Claims Copilot Request a pilot

Independent solution for the Nestlé Vietnam AI Reinvention program, use case NH-1. Not a Nestlé product.

Nutrition ads, checked before they reach Legal.

Upload a draft and get Red, Amber or Green in about a minute. Every finding marks the line in the ad and cites the rule behind it.

Opens the working prototype with sample data. Switch between Marketer, RA, Legal and Administrator from the menu at the top right.

Reviews Immunity supplement launch post RV-0012 Lan, RA reviewerLN
Overall ratingRed: not ready
2 must fix 1 needs a decision 1 needs your input 13 passed
  1. 1Claim is not in the approved library“Mom’s number 1 choice” is a ranking claim with no approved claim behind it.NES-CLM-01
  2. 4Required statement missingVN-HS-02
    In the ad nowNo “not a medicine” statement
    What to doAdd: “Thực phẩm này không phải là thuốc…”

    Decree 342/2025, Art. 5(3)(b) and (c). Law on Advertising, Art. 19(2). Checked by fixed code.

  3. 2Stronger than the approved claimApproved: “supports resistance”, not “superior immunity”.NES-CLM-02
  4. 3Approved wordingNES-CLM-02
Waiting with Marketer, then RA and LegalApproval blocked while Red is open
Synthetic test ad. Rule IDs, wording and sources are the prototype’s own.

Checked againstas of 10 Nov 2026

Law on AdvertisingConsolidated text, 2025
Decree 342/2025Food and supplement ads
Decree 15/2018Ad-content confirmation
Decree 100/2014Products for young children
Nestlé public policies5 rules, published documents
Claims librarySynthetic now, yours in a pilot

See the prototype work, start to finish.

A narrated, step-by-step walkthrough of the working prototype, under three minutes: upload, rule packs, the check, sourced findings, the review path, video, and the rule and claims libraries.

Check every feature yourselfThe live prototype runs with sample data, including the Nestlé key visuals shown in the video. Upload the sample ad, open any review, switch roles, and browse the rule and claims libraries.
Open the live prototype
Recorded from the working prototype on 22 Nov 2026, with an English voice-over and subtitles. It shows a Nestlé key visual taken from Nestlé Vietnam’s public product pages; the findings are the prototype’s own and have not been confirmed by RA or Nestlé. The video ad shown is synthetic. Share it on its own: claims-copilot.vibekaro.dev/demo

One word can send an ad back to the agency.

The rules sit in the Law on Advertising, three decrees, company policy and every product’s approved claims. RA and Legal usually see a draft after it is designed, so a small slip costs a full round.

  • A missing mandatory statement is a Red finding, even on a perfect design.
  • Agencies and creators reword approved claims without meaning to.
One claim, three wordingsChildren’s immunity supplement
Hỗ trợ tăng cường sức đề kháng Helps support resistance
Green: approved wordingNES-CLM-02
Tăng cường miễn dịch vượt trội Superior immunity boost
Amber: stronger than approvedNES-CLM-02
Điều trị dứt điểm táo bón Cures constipation for good
Red: implies treating a diseaseNES-CLM-03Law on Advertising, Art. 8(9) and 19(1)

The model reads. Code applies the rules. People decide.

Everything that sets a rating is fixed code and the approved claims, so the same ad always gets the same answer.

  1. 1Read

    Claude reads the whole ad

    Every line of text, the visuals and claims made by pictures, each with its place on the ad. The model only describes; it never decides.

  2. 2Check

    28 fixed rules, checked by code

    The product category and channel pick the rule packs. Each finding names its rule and source. Where policy is stricter than the law, policy wins.

  3. 3Decide

    RA and Legal make the call

    RA confirms or dismisses each finding with a reason. Legal approves. A Red finding blocks approval until it is fixed.

A refused, cut-off or failed read is never Green. It becomes Manual check, and a person looks at it.

Video, checked frame by frame

Keyframes and on-screen text are read in order. Every finding carries its timecode, so the editor knows which frame to change.

  • MP4 up to 3 minutes, plus images, PDFs and pasted copy
  • Spoken statements on TV and radio are flagged for a person to confirm
  1. 00:07Claim implies treating a disease. Law on Advertising, Art. 8(9) and 19(1)NES-CLM-03
  2. 00:07Superlative needs proof on file. “dứt điểm” is absolute wordingVN-GEN-01
  3. 00:00Differs from the approved wording. Use “Hỗ trợ tiêu hóa”NES-CLM-02
  4. AudioStatement must also be read aloud. Needs your inputVN-HS-02

The tool flags. Your reviewers decide.

Every finding is a suggestion until a named person confirms it. Approval stays blocked while a Red finding is open, and every step is kept with the person and the time.

  • Dismissing a finding needs a reason, and it can be undone
  • Each new version is checked again from scratch
HistoryImmunity supplement launch post
  1. MKMai uploaded version 1Marketer09:12
  2. Check finished Red2 must fix, 1 needs a decision09:13
  3. LNLan confirmed 2 findings, dismissed 1RA reviewerDismissed: “This wording is approved for this product in the March claims update.”10:40
  4. MKMai uploaded version 2 GreenMarketer13:02
  5. MNMinh approved version 2Legal14:05
Reviews screen of the prototype: counts by rating and a table of synthetic test posts with their rating, findings and who each is waiting with
The whole queueEvery material, its rating and who it is waiting with.
New review screen: the upload area and the rule packs that will be checked for the chosen product and channel
Rules shown before uploadPick the product and channel; the rule packs appear first.
Rule library screen: the 28 rules with what each checks and its rating if it fails, and the detail of rule VN-HS-02
One rule file, owned by RAAll 28 rules with their sources and versions.

On 16 held-out test ads, no problem was rated Green.

Synthetic ads never used for tuning. Claude Opus 5.5 through the Claude API, run on 16 Nov 2026.

100%

of known problems found, across the 16 ads

0

ads with a problem rated Green, of 13 with a problem

13of 16

rated exactly as labelled; the other 3 were stricter

21s

per review, about US$0.06 of model use

Every test ad, by label and the tool’s rating

Labelled Red, 5
Labelled Amber, 8
Labelled Green, 3

Rated stricter than its label. None was rated more lenient.

Model and rule card

Model
Claude Opus 5.5. Tests ran on the Claude API; a pilot runs it in Microsoft Foundry.
Rules
Rule file 0.5: 28 rules, sources checked 10 Nov 2026.
Test set
48 synthetic ads; 16 held out and never used for tuning.
Known issue
“Điều trị” is flagged on some medical-nutrition pages. An RA decision is pending.
Not yet
Labels checked by a Vietnamese regulatory reviewer, and a benchmark on past RA decisions.

Built to run in the company’s own Azure.

The app, files and database live in the company’s subscription. Claude runs in Microsoft Foundry, hosted on Azure, so prompts and outputs never leave Azure.

Entra ID sign-inSeparate roles for marketers, RA, Legal and administrators.
Claude in Microsoft FoundryHosted on Azure, as Global Standard or Data Zone Standard (US); there is no Southeast Asia zone yet. Anthropic acts as an independent processor.
Private storageFiles in private Blob storage, data in PostgreSQL, both in the company’s subscription.
No ad text in logsLogs keep IDs, timings and costs only.
Full audit trailEvery upload, decision and approval, with the person and the time.
Cost and upload limitsDaily caps per user, set by an administrator.
Vietnamese data and AI lawBefore any real data, a transfer assessment under the Law on Personal Data Protection (91/2025) and Decree 356/2025, and a risk check under the Law on Artificial Intelligence (134/2025), done with Nestlé IT.
Synthetic material firstThe prototype has only seen synthetic and public material. Real drafts come after Nestlé’s IT-security review.
The tool says what it did not check. These stay with RA and Legal:Is the confirmation number valid?Is the document genuine?Final approval

The program’s six questions, answered.

Short answers for reviewers. The detail is in the sections above and below.

What business problem does it solve?

Claims are cross-checked by hand and late, against rules spread across the law, policy and each product’s approved claims. One slip sends a draft back to the agency.

The problem

What would a pilot prove?

That no draft RA stopped is rated Green, and that RA’s first answer comes sooner with fewer rounds. Measured on past decisions first, then on live drafts.

The pilot plan

What is the solution?

A pre-screen that reads the ad with Claude, checks it with 28 fixed rules in code, and leaves every decision to RA and Legal.

How it works

What data and integrations does it need?

The approved claims library, internal guidelines, advertising requirements and product registration data. Entra ID for sign-in, Claude through Microsoft Foundry. No DAM or CMS connection yet.

Where it runs

Where has it been tested?

On 48 synthetic test ads, 16 of them held out, and on seven public product pages. Not yet with a client.

The results

What does Nestlé keep afterwards?

Its data, the confirmed rule file and claims library, and every audit record, exportable at any time. The software stays with its maker, as the program sets out; terms after a pilot are agreed with Nestlé.

Talk to us

Prove it on past decisions, then on live drafts.

A twelve-week pilot with one brand team, measured against what RA and Legal actually decided.

  1. Now, Nov 2026

    Working prototype

    Full review path, 28 rules, image, PDF and video checks, test sets and deployment templates.

  2. By 15 Dec

    Load the library

    Import the approved claims, confirm the rules with a Vietnamese regulatory reviewer, demo on real drafts.

  3. Weeks 1–4

    Shadow mode

    Run on past drafts with known decisions. The gate: no draft that RA stopped is rated Green.

  4. Weeks 5–12

    Live with one brand team

    New drafts get a pre-screen first; RA still reviews everything. Misses reviewed weekly.

We will measure

  • Missed problems target 0
  • Time to the first RA answer
  • Rounds back to the agency
  • RA agreement with findings
  • Cost per review

What a pilot needs

Azure accessA subscription and two Entra app registrations.
Claude in FoundryA Claude Opus deployment for the pilot.
Approved claimsAn export of the claims library.
Past decisionsDrafts with their RA and Legal outcomes, under NDA.
Regulatory reviewAbout 10–15 hours to confirm the rules.
Two ownersOne in marketing, one in RA for the rule file.

Model cost to Matching Day estimated at US$300–650, billed to the company’s Azure.

What RA and Legal usually ask

Does it approve ads?

No. It advises at every step and never approves. RA confirms or dismisses each finding with a reason, Legal approves, and approval stays blocked while a Red finding is open.

What happens when the model is unsure or fails?

The material becomes Manual check, never a pass. Unclear picture claims are rated Amber so a person looks at them.

What if it misses something?

RA still reviews every material, so the tool adds a check rather than replacing one. In the pilot, misses are reviewed every week and turned into rule or prompt fixes, each tested against past checks before it goes live.

Which formats and languages?

Images up to 10 MB, PDFs up to 32 MB, MP4 up to 200 MB and 3 minutes, or pasted copy. Ads in Vietnamese; reports in English today, with Vietnamese reports planned.

Where does our material go?

In the pilot, nowhere outside the company’s Azure tenant: files in private storage, Claude called through the company’s own Microsoft Foundry deployment, and no ad text written to logs.

How are the rules kept up to date?

RA owns one rule file. A change is tested against past checks and goes live only after a second reviewer approves it. The current legal rules come from translations and summaries and still need about 10–15 hours of a Vietnamese regulatory reviewer.

What does it cost to run?

About US$0.06 of model use per image review on Claude Opus 5.5, measured on the test set and billed to the company’s own Azure. Daily cost and upload limits are built in.

Try it on last quarter’s drafts.

A working prototype today. A pilot with one brand team within twelve weeks of access. No confidential material is needed for the first conversation.