The 30-day first workflow (L2)

The steps

1. 01

   Pick one workflow using the four-part filter

   Tool: Google Sheets

   List every repetitive revenue task your team does weekly. Keep only tasks that pass all four tests: (a) it happens at least 10 times a week, (b) the output is text or a document rather than a judgment call, (c) you can point to what a good version looks like today, (d) a wrong version is embarrassing rather than dangerous. Score each surviving task 1-5 on frequency and 1-5 on pain, multiply, take the top one. Workflows that pass in most companies: writing first-draft follow-up emails after discovery calls, summarizing support tickets into a weekly theme report, turning a call transcript into a CRM note, drafting first-pass responses to RFP or security questionnaire items, rewriting one-paragraph product descriptions for different segments. Workflows that fail and get chosen anyway: forecasting, pricing approvals, anything touching a contract, anything where a wrong answer reaches a customer unreviewed. • Owner: whoever will be accountable for the number at day 30. One name, not a committee. • Tool options: Google Sheets or Excel • Pitfall or what breaks: choosing a workflow that needs judgment; edit rate stays at 70% and everyone concludes AI does not work here. • Definition of done: one workflow is named in a document, along with the three runners-up you explicitly rejected and why.

2. 02

   Time the human baseline before AI touches anything

   Tool: Google Sheets

   Have three people do the task the normal way, five times each. Fifteen data points minimum. Log minutes per instance from start to finished output, including revisions. Then have one person who is not the author score each of the 15 outputs 1-5 against a written definition of good you wrote in step 1. Calculate median minutes and median quality. Do not skip to AI because you "already know it takes about an hour." An estimated baseline gets challenged in the readout and the whole project dies with it. • Owner: the workflow owner • Tool options: a Google Sheet with five columns: date, person, minutes taken, output link, quality score 1-5 • Pitfall or what breaks: skipping the baseline, so at day 30 you have a feeling instead of a number and the project dies in the readout. • Definition of done: the sheet has 15 rows, a median minutes figure, and a median quality score. Screenshot it; that screenshot is half your day-30 readout.

3. 03

   Buy exactly 10 seats of one assistant on a monthly plan

   Tool: ChatGPT Team

   Monthly billing, not annual, no matter the discount offered. Ten seats: the workflow owner, five people who will actually run the workflow, one skeptic, one manager, and two spares. Turn off training-on-your-data in the admin settings on day one; in ChatGPT Team and Claude Team this is off by default on business tiers, but verify rather than assume. Get the DPA from the vendor's trust page and send it to legal the same day so review runs in parallel with your build. If your company already has one of these deployed but unused, skip the purchase and pull seats from the dormant pool instead. • Owner: the workflow owner plus finance • Tool options: ChatGPT Team, Claude Team, Gemini Business or Microsoft 365 Copilot • Pitfall or what breaks: buying a platform, an AI GTM tool, an agent builder, a data-enrichment contract or a consultant in the first two weeks; ten seats of one assistant is the entire budget. • Definition of done: 10 people can log in, data-retention settings are documented in a screenshot, and legal has the DPA in hand.

4. 04

   Write the prompt as a reusable asset, not a chat message

   Tool: ChatGPT Projects

   Create one Project for this workflow. Load four things into its instructions or knowledge: (a) role and goal in one sentence, e.g. "You draft post-discovery follow-up emails for a B2B sales team selling [category] to [buyer]."; (b) rules on what it must never do, such as never invent a customer name, price, date or metric, never promise a timeline, never exceed 150 words, stay at an eighth-grade reading level; (c) context, meaning your one-page product description, your three most common objections with approved answers, and your current pricing structure if the workflow needs it; (d) examples, meaning three real outputs you would be proud to send and two you would not, each labeled with why it fails; the bad examples do more work than the good ones. Then test on five real cases against the same quality rubric from step 2. Iterate the Project instructions, not the chat, so the improvement is permanent and shared. • Owner: the workflow owner plus the best current performer at the task • Tool options: ChatGPT Projects, Claude Projects, or Gemini Gems, the saved-context feature of whichever assistant you bought • Pitfall or what breaks: iterating in chat instead of the Project, so improvements are lost and not shared. • Definition of done: two different people using the same Project on the same input produce outputs a manager cannot tell apart, and quality scores at or above the human baseline median.

5. 05

   Run it live for three weeks with the human approval gate on

   Tool: ChatGPT Projects

   Every instance goes AI-draft, human-edit, send. Nobody sends unreviewed output, all three weeks, no exceptions. Log the same five columns as the baseline plus one more: percentage of the draft the human changed. That last column is your real signal; if people are rewriting 60% of every draft, the prompt is wrong and step 4 is not finished. Hold a 15-minute standup every Friday. One question: what did the model get wrong this week. Fix the Project instructions in the meeting, live, where everyone sees it. • Owner: the five people who run the workflow • Tool options: the Project, plus the same logging sheet from step 2 • Pitfall or what breaks: high edit percentage that does not trend down, meaning the prompt was never actually finished. • Definition of done: 60+ logged instances across three weeks, and the edit percentage has trended down week over week.

6. 06

   Publish the before-and-after and decide with three options only

   Tool: Slack

   Put four numbers side by side: median minutes before and after, median quality before and after. Add hours saved per week across the team and the honest annualized figure. Include the failure list from your Friday standups, unedited; a readout with no failures in it does not get believed and does not get funded. Present exactly three options and pick one: Expand (same assistant, second workflow, run this playbook again; the default answer if quality held and time dropped), Deepen (same workflow, connect it to CRM or the ticketing system so drafts land where work already happens, which is playbook 3 or 5 and moves you to L3), or Stop (cancel the seats; a clean kill after 30 days and $300 is a good outcome, and being willing to make it is why anyone trusts your next readout). Do not pick a fourth option involving a platform demo. • Owner: the workflow owner, presenting to whoever controls the budget • Tool options: one slide or one page, shared in Slack or Teams • Pitfall or what breaks: someone sees an early good result and buys a platform in week two, turning the pilot into an unscoped rollout. • Definition of done: the decision is written down with the number that justified it, and the next playbook is named with an owner and a start date.

7. 07

   Instrument the result

   Tool: Google Sheets

   Track median minutes per instance and median quality score, always together; time saved with quality dropping is not a win, it is a slower problem arriving later. • Where it breaks: three ways in order of frequency: you skip the baseline and end up with a feeling instead of a number; you pick a workflow that needs judgment and edit rate stays at 70%; or somebody sees an early good result and buys a platform in week two before the pilot is scoped. • Visual guidance: two paired bars, before and after, for minutes, with a quality dot plotted on each bar. One chart, both numbers, no way to show the win without showing the quality.

Tools in this playbook

• Google Sheets
• ChatGPT Team
• ChatGPT Projects
• Slack

Next playbooks

• GitHub Copilot for everyday autocomplete (L1)
• Perplexity for daily research (L1)
• Bardeen + Make.com personal workflow agents (L2)
• Cursor + Claude for whole-team coding (L2)