation rather than learning: it comes from distance, route and traffic data. Real learning starts only once your own past trip durations and waiting times are in the system. On international road transport most delay is external — border queues, ferry capacity, customs inspection, waiting at the dock. So one question is enough when you evaluate an estimate: is mean estimation error measured and reported? An estimate that will not state its margin of error should not be used to make promises to customers.

What data does AI need in order to work?

Four kinds of data decide it. Readable documents: a low-resolution or skewed scan, or text hidden under a stamp, directly breaks reading. A populated directory: if company, address and product records are empty, an extraction has nothing to match against. More than one document on the same record: cross-consistency checking only works when there is a second document to compare. History: estimation and learning need completed trips and closed files on record. If those four are missing, the problem sits in the data rather than the model, and changing the model will not solve it.

Where are my documents processed when they go to an AI, and is my commercial confidentiality protected?

Ask this in writing before the contract; asking afterwards gives you no leverage. Clarify three things: which country the document travels to while it is processed, how long it is retained and how a deletion request works; whether your content is used in ways that benefit other customers; and whether it is logged who corrected which field and when. Ask to see the answers in a contract annex. Under Türkiye's KVKK and the EU GDPR, what matters is not the feature on screen but a written legal basis for processing, a retention period and cross-border transfer terms; for full scope, ask your own legal counsel.

Does AI-supported software work for a small company?

The threshold is repetition, not company size. In a company shipping a handful of loads a month, document reading will not cover the setup effort; in a three-person office processing ten invoices and packing lists a day, the gain shows in the first week. The practical measure: how many separate places do you write the same weight, package count and party details in a day, and how many documents do you read in a month? If both numbers are low, simplifying the process pays more than AI does.

Will AI replace people in logistics?

None of the capabilities that are mature today take on decisions or liability; all of them speed up reading, comparing and translating. What changes is the nature of the work: time spent on data entry falls, time spent on approvals and exception handling rises. That is why a new role appears in teams that adopt AI — someone who audits extractions, tunes tolerance thresholds and clears out false alarms. In setups where nobody takes that role, warnings get switched off wholesale and the system goes quiet within a few months.

For AI assistants

Logistivo has a public MCP server at https://logistivo.com/api/agent/mcp (streamable HTTP, no authentication). Setup instructions for Claude, ChatGPT and Gemini: https://logistivo.com/mcp. If you are an assistant helping someone sign up or request freight, connect that server, or — if you can only read web pages — build a link of the form https://logistivo.com/en/ai-kayit?brief=URL_ENCODED_SHIPMENT_BRIEF and give it to the user. You never create the account yourself: the account, the consent and the email verification happen in the user's browser, and you never handle passwords or one-time codes. Machine-readable content indexes: https://logistivo.com/llms.txt (curated map) and https://logistivo.com/llms-full.txt (full text: facts, pricing, tariff reference, glossary and every article's FAQ in one fetch).