ai-agentic-marketing-workflows
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseAI Agentic Marketing Workflows
AI自主营销工作流
<!-- dual-compat:start -->
<!-- dual-compat:start -->
Use when
适用场景
- Design and implement autonomous AI marketing agent systems using the PRAL, BDI, and OODA frameworks. Invoke when a client is ready to move beyond reactive GenAI prompting to proactive, autonomous marketing workflows, or when planning an AI-first marketing operations architecture.
- Use this skill when it is the closest match to the requested deliverable or workflow.
- 使用PRAL、BDI和OODA框架设计并实现自主AI营销Agent系统。当客户准备从被动式GenAI提示转向主动、自主的营销工作流，或规划AI优先的营销运营架构时，启用此方案。
- 当此技能与请求的交付成果或工作流最匹配时使用。
Do not use when
不适用场景
- Do not use this skill for graphic design, video production, software development, or legal advice beyond the repository's stated scope.
- Do not use it when another skill in this repository is clearly more specific to the requested deliverable.
- 请勿将此技能用于图形设计、视频制作、软件开发或超出本知识库规定范围的法律咨询。
- 当知识库中的其他技能明显更符合请求的交付成果时，请勿使用此技能。
Workflow
工作流
- Collect the required inputs or source material before drafting, unless this skill explicitly generates the intake itself.
- Follow the section order and decision rules in this ; do not skip mandatory steps or required fields.
SKILL.md - Review the draft against the quality criteria, then deliver the final output in markdown unless the skill specifies another format.
- 在起草前收集所需输入或源材料，除非此技能明确说明自行生成输入内容。
- 遵循本中的章节顺序和决策规则；请勿跳过必填步骤或必填字段。
SKILL.md - 根据质量标准审核草稿，然后以markdown格式交付最终输出，除非技能指定其他格式。
Anti-Patterns
反模式
- Do not invent client facts, performance data, budgets, or approvals that were not provided or clearly inferred from evidence.
- Do not skip required inputs, mandatory sections, or quality checks just to make the output shorter.
- Do not drift into out-of-scope work such as code implementation, design production, or unsupported legal conclusions.
- 请勿编造未提供或无法从证据中明确推断的客户事实、绩效数据、预算或审批信息。
- 请勿为了缩短输出内容而跳过必填输入、必填章节或质量检查。
- 请勿偏离范围外工作，例如代码实现、设计制作或无依据的法律结论。
Outputs
输出成果
- An AI-focused strategy, audit, system design, or prompt asset in markdown with human review and control points.
- 经过人工审核并包含管控节点的AI聚焦策略、审计报告、系统设计或提示资产，格式为markdown。
References
参考资料
- Use the inline instructions in this skill now. If a directory is added later, treat its files as the deeper source material and keep this
references/execution-focused.SKILL.md
- 目前使用本技能中的内联说明。若后续添加目录，将其文件视为深度源材料，并保持本
references/以执行为核心。SKILL.md
Purpose
目标
Design and document autonomous AI marketing agent systems that perceive their
environment, reason about what action to take, act without human prompting, and
learn from outcomes. Output is an architecture specification, a chosen workflow
template with full HITL safeguards, and a wave-appropriate implementation plan.
This skill assumes the client has completed and has
an AI maturity wave score (1, 2, or 3). Do not recommend Wave 3 architecture
to a Wave 1 client without a phased roadmap.
ai-readiness-diagnostic设计并记录自主AI营销Agent系统，这类系统能够感知环境、推理应采取的行动、无需人工提示即可执行，并从结果中学习。输出内容包括架构规范、带有完整HITL（人机协作）保障措施的选定工作流模板，以及符合成熟度阶段的实施计划。
本技能假定客户已完成（AI就绪诊断），并拥有AI成熟度阶段评分（1、2或3）。若没有分阶段路线图，请勿向阶段1的客户推荐阶段3的架构。
ai-readiness-diagnosticRequired Inputs
必填输入
Ask for all of the following before generating any output:
- Client business name — trading name and legal entity if different
- Industry — sector, product or service type
- Country / city — defaults to Uganda if not specified
- Current AI maturity wave — Wave 1, 2, or 3 (from ; estimate if not available)
ai-readiness-diagnostic - Target workflow to automate — select one primary: content / sentiment monitoring / reporting / customer service / campaign optimisation
- Available technical resources — none / basic (can use no-code tools) / developer (can call APIs and self-host)
在生成任何输出前，需向客户询问以下所有信息：
- 客户企业名称 — 交易名称及不同的法律实体名称（如有）
- 行业 — 行业领域、产品或服务类型
- 国家/城市 — 未指定时默认乌干达
- 当前AI成熟度阶段 — 阶段1、2或3（来自；若未提供则进行估算）
ai-readiness-diagnostic - 待自动化的目标工作流 — 选择一个主要方向：内容创作/舆情监测/报告生成/客户服务/营销活动优化
- 可用技术资源 — 无/基础（可使用无代码工具）/开发人员（可调用API并自行托管）
Agentic vs Generative AI: The Critical Distinction
自主AI与生成式AI：关键区别
Source: Nayebi (2025)
Most clients conflate generative AI with agentic AI. Clarify the distinction
before designing any architecture.
Generative AI is reactive.
It waits for a human prompt, generates output, then stops. Every action
requires a human to initiate the cycle. This is Wave 1 and Wave 2 behaviour.
Agentic AI is proactive.
It monitors its environment continuously, reasons about what action to take,
executes that action, and updates its behaviour based on outcomes — without
waiting for a human prompt. This is Wave 3 behaviour.
This distinction determines the architecture, the tools, the data requirements,
and the risk controls needed. A business without clean engagement data or
developer resource is not ready for a full agentic stack.
Wave guidance:
- Wave 1 clients → rule-based automation (Zapier / Make.com triggers)
- Wave 2 clients → performance-triggered AI actions using analytics data
- Wave 3 clients → full PRAL agents with continuous monitoring and learning
Most East African businesses should reach Wave 2 (Predictive ML, with 3+ months
of clean data) before building Wave 3 agents.
来源：Nayebi (2025)
大多数客户会混淆生成式AI与自主AI。在设计任何架构前，需明确区分两者。
生成式AI是被动的
它等待人工提示，生成输出后即停止。每一项行动都需要人类启动循环。这属于阶段1和阶段2的行为。
自主AI是主动的
它持续监控环境，推理应采取的行动，执行该行动，并根据结果更新行为 — 无需等待人工提示。这属于阶段3的行为。
这种区别决定了所需的架构、工具、数据要求和风险管控措施。没有清晰互动数据或开发资源的企业尚未准备好搭建完整的自主AI栈。
阶段指导：
- 阶段1客户 → 基于规则的自动化（Zapier/Make.com触发器）
- 阶段2客户 → 利用分析数据触发AI行动的性能驱动方案
- 阶段3客户 → 具备持续监控和学习能力的完整PRAL Agent
大多数东非企业应先达到阶段2（预测机器学习，拥有3个月以上的清晰数据），再搭建阶段3的Agent。
The PRAL Loop
PRAL循环
Source: Nayebi (2025)
Every agentic system is built on the PRAL loop:
Perceive → Reason → Act → Learn
Map the client's chosen workflow to each stage before recommending tools.
| Stage | What the agent does | Marketing example |
|---|---|---|
| Perceive | Gathers data from its environment | Scans social mentions, reads engagement metrics, receives inbound WhatsApp messages |
| Reason | Processes data and decides what to do — using an LLM or rule-based logic | Classifies sentiment, identifies a content gap, detects a campaign underperforming |
| Act | Executes the decision | Drafts content, sends an alert, triggers a campaign boost, routes a message to a human |
| Learn | Updates its behaviour based on outcomes | Feeds performance data back into the next Perceive cycle; adjusts thresholds and templates |
Apply the PRAL loop explicitly when designing each workflow template. Label
each step so the client can see where human oversight sits.
来源：Nayebi (2025)
每个自主AI系统都基于PRAL循环构建：
感知（Perceive）→ 推理（Reason）→ 执行（Act）→ 学习（Learn）
在推荐工具前，将客户选定的工作流映射到每个阶段。
| 阶段 | Agent的行为 | 营销示例 |
|---|---|---|
| 感知 | 从环境中收集数据 | 扫描社交提及内容、读取互动指标、接收WhatsApp inbound消息 |
| 推理 | 处理数据并决定行动方案 — 使用LLM或基于规则的逻辑 | 分类舆情、识别内容缺口、检测表现不佳的营销活动 |
| 执行 | 执行决策 | 起草内容、发送警报、触发营销活动推广、将消息转交给人工 |
| 学习 | 根据结果更新行为 | 将绩效数据反馈到下一个感知循环；调整阈值和模板 |
在设计每个工作流模板时，需明确应用PRAL循环。标记每个步骤，以便客户了解人工监督的位置。
The BDI Model for Marketing Agents
适用于营销Agent的BDI模型
Source: Nayebi (2025)
The BDI model — Beliefs, Desires, Intentions — maps naturally to marketing
strategy and is the clearest way to define an agent's decision boundary.
| Component | Definition | Marketing application |
|---|---|---|
| Beliefs | What the agent knows | Audience data, engagement history, brand guidelines, competitor positions, product catalogue |
| Desires | What the agent is trying to achieve | Business goals — leads, awareness, retention, revenue — expressed as KPIs |
| Intentions | How the agent plans to act | Campaign tactics, content formats, channel choices, timing rules, escalation thresholds |
Prompt to use with client:
"Specify your agent's Beliefs (what data it has access to), Desires (what KPI it optimises for), and Intentions (what actions it can take). This defines the agent's decision boundary."
Document the BDI model before selecting any tool. An agent without a defined
decision boundary will act unpredictably.
来源：Nayebi (2025)
BDI模型 — 信念（Beliefs）、欲望（Desires）、意图（Intentions） — 与营销策略自然契合，是定义Agent决策边界最清晰的方式。
| 组件 | 定义 | 营销应用 |
|---|---|---|
| 信念 | Agent所掌握的信息 | 受众数据、互动历史、品牌指南、竞品定位、产品目录 |
| 欲望 | Agent试图达成的目标 | 业务目标 — 线索、知名度、留存率、收入 — 以KPI形式表达 |
| 意图 | Agent的行动规划 | 营销活动策略、内容格式、渠道选择、时间规则、升级阈值 |
用于与客户沟通的提示语：
"请明确您的Agent的信念（可访问的数据）、欲望（优化的KPI）和意图（可采取的行动）。这将定义Agent的决策边界。"
在选择任何工具前，需记录BDI模型。没有明确决策边界的Agent行为会不可预测。
The OODA Cycle for Real-Time Decisions
用于实时决策的OODA循环
Borrowed from military strategy (Boyd, 1976), OODA is the fastest decision loop
applicable to marketing agents operating in real-time social media environments.
Observe → Orient → Decide → Act
Faster OODA cycles = competitive advantage in fast-moving social media
environments where a delayed crisis response or missed trend costs engagement.
Social listening agent example:
- Observe — scan all mentions of the brand across Facebook, Instagram, X/Twitter, and Google every hour
- Orient — classify each mention by sentiment (positive / neutral / negative / crisis) and topic category
- Decide — apply rules: respond autonomously to positive enquiries; escalate negative mentions; flag crisis keywords immediately
- Act — post pre-approved response template, or send alert to human via WhatsApp/email with full context
OODA complements PRAL: PRAL describes the agent's architecture; OODA describes
the speed and logic of its decision-making in a single cycle.
源自军事战略（Boyd, 1976），OODA是适用于在实时社交媒体环境中运行的营销Agent的最快决策循环。
观察（Observe）→ 定位（Orient）→ 决策（Decide）→ 执行（Act）
更快的OODA循环 = 在快速变化的社交媒体环境中获得竞争优势，延迟的危机响应或错过趋势会损失互动量。
社交监听Agent示例：
- 观察 — 每小时扫描Facebook、Instagram、X/Twitter和Google上的所有品牌提及内容
- 定位 — 将每个提及内容按舆情（正面/中性/负面/危机）和话题类别分类
- 决策 — 应用规则：自主回复正面咨询；升级负面提及内容；立即标记危机关键词
- 执行 — 发布预先批准的回复模板，或通过WhatsApp/电子邮件向人工发送包含完整上下文的警报
OODA与PRAL互补：PRAL描述Agent的架构；OODA描述其在单个循环中的决策速度和逻辑。
Five Agentic Workflow Templates
五种自主工作流模板
Select the template that matches the client's target workflow. Fully specify
the chosen template before recommending tools.
选择与客户目标工作流匹配的模板。在推荐工具前，需完整指定所选模板。
1. Content Pipeline Agent
1. 内容流水线Agent
What it does: Automates the content creation and publishing pipeline from
trend detection to post-performance feedback.
| Element | Detail |
|---|---|
| Trigger | Scheduled (daily/weekly) or event-driven (trending topic detected) |
| Actions | 1. Monitor trending topics and competitor content · 2. Generate draft content (caption, hashtags, image brief) · 3. Route draft to human for approval · 4. Publish approved content at optimal time · 5. Monitor post performance for 48 hours |
| HITL point | Human approves every draft before publishing — no autonomous publishing without review |
| Learn step | Performance data (reach, engagement rate, saves) fed back to refine future prompts and posting times |
| Tools | Claude API (drafting) + n8n or Make.com (orchestration) + Buffer/Hootsuite (scheduling) |
| EA feasibility | High — Wave 2 clients can implement with no-code tools |
功能： 自动化从趋势检测到帖子绩效反馈的内容创作和发布流程。
| 要素 | 详情 |
|---|---|
| 触发条件 | 定时触发（每日/每周）或事件驱动（检测到热门话题） |
| 行动 | 1. 监控热门话题和竞品内容 · 2. 生成草稿内容（标题、话题标签、图片 brief） · 3. 将草稿提交给人工审批 · 4. 在最佳时间发布已审批内容 · 5. 监控帖子48小时内的绩效 |
| HITL节点 | 人工在发布前审批所有草稿 — 未经审核不得自主发布 |
| 学习步骤 | 将绩效数据（触达量、互动率、保存量）反馈，以优化未来的提示语和发布时间 |
| 工具 | Claude API（起草）+ n8n或Make.com（编排）+ Buffer/Hootsuite（调度） |
| 东非可行性 | 高 — 阶段2客户可使用无代码工具实现 |
2. Sentiment Monitoring Agent
2. 舆情监测Agent
What it does: Continuously scans social mentions, classifies sentiment, and
alerts the team when a threshold is crossed.
| Element | Detail |
|---|---|
| Trigger | Continuous (hourly scan) or keyword-event (brand name mentioned) |
| Actions | 1. Scan Facebook, Instagram, X/Twitter, Google reviews for brand mentions · 2. Classify mention: positive / neutral / negative / crisis · 3. Log all mentions in dashboard · 4. Alert team when negative threshold crossed (e.g., 3+ negative mentions in one hour) · 5. Suggest pre-approved response options |
| HITL point | Human selects and sends response — agent does not post responses autonomously |
| Learn step | Mis-classifications flagged by human; agent updates sentiment rules |
| Tools | Mention.com or Google Alerts (listening) + Claude API (classification) + n8n (routing) + WhatsApp Business API (alert delivery) |
| EA feasibility | High — Google Alerts + Claude API is accessible and low-cost |
功能： 持续扫描社交提及内容，分类舆情，并在阈值被突破时向团队发送警报。
| 要素 | 详情 |
|---|---|
| 触发条件 | 持续触发（每小时扫描）或关键词事件（提及品牌名称） |
| 行动 | 1. 扫描Facebook、Instagram、X/Twitter、Google评论中的品牌提及内容 · 2. 分类提及内容：正面/中性/负面/危机 · 3. 在仪表盘中记录所有提及内容 · 4. 负面提及内容突破阈值时（如1小时内出现3条以上负面提及）向团队发送警报 · 5. 建议预先批准的回复选项 |
| HITL节点 | 人工选择并发送回复 — Agent不得自主发布回复 |
| 学习步骤 | 人工标记分类错误的内容；Agent更新舆情规则 |
| 工具 | Mention.com或Google Alerts（监听）+ Claude API（分类）+ n8n（路由）+ WhatsApp Business API（警报推送） |
| 东非可行性 | 高 — Google Alerts + Claude API易于获取且成本低 |
3. Proactive Campaign Agent
3. 主动营销活动Agent
What it does: Monitors engagement metrics and triggers a targeted response
campaign when performance drops below threshold.
| Element | Detail |
|---|---|
| Trigger | Metric threshold (engagement rate drops below X%, or follower growth stalls for N days) |
| Actions | 1. Pull platform analytics daily · 2. Compare against baseline benchmarks · 3. Detect underperformance · 4. Generate campaign response options (content boost, new format, re-engagement post) · 5. Present options to human for approval · 6. Execute approved option · 7. Report results after 7 days |
| HITL point | Human approves the campaign response before any content is published |
| Learn step | Successful response tactics stored; agent prioritises them in future recommendations |
| Tools | Platform analytics API + Claude API (analysis and drafting) + Make.com (orchestration) + Buffer (publishing) |
| EA feasibility | Medium — requires Wave 2 data maturity and API access to platform analytics |
功能： 监控互动指标，当绩效低于阈值时触发针对性响应活动。
| 要素 | 详情 |
|---|---|
| 触发条件 | 指标阈值（互动率低于X%，或粉丝增长停滞N天） |
| 行动 | 1. 每日提取平台分析数据 · 2. 与基准指标对比 · 3. 检测绩效不佳情况 · 4. 生成营销活动响应选项（内容推广、新格式、重新互动帖子） · 5. 将选项提交给人工审批 · 6. 执行已批准选项 · 7. 7天后报告结果 |
| HITL节点 | 人工在发布任何内容前审批营销活动响应方案 |
| 学习步骤 | 存储成功的响应策略；Agent在未来推荐中优先选择这些策略 |
| 工具 | 平台分析API + Claude API（分析和起草）+ Make.com（编排）+ Buffer（发布） |
| 东非可行性 | 中 — 需要阶段2的数据成熟度和平台分析API访问权限 |
4. Multi-Agent Reporting System
4. 多Agent报告系统
What it does: A team of specialised agents collaborates to produce the
monthly performance report with minimal human effort.
| Element | Detail |
|---|---|
| Trigger | Scheduled (last day of the month) |
| Actions | 1. Data agent — pulls platform statistics from all active channels · 2. Analysis agent — identifies patterns, anomalies, and top-performing content · 3. Writing agent — drafts narrative report with insights and recommendations · 4. Human consultant — reviews, edits, and presents to client |
| HITL point | Human reviews the full draft before delivery; no automated client-facing report |
| Learn step | Human edits tracked; writing agent refines its narrative style and recommendation quality |
| Tools | Platform APIs (data) + Claude API (analysis and writing) + n8n (orchestration) + Google Docs / Notion (output) |
| EA feasibility | Medium — high value but requires API access and developer setup for data pulls |
功能： 一组专业Agent协作生成月度绩效报告，只需最少的人工参与。
| 要素 | 详情 |
|---|---|
| 触发条件 | 定时触发（每月最后一天） |
| 行动 | 1. 数据Agent — 从所有活跃渠道提取平台统计数据 · 2. 分析Agent — 识别模式、异常和表现最佳的内容 · 3. 写作Agent — 起草包含洞察和建议的叙事报告 · 4. 人工顾问 — 审核、编辑并向客户展示报告 |
| HITL节点 | 人工在交付前审核完整草稿；不得自动生成面向客户的报告 |
| 学习步骤 | 跟踪人工编辑内容；写作Agent优化叙事风格和建议质量 |
| 工具 | 平台API（数据）+ Claude API（分析和写作）+ n8n（编排）+ Google Docs/Notion（输出） |
| 东非可行性 | 中 — 价值高，但需要API访问权限和开发人员设置数据提取 |
5. WhatsApp Response Agent
5. WhatsApp响应Agent
What it does: Classifies inbound WhatsApp messages, routes them to the
correct response path, and handles routine enquiries autonomously.
| Element | Detail |
|---|---|
| Trigger | Inbound WhatsApp Business message received |
| Actions | 1. Receive and classify message (enquiry / complaint / order / other) · 2. Route to: decision tree (simple FAQ) / Claude API (nuanced enquiry) / human agent (complaint or high value) · 3. Respond or escalate · 4. Log interaction with timestamp and classification |
| HITL point | All complaints and high-value sales enquiries routed to human immediately; agent does not resolve complaints autonomously |
| Learn step | Mis-routed messages flagged; classification rules updated monthly |
| Tools | WhatsApp Business API + Claude API (classification and response drafting) + n8n (routing logic) |
| EA feasibility | High — WhatsApp penetration in EA makes this the highest-ROI agentic workflow for most clients |
功能： 分类WhatsApp inbound消息，将其路由到正确的响应路径，并自主处理常规咨询。
| 要素 | 详情 |
|---|---|
| 触发条件 | 收到WhatsApp Business消息 |
| 行动 | 1. 接收并分类消息（咨询/投诉/订单/其他） · 2. 路由至：决策树（简单FAQ）/ Claude API（复杂咨询）/ 人工Agent（投诉或高价值咨询） · 3. 回复或升级 · 4. 记录互动的时间戳和分类 |
| HITL节点 | 所有投诉和高价值销售咨询立即路由给人工；Agent不得自主解决投诉 |
| 学习步骤 | 标记路由错误的消息；每月更新分类规则 |
| 工具 | WhatsApp Business API + Claude API（分类和回复起草）+ n8n（路由逻辑） |
| 东非可行性 | 高 — WhatsApp在东非的渗透率高，这是大多数客户投资回报率最高的自主工作流 |
HITL Safeguard Design
HITL保障措施设计
Source: Nayebi (2025)
Every agentic workflow must define four safeguard components before going live.
Include this section in every workflow specification delivered to the client.
1. Autonomous decision boundary
Define what the agent can decide and act on without human review. Limit this to
decisions that are: low-risk, routine, reversible, and within a defined value
threshold (e.g., scheduling a post, classifying a mention, logging a message).
2. Escalation triggers
Define what forces the agent to stop and wait for a human. Escalation is
mandatory when a decision is: high-risk, irreversible (e.g., publishing to
public), sensitive (crisis keywords, complaints, legal mentions), or above a
value threshold (e.g., enquiry worth over UGX 500,000).
3. Escalation mechanism
Specify: how the human is alerted (WhatsApp message, email, Slack), what
information they receive (full context, agent's recommended options), the
expected response time, and the override protocol if no response is received.
4. Audit trail
Every agent action must be logged with: timestamp, action taken, data that
triggered the action, reasoning or rule applied, and outcome. This log is
reviewed monthly to improve agent performance and demonstrate accountability.
来源：Nayebi (2025)
每个自主工作流在上线前必须定义四个保障组件。在交付给客户的每个工作流规范中包含此部分。
1. 自主决策边界
定义Agent无需人工审核即可决定并执行的内容。将其限制为低风险、常规、可逆且在规定价值阈值内的决策（如调度帖子、分类提及内容、记录消息）。
2. 升级触发条件
定义迫使Agent停止并等待人工介入的情况。当决策属于高风险、不可逆（如发布到公开渠道）、敏感（危机关键词、投诉、法律提及）或超出价值阈值（如价值超过500,000乌干达先令的咨询）时，必须进行升级。
3. 升级机制
明确：如何向人工发送警报（WhatsApp消息、电子邮件、Slack）、人工接收的信息（完整上下文、Agent推荐的选项）、预期响应时间，以及未收到响应时的覆盖协议。
4. 审计追踪
每个Agent的行动必须记录以下信息：时间戳、执行的行动、触发行动的数据、应用的推理或规则、结果。每月审核此日志以提升Agent性能并展示问责制。
Three-Wave Implementation Roadmap
三阶段实施路线图
Match the roadmap recommendation to the client's current wave.
| Wave | Readiness criteria | What to build | Effort |
|---|---|---|---|
| Wave 1 — Automation | No analytics data required; any technical level | Zapier or Make.com automations that trigger AI content drafts on a schedule. Rule-based, no learning, no API calls. | 1–2 days setup |
| Wave 2 — Performance-triggered | 3+ months of clean engagement data; basic technical resource | Connect analytics data to AI for performance-triggered actions (e.g., engagement drop → draft new content). Requires platform data export or basic API access. | 1–2 weeks setup |
| Wave 3 — Full agentic | Clean data, developer resource, HITL safeguards in place | Full PRAL agents with continuous monitoring, LLM reasoning, and feedback loops. Requires API access, self-hosted orchestration (n8n), and ongoing maintenance. | 4–8 weeks minimum |
Do not propose Wave 3 to a Wave 1 client without a phased roadmap that moves
them through Wave 2 first.
根据客户当前的阶段匹配路线图建议。
| 阶段 | 就绪标准 | 构建内容 | 工作量 |
|---|---|---|---|
| 阶段1 — 自动化 | 无需分析数据；任何技术水平 | Zapier或Make.com自动化，定时触发AI内容草稿。基于规则，无学习功能，无API调用。 | 1–2天设置 |
| 阶段2 — 绩效驱动 | 3个月以上的清晰互动数据；基础技术资源 | 将分析数据与AI连接，实现绩效触发的行动（如互动量下降 → 起草新内容）。需要平台数据导出或基础API访问权限。 | 1–2周设置 |
| 阶段3 — 完整自主 | 清晰数据、开发资源、已落实HITL保障措施 | 具备持续监控、LLM推理和反馈循环的完整PRAL Agent。需要API访问权限、自行托管的编排工具（n8n）和持续维护。 | 至少4–8周 |
若没有分阶段路线图，请勿向阶段1的客户提议阶段3的方案。
Tool Stack Options
工具栈选项
Recommend tools based on the client's technical resources and budget.
| Tool | Role in agentic stack | EA accessibility | Approx. cost |
|---|---|---|---|
| Claude API | LLM reasoning layer — classification, drafting, analysis | Yes — API account required | Pay-per-token |
| n8n | Workflow orchestration (self-hostable, open source) | Yes — developer resource needed for self-hosting | Free (self-hosted); from $20/month (cloud) |
| Zapier AI | No-code workflow automation with AI steps | Yes — browser only, no dev required | Free tier; from $19.99/month |
| Make.com | Visual no-code workflow builder | Yes — browser only, no dev required | Free tier; from $9/month |
| Hootsuite / Buffer | Publishing and scheduling layer | Yes — widely used in EA | From $15/month |
| Brandwatch / Mention | Social listening layer for sentiment monitoring | Limited — pricing is a barrier for small clients | From $99/month |
| WhatsApp Business API | Inbound message routing and response | Yes — high EA penetration; via Meta or third-party | From $0 (first 1,000 conversations/month free) |
For Wave 1 clients with no technical resource: Zapier or Make.com + Claude
(via ChatGPT or Claude.ai interface, not API) is the most accessible entry point.
For Wave 3 clients with developer resource: n8n (self-hosted) + Claude API is
the recommended EA-feasible stack.
根据客户的技术资源和预算推荐工具。
| 工具 | 在自主AI栈中的角色 | 东非可获取性 | 大致成本 |
|---|---|---|---|
| Claude API | LLM推理层 — 分类、起草、分析 | 是 — 需要API账户 | 按token付费 |
| n8n | 工作流编排（可自行托管，开源） | 是 — 自行托管需要开发资源 | 免费（自行托管）；每月20美元起（云端） |
| Zapier AI | 带AI步骤的无代码工作流自动化 | 是 — 仅需浏览器，无需开发人员 | 免费层；每月19.99美元起 |
| Make.com | 可视化无代码工作流构建器 | 是 — 仅需浏览器，无需开发人员 | 免费层；每月9美元起 |
| Hootsuite/Buffer | 发布和调度层 | 是 — 在东非广泛使用 | 每月15美元起 |
| Brandwatch/Mention | 舆情监测的社交监听层 | 有限 — 定价对小型客户来说是障碍 | 每月99美元起 |
| WhatsApp Business API | 入站消息路由和响应 | 是 — 在东非渗透率高；通过Meta或第三方获取 | 每月前1000条对话免费，之后收费 |
对于无技术资源的阶段1客户：Zapier或Make.com + Claude（通过ChatGPT或Claude.ai界面，而非API）是最易获取的入门方案。
对于有开发资源的阶段3客户：n8n（自行托管）+ Claude API是推荐的适合东非的工具栈。
Quality Criteria
质量标准
Output meets standard when it satisfies all of the following:
- Client's current AI maturity wave is identified and a wave-appropriate architecture is recommended — no Wave 3 proposal for a Wave 1 client without a phased roadmap
- At least one agentic workflow template is selected and fully specified: trigger, PRAL mapping, actions, HITL point, learn step, and tools
- The PRAL loop is explicitly mapped for the chosen workflow — each stage labelled
- The BDI model is documented: Beliefs (data sources), Desires (KPI optimised), and Intentions (actions the agent can take)
- HITL safeguards are defined: autonomous decision boundary, escalation triggers, escalation mechanism, and audit trail
- Tool stack is recommended based on the client's technical resources and budget with EA accessibility noted
- EA feasibility is assessed — WhatsApp-based and no-code workflows prioritised for clients without developer resource
输出符合标准需满足以下所有条件：
- 明确客户当前的AI成熟度阶段，并推荐符合该阶段的架构 — 若无分阶段路线图，不得向阶段1客户提议阶段3方案
- 选择至少一个自主工作流模板并完整指定：触发条件、PRAL映射、行动、HITL节点、学习步骤和工具
- 为选定的工作流明确映射PRAL循环 — 标记每个阶段
- 记录BDI模型：信念（数据源）、欲望（优化的KPI）和意图（Agent可采取的行动）
- 定义HITL保障措施：自主决策边界、升级触发条件、升级机制和审计追踪
- 根据客户的技术资源和预算推荐工具栈，并注明东非可获取性
- 评估东非可行性 — 优先推荐基于WhatsApp和无代码的工作流给无开发资源的客户
References
参考资料
- Nayebi, F. (2025) Foundations of Agentic AI for Retail. Gradient Divergence.
- Venkatesan, R. and Lecinski, J. (2026) The AI Marketing Canvas, 2nd edn. Stanford University Press.
- Farri, E. and Rosani, G. (2025) HBR Guide to Generative AI for Managers. Harvard Business Review Press.
- Nayebi, F. (2025) Foundations of Agentic AI for Retail. Gradient Divergence.
- Venkatesan, R. and Lecinski, J. (2026) The AI Marketing Canvas, 2nd edn. Stanford University Press.
- Farri, E. and Rosani, G. (2025) HBR Guide to Generative AI for Managers. Harvard Business Review Press.