17 posts tagged with "AI"

Goods arrive with a delivery note — as a PDF in an e-mail, or on paper in the driver's hand — and someone in the warehouse retypes it into a goods receipt, line by line. Structured import solves this only when the supplier can send a clean CSV or XML file; most can't, or won't. When the delivery note is a PDF or a scan, there's nothing to import, so it's keyed by hand — and the stock figures are only as timely as that typing. This is a concrete case where a ready connector doesn't fit and an AI agent that reads the note, matches it to the order and drafts the receipt earns its place.

A lot of B2B orders don't arrive as tidy data — they arrive as an e-mail someone typed, or a PDF attachment, in no fixed format. Structured import and EDI assume the other side can export clean XML or CSV; when your customers just write "send me 20 of the usual plus 5 of the blue ones," there's nothing structured to import, so a person reads each message and keys it in. This is the case where a ready connector genuinely doesn't fit — and where an AI agent that reads the e-mail, extracts the order and posts it (with a human confirming) earns its place.

Most companies don't have one broken tool — they have five good ones that don't talk to each other. Orders arrive by e-mail, stock lives in one system, invoicing in another, and a person retypes between them. The honest answer isn't always "build something custom": if a ready-made connector exists for your combination of tools, use it — it's cheaper. Custom only earns its keep when no connector fits your tools or your process is genuinely non-standard. This guide is about telling those two cases apart — and what we build when it's the second.

Corporate law has a quiet, unglamorous half: keeping dozens or hundreds of client companies formally in order. Filings after every change, beneficial-owner registrations, official-journal announcements that start deadlines whether anyone read them or not. It's high-liability work that runs on spreadsheets, calendar reminders and one very careful paralegal. The pain concentrates in three concrete problems. We show how to build a corporate assistant wired directly to the registers — the Slovak commercial register (ORSR), the register of public-sector partners (RPVS), the official business journals and their Czech counterparts — that watches the sources, drafts the paperwork and keeps the deadlines, with a lawyer approving every step.

Legal AI tools that answer questions about legislation and case law already exist — that layer is covered. What no one covers is the layer where a law firm's real value lives: its own documents. The contracts it negotiated, the filings that won, the opinions partners signed. That know-how sits in folders and in people's heads, and it leaves with them. The pain concentrates in three concrete problems. We show how to build a private firm GPT that attacks all three: retrieval over the firm's own archive, answers with verified citations instead of confident fabrications, and confidentiality by design — hosted in the EU, with a lawyer approving anything that leaves the building.

"AI agents for business" is the most repeated phrase in tech right now — and one of the vaguest. Everything gets sold under it, from a web chatbot to promises of full autonomy. This guide is for companies that want a factual answer: what an AI agent actually does, what it can handle with invoices and orders in systems like Pohoda, Money or Helios — the ERPs Czech and Slovak companies actually run on — how to keep it under control, and how to start so the first deployment delivers a measurable number, not another slide deck.

Invoice data extraction (also called invoice OCR, invoice capture or data mining from invoices) is software that reads supplier invoices — PDFs, scans, photos, e-mails — and turns them into structured data for your accounting system: supplier, amounts, VAT, due date, line items. In 2026 it typically costs from a few cents to about €0.50 per invoice, and modern AI tools read 90–99 % of fields correctly. This guide explains how it works and what you'll actually pay — vendor-neutral, with no tool to sell.

Most companies roll out AI assistants the same way: buy licenses, run a kickoff training, and hope. Three months later nobody can say who actually uses the tool, who is stuck, or whether it brings anything — the honest answer to "is it worth it?" is a feeling, not a number. The pain concentrates in three concrete problems. We show how to build a learning loop around AI adoption that attacks all three: it reads real usage telemetry, gives every developer a personal level and personalized advice, gives managers an aggregate view — and stays firmly on the right side of the line between measurement and surveillance.

A cloud is never set up "insecurely" — it becomes insecure gradually. An instance gets added, a port is opened "just to test," a bucket appears, a service ships under a key nobody ever rotates. A year or two later you have 130 servers, thousands of snapshots, and nobody can tell you whether root is protected, whether the disks are encrypted, or whether you'd even notice someone getting in. We ran a read-only security audit of one mid-sized company's AWS account — through a read-only role, with not a single change to the account — and found exactly this quiet debt: root without MFA, 100% of disks and snapshots unencrypted, threat detection switched off in every region.

But an audit as a one-off PDF report is stale the day it ships — a week later you add a service, AWS ships a new feature, a new CVE lands. So we didn't build a report; we built an auditor that keeps living: every run it measures the same thing deterministically, and between runs it grows its own rule set — proving each new rule before it trusts it.

A lot of real work still starts on paper. Someone in the field fills in a work-order sheet by hand — what was done, where, how much, on which item — signs it, and it gets scanned. Multiply that by hundreds of locations across several branches and month-end billing turns into an archaeology project: someone retypes a thousand noisy scans into a spreadsheet line by line, then reconciles each row against the order list to find the right code, the right item and the right price. It's slow, it's error-prone, and the mistakes don't surface until they're already invoices. We built a controlled system that does this end to end — and it runs in production for a large organization with field teams, processing thousands of hand-filled work-order sheets a month.