IDCRAFT, a vendor-independent RFID and NFC distributor and systems integrator based in Neuhofen near Frankfurt, Germany, has made its IDSCOUT Encode NFC encoding station generally available from 1 September 2026. The station lets an operator describe an encoding job in plain language rather than in memory banks and lock modes, and it runs a language model locally on the device to interpret that description. The design rule underneath it is the part worth attention: the AI never writes.
NFC encoding is the act of writing data onto a tag’s chip: a URL, a Wi-Fi credential, a contact card, a serial number. Setting a job up conventionally means understanding memory banks, lock modes, NDEF record structures and the particularities of individual chips. Label software assumes that knowledge exists in the room. IDCRAFT is aiming the station at label suppliers, media houses and print specialists who want to sell encoding as a service without first hiring an RFID specialist.
How the job actually gets to the chip
The operator types the job as it reads on the job ticket. The station recognises the pattern, pre-fills a form and marks every value with a traffic light: green where the value is literally evidenced in the job text, and it carries the original quote as proof of origin; orange where it is a model suggestion; red where it is missing. Nothing unevidenced is silently filled in.
Encoding then does not start until one label has been physically written and read back, displayed in plain text and confirmed by the operator. IDCRAFT says this first-piece check is enforced and cannot be bypassed. Only after that does the station write the batch, reading back every label as it goes, and produce a serial-number-bound CSV report with a checksum and the software and recipe versions recorded.
The architectural claim sits in the division of labour. The locally running language model, which IDCRAFT describes as open weights under an Apache 2.0 licence, model-agnostic and replaceable, only recognises patterns and extracts evidenced values. The job itself is assembled and written exclusively by deterministic, tested software working from a chip database and vetted recipes. A wrong suggestion has three places to die: validation, the guardrails, or the first-piece check.
That is a meaningful distinction for anyone nervous about putting a language model near a production process. The model is doing comprehension, which is what models are good at, and it is structurally prevented from doing the irreversible bit. A mis-encoded reel of labels is expensive, and the failure is often only discovered at the customer.
The test data, and one run they failed
IDCRAFT published an acceptance run of the language layer with thresholds fixed in advance, dated 16 August 2026 against recipe catalogue v8. It reports 100% slot accuracy at 298 of 298 job fields with every measurement repeated three times, 14 of 14 deliberately planted safety traps caught and stopped, and zero phantom values, build errors or determinism deviations.
The more persuasive number is the series behind it. The company says it has completed ten such runs since July 2026, covering more than 2,200 checked job fields in triplicate, more than 6,600 individual measurements, and that no faulty AI value has ever reached a label across any of them.
It also says run number nine missed the pre-fixed determinism criterion and was not released, the cause was fixed, and the following run passed in full, with the thresholds never moved across all ten runs. Publishing a failed run is rare in a product announcement, and pre-registering thresholds before testing is rarer still. It does not make the numbers independently verified, but it is a materially better standard of evidence than the industry norm.
“We gave the AI exactly one boundary: it never writes,” says Patrick Kochendörfer, founder and Managing Director of IDCRAFT GmbH. “It reads the job and backs every value with the quote it came from. Whatever goes onto the chip is built exclusively by tested, deterministic software, and release happens against a physically read-back first piece. That we withheld release of one acceptance run because it missed a criterion is part of that decision: a gate that never fires is not a gate.”
Offline by design, and what that does for the paperwork
The station runs entirely on the device, is described as air-gap capable, and requires neither a cloud account nor telemetry. It is operated through a browser over the local network. Job data does not leave the premises, which IDCRAFT notes also answers a set of questions that turn up routinely in NIS2 supplier security questionnaires.
The company has published a one-page EU AI Act assessment alongside the launch, which is worth reading if you are working through the same questions. Its position is that the setup assistant is an AI system within the meaning of Article 3(1) of Regulation (EU) 2024/1689, while the executing data lane is rules-based and, following the European Commission’s guidelines on the definition of an AI system, is not. It assesses the product as not high-risk, falling into none of the Annex III areas and not a safety component under Article 3(14), with the self-assessment documented internally under the logic of Article 6(4).
On transparency, the assessment states that the Article 50(1) obligation, in force since 2 August 2026, is met by showing the AI notice visibly in the operator interface at the point of interaction. It also notes that Regulation (EU) 2026/1744, the Digital Omnibus on AI in force since 27 July 2026, postpones the Annex III high-risk obligations to 2 December 2027 and the Annex I obligations to 2 August 2028, but that this has no practical effect here because neither applies, while the Article 50 transparency duty that does apply was not postponed.
The assessment is explicit that it is a manufacturer self-assessment and not legal advice. It also sets out what it leaves with the operator: training operators and documenting that training under the Article 4 AI literacy duty, not removing or obscuring the AI notice in the interface, and observing data-protection duties where job content carries personal data, such as vCards.
What it encodes, and what it costs
At launch the recipe catalogue covers link or text, fixed or with a running number, an individual link per label from a CSV, digital business cards as vCard 3.0, Wi-Fi access using WPS/WSC in WPA2 or open, Smart Poster with a display title, URI presets for phone, e-mail and geographic position, raw block and page writing, and customer-specific house formats.
Supported chips at launch are NTAG213 with 144 bytes of usable capacity, and the ICODE family: SLI and SLIX at 108 bytes and SLIX2 at 312 bytes. That is a deliberately narrow list, covering the common NFC label chips rather than the whole market.
The hardware is an industrial enclosure with a 7-inch touch display and an ACS ACR1552U reader, article number CT-IDS-ENC, developed and assembled in Germany. Price is on request. IDCRAFT also reports a live test under real-world conditions in which 4 of 4 encoded chip payloads were byte-identical to specification, each verified three times.
Encode is positioned as the first of a family. IDCRAFT names Integrate, a development assistant for the readers it distributes, as in progress, Verify, a validation station, as in planning, and Stage, Select and Comply as further members. The stated pattern across all of them is local operation instead of cloud, evidenced answers instead of claims, and an AI that by construction cannot break anything.
The caveats
Every figure in this article is IDCRAFT’s own, from tests IDCRAFT ran and defined. RFID News has not seen the station, has not run the acceptance suite, and has not independently verified any of the results. The AI Act assessment is a manufacturer self-assessment and organisations have their own obligations regardless of what a supplier concludes about its product.
IDCRAFT states that its own market research in August 2026 found no comparable product combining plain-language job intake, per-value proof of origin, an enforced first-piece approval and a fully offline audit trail in one device, and is careful to describe that as a research finding with no guarantee of completeness. We would put it the same way.
More on writing data to tags in the RFID glossary, and further coverage of readers and encoding equipment in the hardware section.
Read more at https://idscout.de
