AI transformation starts with a useful change in how work gets done. Buying an assistant is a tool decision. Transformation means changing the way a team gathers context, makes an artifact, checks it, handles exceptions and learns from the result. This framework turns that idea into a sequence a beginner can follow.
Use one real workflow as your classroom. Marketing and business intelligence are good places to practice because briefs, reports and decisions create visible artifacts. Start with work you understand well enough to judge. Your first goal is a better way to do that work, with evidence that it helps. The framework below is Carroll Consulting’s practical teaching synthesis, not a certification or a promise of a particular return.
What you will leave with
A documented workflow, a bounded pilot, evidence of what improved, and a clear owner for the next decision.
Step 1
Understand AI without the mystique
A model generates a response from its context. A chat is an interface for working with it. An agent adds tools and a loop that can take actions toward a goal. An automation runs a defined process when a condition or schedule triggers it. None of those labels guarantees that the result is correct.
The first mindset shift is from asking for an answer to specifying an outcome. Instead of ‘analyze this campaign,’ name the decision, data, output and checks. The second is from judging fluency to inspecting evidence. A polished explanation of a wrong total is still wrong. Teach people to ask what supports a claim and what could invalidate it. Give beginners sample inputs before connecting company systems.
Try this instruction
Explain the difference between a model, a chat, an agent and an automation using our weekly reporting workflow. For each, show what it can help with, what context it needs and what a human still checks.
Each participant can explain the task, the agent’s permitted actions and how they will judge the result.
Step 2
Map the work before choosing the tool
Pick a repeated task with a named owner, known inputs and an output another person actually uses. Record the current steps, systems, waiting points, corrections and decisions. Include the quiet work: finding the latest file, reconciling totals, getting approval and explaining missing data. Ask the people doing the task where it breaks.
For a marketing report, the current workflow might be export, clean, join, calculate, write, review and circulate. Measure more than generation time. Capture hands-on effort, elapsed time, corrections and whether the reader can make the intended decision. Use representative recent examples and record how you measured them. If a step adds no value, remove or simplify it before automating it.
Try this instruction
Help me map this workflow. Separate steps, inputs, handoffs, decisions, waiting and rework. Ask for missing facts. Identify what could be removed before suggesting AI. Do not invent our baseline.
The current workflow, its owner, its inputs, its output and a credible baseline are written down.
Step 3
Redesign roles and handoffs
Choose where AI should assist a person, where an agent can own a bounded subtask, and where ordinary deterministic software is the better fit. Calculations, access rules and repeated validations often belong in code or formulas. AI can help interpret a brief, draft a narrative or propose hypotheses after those checks run.
Write a future workflow with a single owner for each output. In a BI example, a data check runs first, an analyst calculates results, a writer drafts a decision memo, and a reviewer checks every material claim against the computed evidence. The reviewer cannot begin before a draft exists. Decide what happens when data is missing, tools fail or agents disagree. Define the actions that are drafts and the actions that may change external systems.
Try this instruction
Redesign the mapped workflow. For each step list human responsibility, AI responsibility, required input, output, validation and failure handling. Preserve a human decision where judgment or authority is required. Show what changes compared with today.
The proposed workflow names each owner, dependency, check and exception path.
Step 4
Pilot on bounded, representative work
Use approved sample or appropriately authorized data and a small number of representative cases. Include an ordinary case, a messy case and a case where the correct result is to stop and ask for information. Keep the existing workflow available during the trial. Give the pilot an owner and an explicit definition of done.
Compare equivalent inputs and outputs rather than a carefully polished AI demo with a difficult historical job. Keep an exception log: incorrect claims, missing context, tool failures, manual corrections and extra coordination effort. Have the person who uses the final artifact evaluate it. A report that is fast to produce but cannot support a decision has not met its goal.
Try this instruction
Design a bounded pilot for this workflow. Use cases that include normal work, messy inputs and a stop condition. Specify what we will record and what evidence would justify continuing. Do not invent results or deadlines.
The pilot has actual artifacts, documented exceptions and a review by the intended user.
Step 5
Measure value and make a decision
Compare the baseline with the pilot on task completion, error or correction rate, hands-on effort, elapsed time, tool cost and reviewer effort. Keep the unit of comparison consistent: per report, per campaign brief or per completed request. Include failed attempts and the work needed to get a result ready for use.
Distinguish observed changes from explanations. If a pilot finished faster, that does not by itself prove that the process will work across every team or dataset. Ask whether the gain is repeatable and whether quality stayed acceptable. Choose to adapt, adopt or stop. Stopping a workflow that does not help is a useful outcome. Record the reason so the next trial starts from what you learned.
Try this instruction
Compare the baseline and pilot evidence. Separate measured results, assumptions and unanswered questions. Include failed attempts and review effort. Recommend adapt, adopt or stop, and explain what evidence could change that recommendation.
The decision uses observed results and includes quality, review effort, costs and limitations.
Step 6
Build the operating habit
Turn a successful pilot into a short playbook: when to use it, inputs, approved tools, owners, steps, checks, exception handling and where the output belongs. Store the current instructions in a shared source of truth. Teach another person to run it and ask them to improve the instructions where they get stuck.
Assign responsibility for tool access, data definitions, maintenance and final approval. Revisit the workflow when the task, model, source system or risk changes. Use a review cadence that fits the work rather than treating the original prompt as permanent. Introduce agents or automation only when the handoffs and failure cases are understood. The lasting mindset change is that AI work is designed, checked and improved like any other operating process.
Try this instruction
Turn the accepted workflow into a reusable playbook. Include entry conditions, required inputs, owners, steps, checks, permissions, exception handling, output location and maintenance responsibility. Keep facts separate from assumptions.
Another person can run the playbook, find the output and know how to handle an exception.
When you get stuck
Everyone has an AI subscription, but work still feels the same.
Select a shared workflow and a useful outcome. Map its handoffs and measure the whole process, including review and rework.
The pilot looks impressive but nobody uses the output.
Bring in the actual recipient. Define the decision the output must support and test against that need.
Adding more agents makes the work slower.
Return to one lead. Parallelize independent tasks only, give outputs exclusive owners, and include coordination time in the comparison.
The team keeps correcting the same mistake.
Fix missing context or add a deterministic check where appropriate. Update the shared playbook and test the failure case again.
Terms explained
- Baseline
- A documented observation of how the current workflow performs.
- Pilot
- A bounded trial that produces evidence for a decision.
- Human in the loop
- A specific person with a defined review or decision responsibility, not a vague promise of oversight.
- Source of truth
- The agreed place to find the current version of a fact, instruction or decision.
Sources and further reading
The workflow and templates are teaching guidance. The sources below support the underlying tools and concepts; product setup and availability can change.
Take it into your next task
Your completion checklist
0 / 6 checked
Checks are saved on this browser. Mark them only after you inspect the evidence. These are your checks, not an automated certification.