Majento
ENRU
Get in touch

Teaching Marketing to Check AI: A Six-Review Exercise

6 min readBy Majento

AI trainingMarketing

A steel rolling cart holding finished marketing materials: stacks of brochures with colour blocks, a rolled poster tied with string, cyan and magenta packaging boxes and a colour swatch fan; a blank yellow tag hangs from the handle. Broad magenta, cyan and yellow bands cross the white background.

The session trains one skill: verification. An employee gets a draft from an AI tool. Then they check every quote against the customer reviews, every promise against the brand's own rules, every number against the original source. A mismatch means the line is struck out or flagged.

What the marketing team learns

The public education project AI Masters, run by Sergey Shima, co-founder of Majento, offers three formats for marketing teams. A short course of three two-hour sessions. A full course of six modules. Or a one-to-two-day hackathon, an intensive working session. Programs are listed on the AI Masters site.

The AI Masters blog lays out a sequence for bringing AI into a marketing department. First, list the repetitive tasks in the department. Then pick a few processes for a pilot, a small trial run. Next, build a library of prompts, ready-made instructions you feed to the AI, and templates. Measure time, quality, and output volume. Then gather team feedback. The plan for applying AI in marketing. Majento builds each program around the specific tasks and processes of a given business. No off-the-shelf template.

An AI Masters piece on the CMO's desk in 2026 describes searching meeting transcripts, the written record of meetings. The tool can return a quote and show where it came from. That is the skill this workshop drills: demand a quote from the AI, then open the source yourself.

Three files on the table

Each participant works with three documents. The brand platform, a short text covering what the brand promises, how it speaks, and what it will not say. A set of customer reviews. And a content brief, the instructions for the piece being written. The mechanics of the session and how the facilitator prepares are covered in a breakdown of the team's first workshop.

Final approval on brand materials stays with the person who owns the brand.

This is an editorial recommendation grounded in training programs, not a measured result. The logic is simple: a claim without a source does not make it into the text.

The exercise: review summary and a bakery post

The bakery is called Warm Crust. Fictional brand. All reviews are educational.

Brand platform excerpt: "We bake every morning. We speak simply and warmly, no superlatives. We do not promise what we have not verified: ingredients, health benefits, delivery times."

Six reviews:

  1. "The bread is fresh, but the courier was an hour late."
  2. "I love the rye, I buy it every Saturday."
  3. "The croissants are great, a shame they are gone by nine a.m."
  4. "Delivery arrived cold, the baguette had gone stale."
  5. "The staff are friendly, they always help."
  6. "Prices went up, but the quality is the same."

The assignment: ask the AI to produce a summary of the reviews grouped by theme, with quotes, then write a social media post draft based on that summary and the brand platform.

A wall organizer with two pockets, magenta and cyan. The magenta pocket holds review cards with four or five stars, the cyan one cards with one or two stars. Two hands take a card from each pocket and join them with a black binder clip.

The reference summary, verified by recount:

Topic How many reviews Numbers
Delivery 2 #1, #4
Freshness 2 #1, #4
Assortment 2 #2, #3
Stock availability in the morning 1 #3
Service 1 #5
Price 1 #6

For each topic the participant quotes the reviews with the listed numbers word for word. There are no conclusions beyond those six lines.

Four errors in the AI's answer

Here is what the AI outputs in the educational example:

"Most customers are unhappy with delivery. The best bread in town! 100% natural ingredients, delivered within 30 minutes. As our customer writes: 'Never ate anything tastier.'"

This is a fictional teaching example. It does not describe the behavior of any specific model. The participant must find four errors:

  1. "Most" is wrong. Delivery is mentioned in 2 of 6 reviews.
  2. The quote "Never ate anything tastier" does not appear in any of the six reviews. The AI invented it.
  3. "100% natural ingredients" and "within 30 minutes" are promises absent from the brand platform. The platform explicitly forbids promising ingredients and delivery times without verification.
  4. "Best bread in town" is a superlative the platform forbids.
A young woman with bright pink hair in an ikat-pattern jacket stands at an easel board of colour chips, fabric swatches and a cyan tin sample. She unpins a brief sheet with unreadable lines from the board.

Grading criteria: the participant found all four errors. Every quote in the corrected post is matched to a review number. No new promises appear beyond what the platform allows. The corrected draft goes to the brand owner for approval, not straight to publication.

Independent assignment and what comes next

The participant receives a new fictional brief (say, an announcement of a new bread variety) and three new reviews. They set the task for the AI themselves, check the answer, and attach a verification table: claim, where it came from, status (confirmed / not in source / needs approval). The standard: zero claims marked "not in source" remain in the final text.

Other exercises the department works through: sorting an audience into segments using educational customer descriptions, analyzing a sample ad campaign export (a downloaded file of results) with every number checked against the spreadsheet, and a human review of visuals against brand rules.

After the training, Majento designates AI champions, employees who help colleagues and collect tasks for implementation. If the team identifies a repeatable task that software can genuinely handle, the next step is marketing automation, where software takes over a repeatable step. That is a separate project, not a mandatory follow-up to training. More on automation here.

Majento terms: groups up to 15 people, in-person preferred, online also available. The program is built around the business's actual tasks and processes. Training can be purchased on its own. Implementation is not required afterward. AI training for teams.

Questions and answers

Can we use real customer reviews

Educational reviews are sufficient for the session. Real ones enter only after approval from the data owner inside the company, never with customer names or contact details, and only in a tool the company's rules allow. The verification method is identical.

Who approves the final text

The person responsible for the brand. The participant hands over the corrected draft. The company sets its own approval sequence.

Do we need to automate content right away

No. Automation is a next step if the team finds a task that software actually covers. Training can be ordered independently. The software decision comes when the task, the data, and the verification method are all clear.

Send your brand platform and a couple of anonymized reviews, and we will work through them in a training exercise. Write via Telegram or email.

Sources

  1. aimasters.me publication
  2. aimasters.me publication
  3. aimasters.me publication

View as Markdown