What is an AI marketing workflow?
An AI marketing workflow is a deliberate sequence of steps in which AI helps analyse customer data, segment audiences, plan campaigns, draft content, personalise and measure, while people keep decision rights at defined control points. In Vietnam it must also satisfy two legal frameworks: the Law on Personal Data Protection No. 91/2025/QH15 (in force from 1 January 2026) for data used in advertising, and the Law on Artificial Intelligence No. 134/2025/QH15 (in force from 1 March 2026) for transparency of AI-generated content.
AI makes drafts and analysis cheap, so the bottleneck moves to three questions: is the data lawful, is the message accurate, and can the result be measured? This guide covers the strategic workflow; day-to-day content production is covered in an AI content production workflow for marketing teams.
Map of AI in marketing: what AI does, what people decide
The fastest way to deploy AI in marketing without losing control is to split each stage clearly: what AI or an AI Agent does, what a person must decide, and which legal conditions apply. The table is a reference frame; the right level of automation depends on each company's data maturity and risk.
| Stage | AI assists with | People decide | Preconditions |
|---|---|---|---|
| Data and consent | Collecting, cleaning, flagging consent status per purpose | Processing purpose, legal basis, retention period | Consent under Article 28 of Law 91/2025; no bought or sold data |
| Segmentation | Suggesting customer clusters by behaviour, value, need | Which segments to target and which to exclude | Use only data consented for advertising |
| Campaign planning | Summarising research, sketching messages, proposing channels and timing | Goals, budget, core message, success metrics | A written brief with a measurable hypothesis |
| Content | Drafting variants, checking formats | Fact-checking, labelling, approval | Openable sources; labels under Article 11 of Law 134/2025 |
| Personalisation | Choosing variants per segment, send timing | Frequency caps, topics never personalised | Consent for personalised ads; opt-out mechanism |
| Testing | Designing variants, tracking results | Hypothesis, stop conditions, conclusions | Metrics defined before launch |
| Measurement | Aggregating data, spotting anomalies, drafting reports | Interpretation, next-period budget | One shared metric definition |
Foundation: customer data and consent under Law 91/2025
Every use of AI in marketing runs on customer data, so the starting point is whether that data is lawful. The Law on Personal Data Protection No. 91/2025/QH15 was passed on 26 June 2025 and took effect on 1 January 2026. Its Article 28 deals specifically with personal data protection in advertising services. Based on the Ministry of Public Security's summary of Article 28, the key points are:
- Processing customers' personal data for advertising requires the customer's consent, given with clear knowledge of the content, method, form and frequency of product promotion; customers must have a way to refuse advertising messages.
- Using personal data for advertising must comply with laws on spam messages, spam email and spam calls and with advertising law.
- Data subjects may ask to stop receiving advertising, and the advertiser must provide a mechanism to stop on request.
- Advertising service businesses are responsible for proving how they use customers' personal data for advertising.
- For behavioural, targeted or personalised advertising: data may be collected by tracking websites or apps only with consent; there must be a way to refuse data sharing, a defined retention period, and deletion when the data is no longer needed.
What this means for a marketing team
Article 7 of Law 91/2025 lists prohibited acts, including buying and selling personal data (unless another law provides otherwise). A purchased customer list is not a valid source for a segmentation model or a campaign. In practice, store consent status per purpose (email, messaging, personalised ads) as a data field so AI only uses records with the right flag. When a customer withdraws consent, the flag must update before the next send.
If a campaign touches children's data or sensitive data, the Law has specific provisions (for example Article 24 on children's personal data); involve legal counsel before that data goes anywhere near a model. General data governance principles are in data governance basics.
Segmentation, planning and personalisation with AI
Segmentation: AI suggests, people choose
AI is useful for finding customer clusters with similar behaviour that manual reports miss. But a statistical cluster is not yet a business segment. A marketer needs to name it, describe its need and decide whether to target it. Record explicitly which segments are excluded (for example customers who withdrew consent or have an open complaint) so AI does not pull them back in.
Planning: brief before prompt
A written brief sets audience, message, evidence, the expected action and success metrics. Prompts without a brief produce content that reads well and says nothing. The brief is also the review checklist and the baseline for measuring the campaign afterwards.
Personalisation: with clear limits
AI personalisation (choosing content variants, send times, product suggestions) falls under “behavioural or personalised advertising” in Article 28, so it needs consent and an opt-out. Beyond the legal minimum, set frequency caps and a list of topics that are never personalised (health, personal finances, family circumstances) to avoid content that is lawful but unwelcome.
A 7-step campaign content pipeline
Once a campaign has segments and a brief, content goes through the seven steps below. An AI Agent can handle drafting and format checks within its permissions; fact-checking, labelling, approval and publishing have an accountable owner under the human-in-the-loop principle. Fact-checking is not a formality: Google Search Central's guidance notes that generative models can produce inaccuracies and that manually fact-checking and reviewing AI-generated content before publishing is critical, including titles and descriptions.
| Step | What to do | Owner | Output |
|---|---|---|---|
| 1. Brief | Audience, segment, message, evidence, action, metrics; note whether a real person's likeness or voice is used | Campaign owner | Approved brief |
| 2. Draft | AI drafts variants per segment from supplied sources; log tool and version | Editor or AI Agent within its permissions | Draft with log |
| 3. Fact-check | Every figure, comparison and claim needs an openable source; unsupported numbers are removed, not softened | A checker other than the author | Source list, corrections |
| 4. Label | Decide whether AI-content notification or a label is required under Article 11 of Law 134/2025 and Article 18 of Decree 142/2026 | Editor | Label in its final display position |
| 5. Approve | Approve content, label and recipient list (only records with the right consent) | Team lead; legal when needed | Named, timestamped approval |
| 6. Publish | Publish from company-owned accounts; publishing rights limited to named roles | Person with publishing rights | Live post, platform AI-disclosure settings |
| 7. Measure | Compare results with the metrics defined in the brief; record lessons for the next brief | Campaign owner | Period report, adjustment decision |
When AI marketing content must be labelled
Article 11(3) of the Law on Artificial Intelligence requires deployers to clearly notify the public when AI-created or AI-edited content may cause confusion about the authenticity of events or characters. Article 11(4) and Article 18(2) of Decree 142/2026/ND-CP make an easily recognisable label mandatory for AI-created or AI-edited audio, images or video that simulate a real person's appearance or voice or re-enact a real event, for example an ad using an AI voice resembling a celebrity. Article 18(4) exempts spelling fixes, summaries, translation and quality enhancement that do not change the substance from a visible label. Details and a quick reference table are in labelling AI-generated content.
Guard the human voice
Decide in advance which communications must always be written by a person: apologies, sensitive matters, leadership messages and anything involving real people's circumstances. Automation should extend judgement, not replace it where judgement is the product.
Testing and measurement: define first, compare on the same yardstick
AI makes it easy to produce many variants, so the biggest testing risk is drawing hasty conclusions from too many variants and too little data. Three principles make the numbers more trustworthy:
- Define metrics before launch, in one place, and compare periods on the same definition. Illustrative example: “sign-up” counts only completed and confirmed forms, not form opens.
- Each test has one hypothesis, one main variable and a pre-written stop condition; AI proposes variants, a person makes the call.
- Separate campaign performance metrics (click-through, sign-ups) from process quality metrics (errors caught at fact-check, content corrected or removed after publishing, requests to stop receiving ads). The second set tells you whether AI is helping or adding risk.
When a metric cannot be defined
Where a metric cannot be defined clearly, report it qualitatively and say so, rather than stitching numbers together from different tracking sources. An honest report on data limits is worth more than a neat figure nobody can verify.
Access, accounts and logs
Drafting can be opened to many people and AI Agents. Publishing should be limited to named roles, and platform accounts, domains and analytics properties must be owned by the company, never personal profiles. Give AI Agents least privilege: a drafting Agent does not need publishing rights; a reporting Agent does not need to export the full customer database. Log which person or Agent created, edited, approved and published each item so you can explain decisions when asked.
How EYC approaches it
EYC builds AI marketing and growth and AI for content and media solutions on the principle “Human-led systems. AI-coordinated work.”: AI Agents help analyse, draft and compile figures; people decide segments, messages, approval and publishing. If your team is standardising an AI marketing workflow, contact EYC to discuss a suitable scope.
Frequently asked questions
What are the main uses of AI in marketing?
Common uses include customer analysis and segmentation, summarising market research, drafting content and ad variants, personalising messages, supporting A/B tests and compiling reports. People still decide goals, messages, approval and publishing.
Do I need consent to use customer data for AI marketing in Vietnam?
Yes. Under Article 28 of the Law on Personal Data Protection 91/2025/QH15, processing customers' personal data for advertising requires their consent with clear knowledge of the content, method, form and frequency of promotion, plus a way to refuse. Behavioural or personalised advertising also requires consent for tracking via websites or apps.
Does AI-generated ad content need a label?
It depends. The Law on Artificial Intelligence 134/2025/QH15 (Article 11) and Decree 142/2026/ND-CP (Article 18) require an easily recognisable label when AI audio, images or video simulate a real person or re-enact a real event, and require notification when AI content may mislead. Spelling fixes, summaries and translation that keep the meaning are exempt from a visible label.
Can I run AI campaigns on a purchased customer list?
You should not. Article 7 of the Law on Personal Data Protection 91/2025/QH15 prohibits buying and selling personal data unless another law provides otherwise. Data for segmentation and personalisation should come from customers who consented directly to your business.
Which metrics show whether an AI marketing workflow works?
Track two groups: campaign metrics defined in the brief (for example sign-ups, click-through) and process quality metrics (errors caught at fact-check, content corrected after publishing, requests to stop receiving ads). Define metrics before launch and keep definitions stable across periods.
Can an AI Agent publish ads on its own?
Technically yes, but it should be limited. We recommend letting AI Agents draft, check formats and propose labels, while approval and publishing stay with people who hold those rights, with every action logged.