AI Lab

Change how your team works with AI.

A practical transformation framework for marketing and business intelligence. Start from zero. Redesign one workflow, prove it helps, and build from there.

Start with the framework Find a useful workflow
On this pageStart where the work is.

A framework, practical workplace playbooks and tools you can use today.

A different way to work

From first experiment to a working habit.

AI becomes useful when the goal, context, responsibilities and checks are clear. Work through one real task before expanding.

  1. Understand the tools

    Know what to delegate and what to verify.

    You leave with: a shared vocabulary and clear boundaries

  2. Map the work

    Find a repeated task with a real owner and a visible problem.

    You leave with: a current workflow and a baseline

  3. Redesign the workflow

    Change the sequence, responsibilities and handoffs.

    You leave with: a proposed workflow with review built in

  4. Run a small pilot

    Try the new approach on representative work.

    You leave with: real artifacts and an exception log

  5. Measure what changed

    Compare quality, effort, time and cost.

    You leave with: a decision to adapt, adopt or stop

  6. Make it repeatable

    Give the workflow an owner, instructions and a review habit.

    You leave with: a working playbook your team can reuse

The mindset shift

From asking for an answer to defining a useful outcome. From trusting confident prose to checking evidence. From giving everyone a tool to redesigning the work together.

How this connects to my method

Practical workplace use cases

Start with work you already need to do.

Each playbook defines the inputs, artifact, agent’s role and human checks. Use approved data and leave with something your team can use.

Marketing

Business intelligence

Start with the work

What do you need to get done?

Turn a goal into instructions you can use with your AI assistant.

Plan a marketing campaign

Turn a real offer and audience into a brief your team can execute.

Create a brief without connecting an account.

Design the work

Give the workflow a plan.

Map the process you want to improve, choose only the infrastructure you need, and give every agent a clear job.

Redesign one workflow. Keep the human purpose.

Map the work you do today, decide where AI can help and create a small pilot you can evaluate. Begin with an illustrative example or your own task.

Examples are fictional planning inputs, not measured results. Switching the starting point replaces the fields below.

Understand the work today

One step per line, including review and handoffs.

Describe sources and definitions. Do not paste passwords, keys or private source records.

Define a bounded pilot

Choose observable quality and effort measures, not an invented maturity score.

Leave blank if unmeasured; the charter will make that gap explicit.

Processed in this page. These fields are not sent to an AI service or saved by this canvas. Copy or download your charter before leaving.

Prepare the weekly campaign review

A sourced review that helps the marketing lead choose one action to test.

Accountable human: Marketing lead

Current work
  1. Export campaign results
  2. Check dates and metric definitions
  3. Reconcile totals
  4. Write the performance summary
  5. Review claims and choose an action

Friction: Repeated copy-and-paste and unsupported explanations slow the review.

Proposed pilot

AI helps with a bounded draft or analysis. You choose the inputs, check the output and take the next action.

  1. Marketing lead confirms the useful outcome and the acceptance checks.
  2. Prepare only the authorized inputs, record their dates and definitions, and keep a copy of the current workflow for comparison.
  3. Give AI one bounded draft or analysis step. Keep the human in control of each next step.
  4. Check the artifact against the actual sources and the stated measure. Return corrections to the output's owner.
  5. The accountable human makes the review decision before any separately authorized external action.
  6. Compare the pilot with the baseline. Keep the useful steps, change what failed and save the corrected brief for another attempt.
Success measure
Time to a usable brief, factual corrections needed and whether the chosen action has supporting evidence.
Baseline
Not measured yet. Time one current review and count corrections before the pilot.
Permitted actions
Read the provided exports and write a draft in the pilot folder. Do not change campaigns, budgets or publish messages.
Human review
Marketing lead verifies claims, approves the final brief and decides any campaign changes.
Checks before repeating or scaling
  • Confirm the outcome helps the person doing the work, not just that AI produced an answer.
  • Record the baseline and pilot under comparable conditions. Include review time, corrections and relevant tool costs.
  • Trace factual claims and numbers to the inputs. Reconcile counts, totals and definitions when data is involved.
  • Try a missing or conflicting input. Check that the workflow reports the gap instead of inventing an answer.
  • Inspect what changed, who reviewed it and which actions actually occurred.
  • Choose to keep, revise or stop based on the evidence. Repeat successfully before scaling access or automation.

Read or select the export text

Choose your smallest useful setup.

Start with the job. Get a tool recommendation, a sequence to follow and checks you can actually perform.

What does the first version need?

Git + GitHub + Vercel

Get a useful page online before adding accounts or a database.

Git + GitHub
Git records inspectable local checkpoints. GitHub holds the pushed copy and a place to review changes.
Vercel
Build and publish the page or app. Inspect the local version and verify the deployment target; a new project’s first deployment may be production.
Do this in order
  1. Write one small outcome and the observable checks that define done.
  2. Build locally with sample content. Save a working Git checkpoint, inspect the diff, then push it to GitHub.
  3. Verify the intended deployment environment and who can access it before publishing; plan phone and keyboard checks for the actual URL.
  4. Set the required environment variables in the host. Keep secret keys out of browser code and the repository.
  5. Publish deliberately, test the real URL or job result, and record how to recover.
Check the result before calling it done
  • A fresh checkout contains the files needed to build the project, without secrets.
  • The main action works at the deployed URL and at a narrow screen width.
  • You have checked current plan conditions and know how to stop the resources you enabled.
Why these tools, and what to verify
  • Vercel and Railway overlap: both can host web applications. This recommends one starting host, not a mandatory frontend/backend pair.
  • No database or login is required for this starting point. A background task can use an existing authorized source and export a result without adding Supabase.
  • This is a suggested minimum, not a price estimate. Verify current limits, billing, backup coverage and runtime requirements before connecting accounts.

Read or select the export text

Give every agent a clear job.

Turn a real task into named outputs, permissions and handoffs. These are editable instructions for your chosen tools; no agents run here.

Start with one agent. Add helpers for independent work, not because every task needs a team.

Your choices are processed in this page. The planners do not connect accounts or send a brief to an AI service.

Marketing campaign plan

A usable campaign brief with audience, offer, channels, draft messages, measurement plan and clearly labeled assumptions.

Bring these inputs
  • Business goal, audience and offer
  • Approved brand facts and example messages
  • Budget range and dates, only if supplied
  • Existing channel performance with source, date range and attribution definition
One agent, a clear sequence
  1. One agent, sequential passes

    Confirm the brief, then research, prepare, draft and check in order. Review the supplied performance data and constraints. Separate measured results from proposed targets; identify missing budget or tracking inputs.

    Inputs
    Business goal, audience and offer; Approved brand facts and example messages; Budget range and dates, only if supplied; Existing channel performance with source, date range and attribution definition
    Output
    campaign-plan.md, with sources, check results and unknowns
    Starts after
    Finish each pass before beginning dependent work; check the draft only after it exists.
    May do
    Read authorized inputs and create the draft. Do not send, publish or change connected systems.
How to check this task
  • Each factual or performance claim has a source; no invented testimonials, audience sizes or results.
  • Targets and forecasts are labeled as proposals, with assumptions and no outcome guarantee.
  • Each channel has an owner, an action, a measurement and a stated dependency.
  • Budget totals reconcile to the supplied allowance; unknown costs or dates stay unknown.
Rules that prevent collisions
  • One writer owns each output. Reviewers return findings rather than silently editing someone else's file.
  • A shared folder does not share chat history. Pass the brief, exact artifact paths, source notes and current status explicitly.
  • Keep secrets and private source records out of public prompts and exports. Use only the folders, data and tools authorized for this task.
  • Drafting does not authorize sending, publishing, changing accounts or spending money. Name any external actions separately in the brief.
  • When agents share a desktop, only one controls a particular window at a time. Independent background research can continue.
  • A self-check is useful but is not an independent review. Ask for a separate reviewer when the stakes justify it.

Use native delegation only when your chosen product supports it. Separate desktop AI apps do not automatically talk to each other: copy or explicitly share the brief and artifact paths, confirm receipt, then continue. This planner generates instructions; it does not run or connect agents.

Read or select the export text

Learn by checking real calculations

Calculate first. Explain second.

Use the sample or a compatible CSV to inspect marketing metrics. Then give an AI analyst a brief grounded in the calculated result.

Check the numbers before writing the story.

Try the synthetic sample or load a campaign CSV. This workbench recomputes campaign totals and rates so you can inspect the evidence before asking AI to interpret it.

Your CSV stays in browser memory. This tool sends no file contents to a server, model, analytics event, or browser storage. Reloading clears the uploaded data and returns to the sample.

Up to 200 KB and 2,000 data rows. Required column names:

campaign,spend,impressions,clicks,conversions,revenue

Use plain nonnegative numbers, one currency, and one reporting period. Extra named columns are ignored. Blank rows are skipped; malformed rows reject the whole file.

Synthetic practice data: Synthetic campaign sample. 4 input rows, 3 campaign groups.

CTR
1.64%
CPC
0.65
CVR
7.53%
CPA
8.62
ROAS
3.48×

Rates use the summed inputs, not the average of row percentages. A zero denominator is shown as unavailable.

Synthetic sample calculations. Repeated campaign labels are combined. Amounts use the input’s currency units.
CampaignSpendImpressionsClicksConversionsRevenueCTRCPCCVRCPAROAS
Search brand200.005,000.00500.0040.001,200.0010.00%0.408.00%5.006.00×
Community newsletter60.0012,000.00120.0012.00360.001.00%0.5010.00%5.006.00×
Social prospecting240.0030,000.00150.006.00180.000.50%1.604.00%40.000.75×
All campaigns500.0047,000.00770.0058.001,740.001.64%0.657.53%8.623.48×

Display values are rounded to two decimals. Calculations use unrounded inputs. No performance recommendation is inferred from these numbers.

Metric definitions and checks before interpretation
CTR: Clicks ÷ impressions
Click-through rate, shown as a percentage.
CPC: Spend ÷ clicks
Cost per click, in the same currency units as spend.
CVR: Conversions ÷ clicks
Click-based conversion rate. Other systems may use a different denominator.
CPA: Spend ÷ conversions
Cost per recorded conversion, in the same currency units as spend.
ROAS: Revenue ÷ spend
Return on ad spend, shown as a multiple. This is not a profit calculation.
  • Use one reporting period, currency, and attribution basis. These columns do not verify that those match.
  • Repeated campaign labels are combined exactly and case-sensitively. Duplicate records are not removed automatically.
  • Aggregate rates are recomputed from summed inputs, never averaged from row rates. A zero denominator is unavailable.
  • CVR uses clicks as its denominator. ROAS compares revenue with spend and is not profit. Attribution does not establish causation.
  • Only numeric format and range are checked. Verify source completeness, metric definitions, missing records, and duplicate exports before drawing conclusions.

Copy these instructions into your assistant, then decide which approved report or sources to provide. Copying this brief does not include your CSV or its campaign values.

The reference library

Learn the tools when you need them.

Beginner explanations, complete workflows, recovery paths and source links. Every guide has a saved checklist, printable view and downloadable copy.

Build your first useful page with AI

Go from a clear idea to a working page, with saved checkpoints and a way to recover when a change breaks.

Put AI to work in marketing and business intelligence

Turn approved campaign material and performance data into a useful brief, with clear roles and evidence you can check.

Choose the tools your project actually needs

Understand hosting, services, data, and access without signing up for an entire stack on day one.

Save your work and make changes with Git

Learn checkpoints, branches, diffs, and review so an AI-assisted project remains understandable and recoverable.

How to coordinate AI agents without losing the thread

Give desktop agents clear roles, useful handoffs and separate ownership, then check the combined result.

Write an AI project brief that leads to useful work

Turn a vague request into clear scope, useful examples, allowed actions and observable checks.

Deploy your first website and verify the real result

Publish a small AI-built site with GitHub and Vercel, understand deployment environments, and practice a controlled update.

A practical Supabase security checklist for beginners

Check keys, row-level access, storage and recovery before trusting an AI-built app with private data.

Debug AI-generated code one observable failure at a time

Reproduce the problem, collect useful evidence, make a focused repair and prove the original failure is fixed.

The AI project launch checklist: from demo to dependable

Check real behavior, access, data recovery, costs and operations before sharing an AI-built project.

Use it with your team

Show the work. Let someone else try.

Pick a playbook, demonstrate one task, then let a colleague run a second example. Compare the output, the effort and the corrections. Save the version that works and give it an owner.

This lab brings together Joseph Carroll’s focus on useful outcomes, evidence and accountable execution. Read the framework, use the templates, and adapt the process to your team.

About Joseph Carroll

A practical teaching session

  1. Bring a real task and define a useful result.
  2. Map the current work and agree on the checks.
  3. Try a bounded workflow with approved inputs.
  4. Review the artifact and document what changed.
  5. Repeat with a new example before expanding.
Take the framework with you