Casebook

The Casebook is a catalogue of documented deployments of artificial intelligence in viticulture, winemaking and the wine business. It holds 299 records; behind each is a source that was opened and read, and a statement of what that source does and does not support.

It is a register of cases rather than a survey of the market: it measures nothing, ranks no one, predicts nothing. It records what was done, by whom, with what result, and how well that is known. The corpus was assembled in August 2026 in three waves: the main regions and domains, then the geographies and topics those missed, then the Russian market and the industry’s business edges — insurance, logistics, regulators. 177 records are about the vineyard, 47 about the cellar and the laboratory, 75 about business, marketing and hospitality. The tilt towards the vineyard is not an editorial choice: a satellite, a drone and a weather station produce thousands of observations a season; fermentation, dozens a year.

What counted as a case

Three conditions, all required. The record concerns wine, wine grapes or the wine business. It has an algorithmic component. And it is backed by a source that was actually opened and read — not a press release known by paraphrase, not a link that leads nowhere. The third excluded more than the first two: unverified leads never entered the catalogue at all.

The boundary around AI is drawn deliberately wide. The industry uses the word loosely, and an unstated definition would make the catalogue impossible to check — so it is written out in full, in both directions.

• Counts as AI: machine learning of any kind, from logistic regression and random forest to deep networks; computer vision and image recognition; time-series forecasting on weather and sensor data; chemometrics and spectroscopy paired with classifiers; large language models; recommender systems; autonomous navigation with machine perception; optimisation learned from data.
• Does not count: measurement without prediction — refractometry, chromatography, biosensors, differential thermal analysis; cryptography and blockchain; arithmetic on fixed coefficients; classical BI reporting; GIS layers overlaid without a predictive model.

The boundary does not fall where the industry puts it. Seven well-known systems sit outside it though described as AI almost everywhere: Amorim NDtech’s gas chromatography, Anton Paar’s inline refractometry, Selinko’s cryptographic tag, eProvenance’s fixed rules, Sentia’s biosensor strips, the IWCA emissions calculator and WinePulse’s BI reporting. None learns and none predicts. The reverse also happens — the word “AI” comes from a journalist rather than the company: neither “artificial intelligence” nor “machine learning” appears once on Oculyze’s or Vinescapes’ own pages, and the first of the two does use AI by our definition.

Under that frame 287 records are AI cases proper. A further 5 are borderline, and 7 describe adjacent technologies with no AI component; they are kept and labelled rather than deleted, because the absence of AI where the industry claims it tells the reader as much as its presence.

How to read a record

Two fields should be read before the rest, and both are filters in the register.

The confidence level rates the source, not the case. A — a peer-reviewed publication, an official EU or ministry report, or independent press with figures (109). B — trade press or a company’s own announcement with checkable detail (110). C — a vendor marketing claim with no independent confirmation (61); read those figures as “according to the company”. 19 records carry no level: they are the peer-reviewed research section, where the paper is the source. Neither A nor B means “verified”: it means the source is respectable, and the check found corrections in both categories.

The maturity stage says where a case stands, not how good it is: research (83), pilot (54), commercial operation (129), scaled (18), and closed, acquired or wound down (11). A sixth value, stage not established (4), is not a point on that ladder but the absence of one: those records’ own sources do not establish a stage, and until the audit they were filed as commercial beside a caveat saying exactly that.

Commercial operation and scale together are just under half the catalogue; most of the rest has not left the laboratory or the pilot, or has already ended. Several of those pilots are in their fifth to eighth year, a finding in itself: the road from a working model to a working estate is longer than the road to the model. The register also filters by domain, by 12 technologies and by 11 geographical groups, and adds full-text search and sorting.

What the audit found

All 299 records were reopened against their sources by seventeen independent checkers, briefed to find the error rather than confirm the record.

The first result: not one fabricated source, not one non-existent company, not one non-existent paper. Every link opens or is unavailable for a stated reason, every company named is real, every publication cited exists.

The second result is less comfortable and no less important. 181 records were confirmed in full. In 92 the discrepancy was corrected in the text, 16 were reclassified as vendor claims and attributed to the company, in 7 the figure turned out not to be about grapes, and 3 described a target rather than an achieved result. In total, 118 records out of 299 — close to forty per cent — carried a discrepancy in some detail. Not one was an invention; all belong to five recurring types.

• The figure cited is simply absent from the source — the most common type by a distance. Neither “1,300 acres a season” nor “$11.2m” appears in the Saga Robotics article.
• An adopter or partner is named in the catalogue but not in the source. Jackson Family Wines were a separate news item on the same page, unconnected to CladisIQ.
• The stage is overstated: a pilot or a plan presented as operation. Fujitsu’s single winery, running since 2011, was recorded as “scaled”.
• The figure describes a whole platform or region rather than grapes. Ningxia Nongken’s 606,000 mu and 136 wineries are the region’s numbers, not the corporation’s.
• The link is attached to the wrong page.

The check is visible in the records themselves: 85 records carry a caveat beside the result stating what the cited source does not support, and the other corrections went into the text of the record. Eight records were dropped outright: three duplicates; two with no source at all and no confirmed wine application; one research record whose link led to a journal’s front page, not the paper; one whose article was about other companies; and one that was not a case but a general industry claim with no estate, vendor or sample named. The audit’s own columns — the verdict, the action required, what the check found — are not published: they record how the corpus was checked, not the corpus. What is published is the consequence: corrected text, and a caveat beside the figure.

What is deliberately not here

Forecasts, market sizing and vendor rankings. The confidence level rates a source, not a product; the stage says where a case stands and is not a score. Nor are unverified leads here — a mention without a link, a link without a page and a page without grapes do not become a record, however interesting they sound.

Absence is recorded too. Ten business edges of the industry were checked on purpose and eight came back empty: vineyard valuation, auction estimation of bottles, nurseries and planting material, bottle lightweighting, shelf analytics, water and carbon reporting, dealcoholisation, and vineyard labour management. None of the eight held a deployment with a learned model and a checkable result. There are no records for them in the catalogue, and that is a statement rather than a gap.

The register itself is the Cases page: every record on one sheet, with filters, sorting, search and a source link on each. The catalogue is knowingly incomplete — it holds what is published and checkable, not everything that happens — so the same page carries a form: if you know a case we are missing, send the link. We will open it and read it, and if it survives the same three conditions it goes into the next data release.