There’s a difference between using AI and using AI well.<\/p>\n
Most project managers are in the first camp. They open ChatGPT, type a question, get an answer, and move on. That works for simple tasks \u2014 but it misses what AI is actually capable of when you understand how it reasons.<\/p>\n
Two frameworks changed how I use AI in project delivery: ReAct and Chain-of-Thought (CoT). Neither requires you to be a developer. But understanding them will make you a significantly better AI user \u2014 and a more effective PM.<\/p>\n
Chain-of-Thought is a prompting technique that guides an AI through a reasoning process step by step, rather than jumping straight to an answer.<\/p>\n
Think of it as the difference between asking a junior analyst “what’s your recommendation?” versus asking them to walk you through their analysis first, then give you a recommendation. The second approach produces better output \u2014 and the same is true for AI.<\/p>\n
With CoT, instead of prompting:<\/p>\n
“What are the risks of this software integration project?”<\/p>\n
You prompt:<\/p>\n
“Walk me through the risks of this software integration project step by step. First, identify the technical risks. Then the stakeholder risks. Then the timeline risks. Finally, rank them by probability and impact and give me your top recommendation.”<\/p>\n
The AI reasons through each layer before concluding. The output is dramatically more useful.<\/p>\n
CoT directly maps to how good project management thinking works. We don’t jump to conclusions \u2014 we analyze, sequence, and reason through complexity.<\/p>\n
The three things CoT improves:<\/p>\n
Accuracy \u2014 breaking down problems into smaller steps reduces the chance the AI misses something. I’ve used this for risk registers, and the structured approach catches edge cases a simple prompt would miss.<\/p>\n
Reasoning quality \u2014 the AI’s “thinking” becomes visible. You can see where it’s making assumptions and correct them before they become deliverables.<\/p>\n
Explainability \u2014 when you need to present AI-assisted analysis to a stakeholder, CoT gives you a reasoning trail. You’re not just presenting an answer \u2014 you’re presenting a process.<\/p>\n
When I use CoT for project planning, my prompt structure looks like this:<\/p>\n
“Think through this step by step. [Context about the project]. Step 1: identify the key dependencies. Step 2: map the critical path. Step 3: identify the top 3 risks to the critical path. Step 4: recommend mitigation strategies for each. Present your reasoning at each step before moving to the next.”<\/p>\n
That last line \u2014 “present your reasoning at each step” \u2014 is the key trigger for CoT behavior. It forces the AI to show its work.<\/p>\n
ReAct (Reasoning + Acting) takes CoT further. While CoT focuses on internal reasoning, ReAct allows AI to interact with the external world \u2014 searching the web, querying databases, using tools \u2014 and incorporate that information into its reasoning.<\/p>\n
The cycle is:<\/p>\n
Think \u2192 the AI reasons about the problem
\nAct \u2192 the AI takes an action (web search, database query, tool use)
\nObserve \u2192 the AI receives results from that action
\nRespond \u2192 the AI generates a response, or loops back to Think<\/p>\n
This is what’s happening when you use an AI agent with web browsing enabled, or when Claude or ChatGPT can access real-time data.<\/p>\n
ReAct is what transforms AI from a knowledge tool into a delivery tool. Instead of the AI reasoning only with what it already knows, it can go get information \u2014 and that changes what’s possible.<\/p>\n
At Bartronics, we integrated AI agents that use ReAct-style loops to monitor data sources, detect changes, and trigger actions automatically. The PM role shifts from executing tasks to designing the decision logic the agent follows.<\/p>\n
Three applications where ReAct changes the game for project managers:<\/p>\n
Real-time status reporting \u2014 an AI agent can pull data from Jira, Monday.com, or your project tool, reason about it, and generate a status report without you building it manually.<\/p>\n
Risk monitoring \u2014 an agent can monitor external signals (vendor announcements, regulatory changes, market shifts) and flag risks proactively, rather than waiting for you to discover them.<\/p>\n
Stakeholder research \u2014 before a client meeting, an agent can gather current information about the client’s business, recent news, and industry context, and brief you in minutes.<\/p>\n
They solve different problems, and the best approach often combines both.<\/p>\n
Use CoT when:<\/p>\n
You need structured reasoning on information you already have
\nYou’re generating documents: risk registers, project charters, stakeholder plans
\nYou want the AI to think through a complex decision before giving you a recommendation
\nYou need to show your reasoning process to a stakeholder<\/p>\n
Use ReAct when:<\/p>\n
You need current or external information
\nYou’re building or using AI agents that interact with tools
\nYou want the AI to take actions, not just generate text
\nYou’re automating a workflow that involves multiple data sources<\/p>\n
Use both when:<\/p>\n
You want an agent that can gather information AND reason carefully about it before acting
\nYou’re building automation that needs to make nuanced decisions based on real-time data<\/p>\n
This combination \u2014 gather information with ReAct, reason about it with CoT \u2014 is what separates basic AI use from genuinely AI-augmented delivery.<\/p>\n
At Capgemini, where I manage B2B API integrations for Cisco, I use CoT-style prompting for everything that requires structured thinking: risk assessment, stakeholder communication drafts, requirement analysis. The outputs go directly into our documentation workflow.<\/p>\n
At Bartronics, we’ve built ReAct-style agents into our automation infrastructure. The agents monitor data sources, apply reasoning logic, and trigger actions \u2014 the 120 bots running in our system operate on this principle. My role is to define the decision logic and review edge cases, not to execute the repetitive steps.<\/p>\n
The result: I can manage a multi-project program at the complexity level of Bartronics while simultaneously handling senior PM responsibilities at Capgemini. That’s not superhuman capacity \u2014 it’s AI leverage applied deliberately.<\/p>\n
You don’t need to build agents to benefit from these frameworks right now.<\/p>\n
Start with CoT in your next complex AI prompt. Add the phrase: “Think through this step by step and show your reasoning at each stage before giving me a final answer.”<\/p>\n
Run it on your next risk assessment, your next project planning session, or your next stakeholder communication challenge. Compare the output to your usual prompt. The difference will be immediate.<\/p>\n
Once you’ve internalized CoT, explore AI tools with agent capabilities \u2014 Claude, ChatGPT with browsing, or Gemini with extensions. That’s where ReAct becomes tangible.<\/p>\n
The PMs who will lead in the next five years are the ones who understand not just what AI can do, but how it thinks \u2014 and how to direct that thinking toward project outcomes.<\/p>\n
Alejandro Barahona is a PMP\u00ae-certified Senior Project Manager specializing in software integrations and AI-augmented delivery. He leads integration projects at Capgemini Engineering and is co-founder of Bartronics. Available for remote roles \u2014 connect on WhatsApp or view portfolio.<\/p>\n","protected":false},"excerpt":{"rendered":"
There’s a difference between using AI and using AI well. Most project managers are in the first camp. They open ChatGPT, type a question, get an answer, and\u2026<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[17,15,16],"class_list":["post-14","post","type-post","status-publish","format-standard","hentry","category-ai-project-management","tag-framework","tag-ia","tag-prompt"],"yoast_head":"\r\n