{"id":26103,"date":"2026-09-28T19:15:53","date_gmt":"2026-09-28T23:15:53","guid":{"rendered":"https:\/\/marktg.ai\/?page_id=26103"},"modified":"2026-10-27T11:16:52","modified_gmt":"2026-10-27T15:16:52","slug":"comment-un-modele-mai-est-construit","status":"publish","type":"page","link":"https:\/\/marktg.ai\/fr\/comment-un-modele-mai-est-construit\/","title":{"rendered":"Comment un mod\u00e8le mAI est construit, test\u00e9 et gouvern\u00e9 | marktgAI"},"content":{"rendered":"

mAI Technology \u00b7 Build, test, govern<\/p>\n<\/div>

<\/div><\/div><\/span>

How an mAI model is built, tested and governed<\/h1><\/span>
<\/div><\/div><\/div>

Every mAI Custom AI Marketing Model starts from one documented framework and is baselined to one company. This page shows how, so you can inspect the method before you commit to anything.<\/p>\n<\/div>

Start a Strategic Conversation<\/span><\/a><\/div>
Review Evidence & Governance<\/span><\/a><\/div><\/div><\/div>
\n

At a glance<\/p>\n

One master framework.<\/strong> Method, evidence rules and governance, documented once.<\/p>\n

One company baseline.<\/strong> Your ICP, claims, KPIs, roles and risk limits.<\/p>\n

Tested before release.<\/strong> Safety behaviors are tested repeatedly, then retested when the framework changes.<\/p>\n

People approve what matters.<\/strong> Strategy, audiences, budgets and claims stay with your team.<\/p>\n<\/div>\n<\/div><\/div><\/div><\/div><\/div>

The build route:<\/strong> baseline one company \u2192 build from the master framework \u2192 test before release \u2192 govern in operation \u2192 measure against your baseline \u2192 learn under human command.<\/p>\n<\/div><\/div><\/div><\/div><\/div>

01 \u00b7 One framework, one company<\/p>\n<\/div>

<\/div><\/div><\/span>

Custom where it should be, consistent everywhere else<\/h2><\/span>
<\/div><\/div><\/div>

mAI is built in two parts. A master framework<\/strong> holds the operating method, the evidence rules and the governance controls. A company baseline<\/strong> holds what is specific to you. Every field is labeled, so it is always clear what is fixed, what is a default you can change, and what belongs to your company.<\/p>\n<\/div>

A master framework and a company baseline combine into one mAI model<\/title>The master framework contains fixed rules and adjustable defaults. The company baseline contains company-specific context. Together they form your mAI model.<\/desc><\/marker><\/defs>MASTER FRAMEWORK<\/text>FIXED<\/text>Method, evidence rules, approval list<\/text>DEFAULT, ADJUSTABLE<\/text>Risk tiers, budget thresholds, templates<\/text>+<\/text>COMPANY BASELINE<\/text>COMPANY-SPECIFIC<\/text>ICP and positioning<\/text>Approved claims and brand rules<\/text>KPIs, baselines and roles<\/text>Risk limits and approvers<\/text>YOUR mAI CUSTOM AI MARKETING MODEL<\/text><\/svg><\/div>\n<\/div>
\n
    \n
  1. Master framework<\/strong>Fixed: method, evidence rules and approval list. Adjustable defaults: risk tiers, budget thresholds and templates.<\/li>\n
  2. Company baseline<\/strong>Company-specific ICP, positioning, approved claims, brand rules, KPIs, baselines, roles, risk limits and approvers.<\/li>\n
  3. Your mAI model<\/strong>The master framework and company baseline combine into your custom AI marketing model.<\/li>\n<\/ol>\n<\/div>\n<\/div><\/div><\/div><\/div><\/div>

    02 \u00b7 From company URL to a governed model<\/p>\n<\/div>

    <\/div><\/div><\/span>

    Part automated, part approved by your people<\/h2><\/span>
    <\/div><\/div><\/div>

    Onboarding starts from your public footprint and ends inside your approval workflow. Automation does the first draft. Your team decides what the model is allowed to treat as fact.<\/p>\n<\/div>

    Six onboarding steps with two human approval points<\/title>Company URL, research draft, your approval of company facts, baselining, testing, and go-live inside your approval workflow. Steps three and six are approved by your team.<\/desc>AUTOMATED<\/text>1<\/text>Company URL<\/text>and public sources<\/text>AUTOMATED<\/text>2<\/text>Research drafts<\/text>the company card<\/text>YOU APPROVE<\/text>3<\/text>Your team confirms<\/text>what counts as fact<\/text>BUILT<\/text>4<\/text>Baselined to<\/text>card, KPIs and rules<\/text>TESTED<\/text>5<\/text>Tested against<\/text>your risk limits<\/text>YOU APPROVE<\/text>6<\/text>Goes live inside<\/text>your approval flow<\/text><\/g>Anything not yet confirmed stays marked as unverified until a person approves it.<\/text><\/svg><\/div>\n<\/div>
    \n
      \n
    1. 1 \u00b7 Sources<\/strong>Company URL and public sources.<\/li>\n
    2. 2 \u00b7 Research<\/strong>Research drafts the company card.<\/li>\n
    3. 3 \u00b7 Human approval<\/strong>Your team confirms what counts as fact.<\/li>\n
    4. 4 \u00b7 Baseline<\/strong>The model is baselined to the company card, KPIs and rules.<\/li>\n
    5. 5 \u00b7 Testing<\/strong>Tested against your risk limits.<\/li>\n
    6. 6 \u00b7 Human approval<\/strong>Goes live inside your approval workflow. Unconfirmed facts remain unverified until approved.<\/li>\n<\/ol>\n<\/div>\n<\/div>
      <\/div><\/div><\/span>

      What you provide<\/h4><\/span>
      <\/div><\/div><\/div>

      Company context, approved claims, KPIs and baselines, named approvers, and the risk limits you want enforced.<\/p>\n<\/div><\/div><\/div>

      <\/div><\/div><\/span>

      What marktgAI does<\/h4><\/span>
      <\/div><\/div><\/div>

      Researches, drafts, baselines and tests the model, and documents every decision so it can be inspected later.<\/p>\n<\/div><\/div><\/div>

      <\/div><\/div><\/span>

      What stays with your people<\/h4><\/span>
      <\/div><\/div><\/div>

      Strategy, audiences, budgets above threshold, brand-critical creative and regulated claims. Always.<\/p>\n<\/div><\/div><\/div><\/div><\/div><\/div><\/div><\/div>

      03 \u00b7 Two layers, one loop, human command<\/p>\n<\/div>

      <\/div><\/div><\/span>

      Every model runs the same operating loop<\/h2><\/span>
      <\/div><\/div><\/div>

      Whatever the company, the model plans, executes, measures and optimizes in that order, and never separates strategy from execution or measurement.<\/p>\n<\/div>

      <\/div><\/div><\/span>

      AI Marketing OS<\/h4><\/span>
      <\/div><\/div><\/div>

      How the work runs: Plan \u2192 Execute \u2192 Measure \u2192 Optimize, with explicit inputs, outputs and handoffs.<\/p>\n<\/div><\/div><\/div>

      <\/div><\/div><\/span>

      AI Marketing Brain<\/h4><\/span>
      <\/div><\/div><\/div>

      How the model reasons: prioritize, predict, explain, learn and recommend the next best action from approved context, with confidence and risk made explicit.<\/p>\n<\/div><\/div><\/div>

      <\/div><\/div><\/span>

      Human Command<\/h4><\/span>
      <\/div><\/div><\/div>

      Who approves what, before anything material happens. People lead. The model accelerates.<\/p>\n<\/div><\/div><\/div><\/div>

      Inspect the full architecture →<\/a><\/p>\n<\/div><\/div><\/div><\/div><\/div>

      04 \u00b7 Rules the model cannot outrank<\/p>\n<\/div>

      <\/div><\/div><\/span>

      A fixed order of precedence, and four plain-language flags<\/h2><\/span>
      <\/div><\/div><\/div>

      When instructions conflict, higher outranks lower. A lower rule can never override a higher one, and the task at hand sits last.<\/p>\n<\/div>

      Seven-tier order of precedence<\/title>From highest to lowest: law and platform policy, governance and approval rules, approved company knowledge, approved company strategy, market and channel intelligence, methods and playbooks, and the task at hand.<\/desc>1<\/text>Law, regulation and platform policy<\/text>2<\/text>Governance and approval rules<\/text>3<\/text>Approved company knowledge<\/text>4<\/text>Approved company strategy<\/text>5<\/text>Market, customer and channel intelligence<\/text>6<\/text>Methods, best practices and playbooks<\/text>7<\/text>The task at hand, working within tiers 1 to 6<\/text>Higher outranks lower.<\/text><\/g><\/svg><\/div>\n<\/div>
      \n
        \n
      1. 1 \u00b7 Law and policy<\/strong>Law, regulation and platform policy.<\/li>\n
      2. 2 \u00b7 Governance<\/strong>Governance and approval rules.<\/li>\n
      3. 3 \u00b7 Knowledge<\/strong>Approved company knowledge.<\/li>\n
      4. 4 \u00b7 Strategy<\/strong>Approved company strategy.<\/li>\n
      5. 5 \u00b7 Intelligence<\/strong>Market, customer and channel intelligence.<\/li>\n
      6. 6 \u00b7 Methods<\/strong>Methods, best practices and playbooks.<\/li>\n
      7. 7 \u00b7 Task<\/strong>The task at hand operates within tiers 1\u20136. Higher tiers outrank lower tiers.<\/li>\n<\/ol>\n<\/div>\n<\/div>
        <\/div><\/div><\/span>

        Approval required<\/h4><\/span>
        <\/div><\/div><\/div>

        The action cannot proceed without a named person’s approval.<\/p>\n<\/div><\/div><\/div>

        <\/div><\/div><\/span>

        Review required<\/h4><\/span>
        <\/div><\/div><\/div>

        A person checks the work before it is used.<\/p>\n<\/div><\/div><\/div>

        <\/div><\/div><\/span>

        Compliance check required<\/h4><\/span>
        <\/div><\/div><\/div>

        A compliance or legal review applies. The model flags it and pauses.<\/p>\n<\/div><\/div><\/div>

        <\/div><\/div><\/span>

        Unverified<\/h4><\/span>
        <\/div><\/div><\/div>

        A fact the model could not support. It says so instead of filling the gap.<\/p>\n<\/div><\/div><\/div><\/div><\/div><\/div><\/div><\/div>

        05 \u00b7 Governed decisions<\/p>\n<\/div>

        <\/div><\/div><\/span>

        Recommendations come with evidence, risk and an approver<\/h2><\/span>
        <\/div><\/div><\/div>

        Each material recommendation follows the same record: context, signal, hypothesis, action, confidence and risk, then measurement and learning. Risk decides what the model may do alone.<\/p>\n<\/div>

        Low risk: proceed.<\/strong> Internal summaries, first drafts, ideation, alerts.<\/p>\n

        Medium risk: recommend, then a person reviews.<\/strong> New campaign variants, audience ideas, moderate budget shifts.<\/p>\n

        High risk: flag and pause.<\/strong> Strategy shifts, regulated claims, major budget moves, sensitive segments.<\/p>\n<\/div><\/div><\/div>

        \n

        Illustrative decision record<\/p>\n

        Constructed to show the mechanism. Not a client record. Thresholds are examples.<\/p>\n

        Context:<\/strong> a cost-per-lead target is set. Budget moves up to 5% proceed, 5 to 10% are reviewed, above 10% need explicit approval.<\/p>\n

        Signal:<\/strong> cost per lead has drifted above baseline for two weeks. Attribution is platform-reported.<\/p>\n

        Hypothesis:<\/strong> shifting 8% of budget to high-intent terms lowers cost per lead. Test against a holdout for 14 days.<\/p>\n

        Confidence and risk:<\/strong> platform-reported evidence, moderate confidence. 8% exceeds the automatic threshold, so a named approver confirms in writing.<\/p>\n

        Learning:<\/strong> the result is compared with the holdout and logged. The playbook changes only if the result repeats.<\/p>\n<\/div>\n<\/div><\/div><\/div><\/div><\/div>

        06 \u00b7 Tested before release<\/p>\n<\/div>

        <\/div><\/div><\/span>

        Safety behaviors are tested repeatedly, not once<\/h2><\/span>
        <\/div><\/div><\/div>

        The framework ships with a structured test matrix across every module. Tests are classed by what is at stake, and the classes carry different bars.<\/p>\n<\/div>

        <\/div><\/div><\/span>

        Safety-class tests<\/h4><\/span>
        <\/div><\/div><\/div>

        Approvals, claims, privacy, budget limits and evidence handling. These are repeated across several runs before a model is released.<\/p>\n<\/div><\/div><\/div>

        <\/div><\/div><\/span>

        Quality-class tests<\/h4><\/span>
        <\/div><\/div><\/div>

        Structure, clarity, use of brand terminology and completeness of deliverables. These are reviewed and improved continuously.<\/p>\n<\/div><\/div><\/div><\/div>

        When the framework changes, the affected tests run again.<\/strong> After launch, live outputs are sampled for human review. Test logs are shared during onboarding or a formal review, not published.<\/p>\n<\/div><\/div><\/div><\/div><\/div>

        07 \u00b7 Portable by design<\/p>\n<\/div>

        <\/div><\/div><\/span>

        Model-agnostic, so your knowledge stays yours<\/h2><\/span>
        <\/div><\/div><\/div>

        mAI is an intelligence and operating layer. It orchestrates and governs leading third-party AI models. marktgAI does not develop or train foundation models. See the platforms and workflows this connects to in Platform integrations<\/a>.<\/p>\n<\/div><\/div><\/div>

        What stays with you when the model changes:<\/strong><\/p>\n