TUNA SCOPE

Read the campaign story

TUNA SCOPE turned expert tuna assessment into an image-based tool and then made that tool part of a consumer-facing food story. Dentsu’s September 2019 account describes development with ISID and Sojitz and implementation at a Yaizu processing facility earlier that year. The system used images of tuna tail cross-sections and expert judgement as training material. A later Dentsu case describes adoption in Kura Sushi’s procurement and restaurant communication.

Editorial interpretation: the marketing proposition is inseparable from an operational change. Instead of using technology only to personalise an advertisement, the project makes a normally hidden selection process visible. A restaurant can then invite diners to consider how quality is chosen. This creates a distinctive proof story, but the story must accurately describe what the system evaluates and what remains outside its scope.

The developer’s accuracy and commercial claims are not independently validated in this study. Image-based quality assessment is not automatically food-safety certification, proof of sustainable fishing or a replacement for every expert responsibility. The practical lesson is to connect a technical capability with a customer-relevant explanation while keeping model performance, operational usefulness and commercial outcomes separate.

Start with a decision that can be examined

An artificial-intelligence story is stronger when the underlying task is precise. Define what a human assessor currently judges, what information is available and where disagreements occur. Ask whether the proposed output is a recommendation, a ranking or a decision that triggers action. Preserve expert review where consequences require it. In a new project, compare the tool with an appropriate baseline and include difficult examples rather than selecting only clear demonstrations. A visually impressive interface cannot compensate for an ambiguous task or a dataset that does not represent the conditions of actual use.

Translate capability without expanding the claim

Consumers may hear AI-selected and infer superior taste, safety, consistency and sustainability all at once. Decide which claim the evidence supports and explicitly separate the others. Explain the role of the image and the human team in plain language. If the system assesses one visible feature, do not imply that it sees every relevant property of the product. Have technical and category specialists review the advertising together. The strongest demonstration makes the limits understandable, allowing customers to value a specific capability without being invited to believe in an all-purpose technological guarantee.

Measure the workflow after the demonstration

A system can perform well in a curated test and still create work in everyday operations. Track image quality problems, cases that require escalation, time spent correcting outputs and differences across suppliers or conditions. Give staff a way to disagree without being penalised for slowing the process. Monitor whether the tool changes the distribution of work rather than simply removing it. Commercial evaluation should include integration, training and maintenance costs. If the consumer story depends on a technical process, there must be a reliable way to update that story when the process or its limitations change.

A bounded test for another brief

Proposed adaptation: a retailer could test an image-assisted sorting tool on non-food, low-value display samples, such as grouping fabric swatches by visible colour family. Use consented or owned images and a clearly labelled prototype. Compare suggestions with staff classifications, including ambiguous lighting and mixed-colour examples. Require a person to approve every final label and record disagreements. Show a small customer group two explanations of the tool and ask what they think it can and cannot determine. This tests operational clarity and communication without making food-quality, safety or sustainability decisions, and without treating a model output as an unquestionable expert verdict.

Measure agreement on the defined task, correction time, escalation rate, robustness to image conditions and customer understanding of limits. Evaluate operational cost separately from attention to the technology story or any presumed sales effect.

Do not deploy consequential quality or safety decisions without appropriate validation and oversight; stop communication that implies capabilities the system has not demonstrated.