Where to Start or Continue with AI
Five states, two transitions, and one step for this week.
In industry news, everyone already lives in a world of autonomous agents—while on your screen it is the same work as a year ago: ERP or Excel spreadsheets, long email threads, analytical memos, contract approvals, and budget defenses.
The gap between AI hype and real practice happens because people try to skip stages. Someone who hasn’t yet let language models near work documents doesn’t need to build complex agents—they need one fifteen-minute approach to a safe draft. And someone running dozens of prompts a day in ChatGPT or Claude needs to stop using a powerful analytical tool like an expensive typewriter.
This page is the working navigator of the Extended Mind project. Honestly identify your current stage of working with AI based on your actual habits, and pick up a concrete step forward for the coming week. Then see what else awaits you on the path of adopting AI in your professional life.
Where to Aim: Two Capabilities and Two Dead Ends
Adopting AI is not collecting someone else’s prompt templates in your Telegram “Saved Messages” or chasing new model releases. At the top of the route, a professional relies on two capabilities that change the very scale of their work:
-
Thinking together with AI, not letting it think instead of you. Search engines, reference databases, and spreadsheets simply returned information and handed control back to the human. Language models participate in reasoning itself for the first time: they help shape an argument’s structure, spot vulnerabilities in your position, and ask uncomfortable questions. When you integrate a model into your thinking loop, every new generation of systems makes you stronger as an expert.
-
Designing work around AI. Instead of opening a chat window ad hoc, you assemble a managed process: which steps to leave to ordinary spreadsheets, where to bring in a language model, and what criteria to use when accepting the result. Here it is crucial to separate two layers:
-
Technical wiring (“buttons”): how to connect services and configure data routing. Useful today, but by 2028–2029 systems will largely handle setup on their own.
-
Strategic judgment: what exactly to entrust to a machine, how to distinguish polished defects (an answer that sounds confident and fluent yet contains an error) from a usable result, and where the boundary of your personal accountability lies. This layer remains human for a long time.
-
The measurable outcome of this path is not merely saving half an hour on emails, but a new class of tasks you can take on: for instance, running several analytical workstreams in parallel or reviewing eighty documents in a single cycle instead of twenty.
Along the way, it is easy to turn down two dead-end branches that look very much like success from the outside:
-
The Erudite’s Dead End (“literacy instead of practice”). You know everything about the latest versions of ChatGPT, Claude, and DeepSeek and read all the reviews, yet in real work you type one-line queries into chat as if it were a search bar, get a generic reply, and close the tab.
-
The Passenger’s Dead End (“handing over work with your brain turned off”). You use AI daily, but accept its outputs by default. When asked “why did we choose this approach?”, the answer is “that’s what the AI generated.” As a result, hallucinations slip into documents under your signature while your own expertise quietly erodes.

Self-Check: 5 States of Practice
Think back to your most time-consuming work tasks over the past two weeks. Pick the description where you recognize what you actually did with your own hands—not articles you read or home experiments with recipes or vacation planning. If you see yourself in two adjacent states, start with the earlier one.
1. AI Outside Work
-
How to recognize yourself in tools: over the past two weeks, AI has not been involved in a single one of your work tasks. You may have seen an “Assistant” button in your workspace suite, heard colleagues talk about ChatGPT, or even installed a chatbot app on your phone, but you prepared every work email in Outlook, policy document in Word, and summary table in Excel entirely the traditional way.
-
What usually holds you back: healthy professional skepticism (“my work demands precision and regulatory compliance, while neural networks confidently make things up”) or a sense of corporate prohibition (“we’re not allowed to use that here”).
-
Step for this week: separate external restrictions from internal inertia—ask yourself: “If an official permission slip landed on my desk tomorrow morning, which task would I open first?” Set aside 15 minutes for one small task free of trade secrets or personal data: an outline for an upcoming report, a plan for a difficult email, or an answer to a recurring question from a colleague. Open any available chat (ChatGPT, Claude, Gemini, DeepSeek, or an approved corporate deployment where external services are restricted) and describe the situation: who the document is for, why it is needed, and what it must cover. Evaluate the output the way you would review a new intern’s draft: what to keep, what to discard, and what to ask to be redone.
-
How you know the step worked: you opened the tool on a real work task yourself and can point specifically to where it helped and where it produced nonsense.
2. First-Answer Lottery
-
How to recognize yourself in tools: you open ChatGPT, Claude, or DeepSeek a couple of times a month when you hit writer’s block. You type a short phrase without context—exactly as if it were a Google search bar. Out comes a smooth but watered-down template. You conclude that “it can’t handle our industry specifics,” close the tab, and finish the job manually in Word.
-
What usually holds you back: the habit of treating a language model like a search engine and expecting a finished result from a single prompt.
-
Step for this week: introduce a follow-up rule. Never judge AI’s first answer—it merely highlights the constraints you forgot to state out loud. All week long, make at least 2–3 clarifying replies in the same chat before closing the window. For example, when drafting a difficult email to a client about a schedule slip, follow up after the first stiff reply with concrete details: “The two-week delay happened because the client approved the specification late; we cannot blame them outright, and in return we are ready to ship the first batch on the original date.” Save phrasing that works well in a dedicated note (Notion, Obsidian, Apple Notes, or Telegram Saved Messages).
-
How you know the step worked: after your follow-ups, the draft actually went to use, and opening an AI window at the start of a new task no longer requires a separate act of willpower—you have crossed the first key transition and built a habit.
3. Tamed Routine (The Main Plateau — 80–85% Get Stuck Here)
-
How to recognize yourself in tools: you are an active user. You look up background facts in Perplexity instead of traditional search, condense call recordings from Zoom or corporate video platforms into brief summaries, translate contracts in DeepL or ChatGPT, and polish email tone in Claude. You might rely on two or three favorite techniques or maintain a collection of thirty saved prompts and custom GPTs. Yet all this activity revolves around text packaging and minor routine, while complex substantive work—defending a budget, diagnosing a performance drop, preparing a negotiation position—still happens entirely the old way.
-
What usually holds you back: the comfort trap. In editing and summarizing, results are fast and predictable, whereas letting a machine into the substantive core of your work feels either risky or pointless.
-
Step for this week: once a week, pick one “stretch task” beyond text packaging. Keep the structure and final decision for yourself, but assign AI the narrow role of a critic. Upload an anonymized draft of a business case or decision memo into Claude, ChatGPT (with Thinking / o-series reasoning enabled), DeepSeek R1, or NotebookLM with this prompt: “Here is my proposal using placeholder numbers. Read it through the eyes of a skeptical CFO: in which three places is the logic vulnerable, and what data is missing to defend the conclusion?”
-
How you know the step worked: the model’s critique helped you spot a blind spot ahead of time and strengthen at least one argument in the document before the actual meeting.
4. Joint Thinking
-
How to recognize yourself in tools: you know how to set up workspaces in Claude Projects or ChatGPT Projects, upload batches of regulatory documents into NotebookLM, use reasoning models to hunt for flaws, and understand why a system missed the mark (for instance, missing context in the instructions). Yet the stall here is almost invisible: you still come up with the solution in your own head first, and only then call on AI to check, format, or expand a thought you have already settled on.
-
What usually holds you back: protecting your own expertise (“thinking is my job”) and the fear of falling for a persuasive hallucination.
-
Step for this week: practice “thinking out loud with AI”—bring a reasoning model into the loop 10 minutes earlier than usual, while the problem is still raw. For example, after exporting an anonymized summary of a quarterly segment drop from your ERP, CRM, or BI system into Excel, do not look for the cause on your own first. Feed the summary into chat along with your initial working hypothesis and ask two questions: “What else could explain this pattern?” and “What specific numbers in our reports would disprove each of these hypotheses, including mine?” Then open your source spreadsheets in Excel and test the hypotheses against hard facts.
-
How you know the step worked: you have crossed the second qualitative transition—at least one hypothesis from the dialogue prompted you to inspect a slice of data you wouldn’t normally open, and your final conclusion rests on a solid check against the numbers.

5. Process Architecture
-
How to recognize yourself in tools: you regularly work through complex problems in sparring sessions with models, yet the entire workflow lives exclusively inside your head. Every week or quarter you manually export spreadsheets, copy context from old chats, and reconstruct your verification rules from memory. It is tempting to immediately build a bot or wire up an automation chain in n8n, Make, Albato, or through corporate IT. But until the process is extracted from your head onto paper, any automation merely scales chaos.
-
Step for this week: write a “process contract”—document one recurring task on a single page in an ordinary document (Word, Google Docs, Notion) clearly enough that a colleague could execute it without verbal coaching from you. Include five mandatory blocks:
-
Inputs: where the data comes from and in what format (e.g., a plan-vs-actual export from ERP into Excel).
-
Steps and tools in order: what an ordinary formula calculates in Excel / Google Sheets (you don’t need a language model to do arithmetic across sheets of numbers), at which step Claude / GPT drafts a list of hypotheses, and at which step it assembles a draft memo after your verification.
-
Acceptance criteria (“what good looks like”): every number in the text reconciles with the summary table, and every stated cause includes the result of its data verification.
-
“Depends on context” fork: where mechanical rules end and your judgment is required (e.g., choosing the comparison baseline—last quarter or last year).
-
What always stays yours: judging whether the evidence is sufficient, the final conclusion, and your personal signature.
-

- How you know the step worked: looking at this one-pager, a colleague clearly sees the boundary between a mechanical rule and expert judgment, and you have a solid foundation for moving individual steps into semi-automated mode.
Three Rules for Any Step
-
Data and security. To get familiar with a technique, placeholder numbers and fictitious names are plenty. On a real task, work either with an anonymized document structure stripped of trade secrets and personal data (replacing “Acme Corp, revenue 420M” with “Client A, index 100”) or inside an approved corporate environment (self-hosted models, an enterprise cloud workspace, or a secure gateway). When in doubt, do not send.
-
Honest time accounting. Count not the seconds until text appears in chat, but the entire path to a finished document: preparing the export, running the dialogue, fact-checking, and manual editing. At the same time, avoid the perfectionist trap: compare AI’s draft not against an imaginary ideal, but against what you could have written yourself in the same 15 minutes.
-
Three verification buckets and the right to say “no.” Double-checking every single line after AI everywhere wipes out your entire gain; checking nothing erodes your expertise. Sort tasks by the cost of an error into three buckets: always verify, spot-check, and do not verify. And remember: concluding “AI isn’t helpful for this task” is an excellent outcome if you gave the model context, ran 2–3 follow-ups, and understand why it fell short.

How I Walk This Path with You in Extended Mind
You can follow this route on your own—the page you are reading was built specifically as an open map. And the Extended Mind project (this site and my Telegram channel) is designed as a working laboratory where, free of marketing noise, I break down every transition on real managerial and analytical tasks.
Across all articles and case breakdowns, I examine practice through four cross-cutting dimensions:
-
Thinking and method — how to frame problems, conduct analytical sparring with a model, and separate mechanical steps from human judgment (see When AI Eats Judgment and How to Turn Raw Thoughts into a Proposal for Your Manager).
-
Tools beyond button-pushing — when a fast chat is enough, when to switch on reasoning models (Reasoning), when to assemble a project knowledge base in Claude Projects / NotebookLM, and when to close AI altogether and solve the problem with an ordinary formula in Excel (model landscape breakdown: What Kinds of Neural Networks Exist and Access Options; how models work under the hood: the “AI: Foundation” series).
-
Reliability and information security — how to anonymize work materials in two minutes, how personal verification checklists work, and how to protect workflows from hallucinations and vulnerabilities (see 42 Questions About AI and Information Security and When AI Is Banned in Your Organization).
-
Team architecture — how to turn personal discoveries into clear technique cards for colleagues and transition proven steps into managed semi-automation (illustrated with meetings: The “Phantom Action Items” Trap and Meeting Summary in Five Minutes).
Navigating Stages Inside the Project
You can join the breakdowns at the exact stage where you are today (and comfortably stop at the level that is sufficient for your work tasks):
1. AI Outside Work: Safe Start
Breakdown focus: Data hygiene rules in corporate environments, a fast method for anonymizing documents, and walkthroughs of your first 15-minute approaches on recurring tasks.
Outcome: Your first working “before / after” example and a realistic map of 2–3 safe tasks where AI genuinely saves effort.
2. First-Answer Lottery: Working Method
Breakdown focus: Iteration technique (the 3-step rule), setting quality criteria before you start, and comparing models on identical work inputs.
Outcome: A steady habit of opening AI by default, a personal bank of working prompts, and quick verification rules.
3. Tamed Routine: Systematic Practice
Breakdown focus: Auditing your workweek, shifting AI from “text editor” to critic, sparring partner, and analyst on complex domain tasks.
Outcome: A set of deep analytical workflows, technique cards for sharing with colleagues, and a broader range of tasks.
4. Joint Thinking: Personal Work System
Breakdown focus: Assembling context bases, working with reasoning models before making a decision, writing a “process contract,” and preparing 5–7 verification test cases.
Outcome: A repeatable process that a colleague can execute from your one-page description without verbal hints.
5. Process Architecture: Managed Automation
Breakdown focus: Dividing steps between spreadsheets and models, designing semi-automated workflows (Make, n8n, scripts), setting up circuit breakers, and writing clear specifications for IT.
Outcome: A designed pipeline under your control, protected against silent defects, and a well-structured requirements package for developers.
Where to Start Today
-
At your own desk: return to whichever of the five levels above matched your past two weeks, pick one task for tomorrow, and block 15–20 minutes on your calendar for this week’s step.
-
Together with me: read the in-depth breakdown of transition mechanics and automation in my flagship article “Why 85% of Professionals Get Stuck Halfway with AI” or open the next installment of the “AI: Foundation” series — “What Lies on the Other Side of the Chat Window” (full series syllabus), and join the discussion of real-world cases in the Extended Mind Telegram channel.


