What the case actually confirms

An OECD discussion paper on AI adoption by small and medium-sized enterprises describes a small coffee roaster in San Francisco. The company already relied on digital tools before the generative AI wave: it sold coffee directly to customers in the United States and Canada through its online store, and the owner said e-commerce represented 45% of the business.

Generative AI did not replace that infrastructure. It became a layer on top of an existing process. The owner used off-the-shelf products, including ChatGPT, for product descriptions, SEO updates, marketing emails, and shipping-cost analysis. In a more experimental use, the owner asked the model to explain how to automate the removal of static buildup during coffee grinding, a previously time-consuming task.

The owner described the main benefit as using time more efficiently, while also stating the central limitation: outputs cannot be fully trusted and need careful checking. OECD does not report revenue impact, hours saved, or the measured result of the physical-process experiment. The case therefore does not prove a specific model's ROI. It demonstrates an adoption pattern: an existing digital process, narrow tasks, a human owner of the outcome, and mandatory verification.

Why a digital process comes first

The roaster already had a website, online orders, label printing, and shipping management. Those systems provide the context that makes AI useful. A model does not know the current carrier price, packaging inventory, batch margin, or delivery promise unless the company supplies verified data from its systems.

For a small business, the sequence matters:

  • define the source of truth for products, prices, orders, and rates;
  • choose one repeatable area where drafting or analysis consumes time;
  • give AI only the required data and a precise output format;
  • keep publication, email sending, or rate changes under an accountable employee.

If product data is scattered across files, shipping rates are updated manually, and costs are calculated with inconsistent formulas, generative AI will accelerate inconsistencies. The first project should then be reference-data cleanup, not an AI agent.

Four tasks, four risk levels

The tasks in the case appear similar, but they need different controls.

**Product descriptions.** A model can draft from structured fields: variety, origin, process, roast profile, tasting notes, package size, and brewing method. A person verifies facts and brand voice. Medical claims, unsupported certifications, and invented provenance should be prohibited.

**SEO and marketing emails.** AI proposes headlines, metadata, and copy. The system supplies only approved facts and currently valid campaign terms. Before publication, an employee checks price, timing, audience segment, and false urgency.

**Shipping-cost analysis.** This is a numerical task. A language model should not calculate final totals from free-form text. Rates, weight, zone, packaging type, and carrier invoices should be processed by deterministic code or a spreadsheet; AI can explain variances and group hypotheses. A person decides on rate changes after checking invoice samples.

**Advice about a production operation.** A model's answer about static buildup is an idea source, not an instruction for immediate equipment changes. The company needs manufacturer documentation, a safety review, a small isolated test, and rollback. The closer advice gets to a physical process, the less authority the model should have.

A minimal architecture without an “agent swarm”

This scenario needs a simple pipeline.

1. The product system, online store, and carrier export send only required fields to a working folder or internal service.
2. A task template fixes the goal, allowed sources, output schema, and prohibitions.
3. The model creates a draft or explanation but cannot publish it.
4. Deterministic checks validate mandatory fields, numbers, links, and campaign conditions.
5. An accountable employee sees source data beside the result and approves or returns it.
6. A log records template version, inputs, output, corrections, and approver.

RAG is useful when the company regularly uses a variety catalogue, editorial rules, equipment manuals, and carrier contracts. The index should contain approved documents with effective dates and access controls. For a few dozen stable rules, attaching a compact verified excerpt to each request may be simpler than operating a vector database.

A local model can be justified when prompts contain purchase prices, customer data, unpublished supplier terms, or internal production instructions. Locality, however, does not repair bad reference data or remove the need for logs. When data is not sensitive and volume is low, an external API with redaction and contractual safeguards may cost less than a dedicated GPU.

Data, access, and quality

A pilot does not need a full CRM export. A minimum dataset is enough:

  • product ID and approved attributes;
  • current price and campaign period;
  • shipment weight, zone, and service;
  • anonymised shipping-cost history;
  • editorial rules and prohibited claims;
  • reference examples of good product cards and emails.

Customer names, payment data, and full addresses are unnecessary for product-description generation. Shipping analysis can use zone, weight, dimensions, and price instead of identity and address. Purchase margins and carrier contracts should remain accessible only to roles that already require them without AI.

Measure quality separately for each task. For product cards: factual error rate, editing time, and acceptance rate. For emails: preparation time, correction count, and controlled A/B-test outcomes. For shipping: accuracy of detected variances and savings confirmed by invoices. For technical advice: ideas that pass a safety review and a small experiment.

The owner's “triple check” warning is a useful management signal, but the process should make it concrete: automatic field validation, employee review, and sampled managerial audit. Three identical human readings protect less than three different controls.

A modeled pilot economy

Consider a small company producing 40 product cards and emails each month and reviewing 300 shipments. Assume each text takes 25 minutes before AI and ten minutes to edit from a draft, saving ten hours. Manual shipping analysis takes another eight hours; a prepared variance list reduces it to three. Total released capacity is 15 hours.

At a fully loaded labour cost equivalent to RUB 1,500 per hour, the resource effect is RUB 22,500 per month. If the service, integration, and quality control cost RUB 12,000, the modeled net effect is RUB 10,500. A RUB 100,000 pilot would take roughly ten months to pay back under these assumptions.

This is an illustrative calculation, not a reported result of the OECD coffee-roaster case. Revenue growth is excluded because sales cannot be attributed to AI without a controlled comparison. If employees rewrite every draft, carrier rates rarely change, or content volume is low, the project will not pay back. The first four weeks should therefore track confirmed time reduction at equal or better quality.

A four-week pilot

**Week 1.** Choose one product category and one email type. Assemble approved facts, ten strong examples, and prohibited claims. Measure baseline time and errors without AI.

**Week 2.** Configure the template and create 20 drafts without automatic publishing. The reviewer classifies every correction: fact, style, price, legal wording, or unsupported promise.

**Week 3.** Add an anonymised shipping export. Code performs calculations; the model groups and explains variances. Verify hypotheses against original invoices.

**Week 4.** Summarise time, errors, accepted drafts, and confirmed savings. Scale only tasks with a measurable benefit and an accountable reviewer. Keep production advice in a separate experiment lane with no authority to alter equipment.

The practical lesson is not that every coffee shop needs a chatbot. AI becomes useful when a digital trail already exists, the task is narrow enough, and a person can verify the output against primary data. In that design, the model saves time while decisions and accountability remain with the business.