YOUR DEPLOYMENT
PLAYBOOK

Here's the thing about AI: almost everyone knows they should be doing something with it. Almost nobody has started. And it's not because they're lazy, or behind, or not smart enough. It's because of a very specific set of feelings that get in the way.

8 REASONS PEOPLE STALL — SEE IF ANY SOUND FAMILIAR

• Too many options, none chosen: There are thousands of AI tools. When everything is an option, nothing gets picked. You open a browser tab, read for 20 minutes, close it, and do nothing.
• Not knowing where to start: The hardest part isn't learning the tool—it's figuring out what to point it at first. Most people stall here, not because they can't execute, but because they never decided what to try.
• It feels like it's for tech people, not you: Words like "API," "model," and "workflow" make it sound like you need a computer science degree. You don't. But the language makes it easy to assume you do.
• Fear of breaking something: What if it sends an email you didn't mean to send? What if it deletes something? This fear is reasonable—and there are easy ways to address it. But most people never get that far.
• Waiting for the right moment: "I'll try it when things slow down." Things don't slow down. The wait becomes indefinite.
• Not being able to picture the payoff: Hiring someone is easy to justify—you know what they'll do and what it costs. AI feels murkier. You can't quite picture what changes if it works.
• Afraid of looking foolish: The stories about AI making mistakes—hallucinating facts, misreading instructions—make people nervous about trusting it with anything real.
• Nobody around them is doing it: If none of your friends, colleagues, or competitors are using AI agents, it still feels experimental. Like you'd be the guinea pig. Most people wait for someone else to go first.

The fastest way to unlock imagination: see it in your context. Pick your industry.

THE 10-10-10 RULE — IDENTIFY AUTOMATION CANDIDATES

• Does this task take more than 10 minutes each time?
• Does it happen more than 10 times per month?
• Could it be described in a 10-step (or fewer) process?

Yes to all three = strong automation candidate.

TOP 5 QUICK-WIN FIRST TASKS BY ROI SPEED

1. Email triage and drafting — 1–2 hours/day saved immediately
2. Meeting notes and action item extraction — 30–60 min per meeting
3. Report generation from existing data — 2–5 hours per report cycle
4. Data entry and categorization — 5–10 hours/week
5. Research and competitive intelligence — 3–8 hours per cycle

REAL EARLY-ADOPTER EXAMPLES

• Klarna: AI handled 2/3 of all customer service chats in its first month—equivalent to 700 full-time agents. Started with simple FAQ resolution.
• BCG consultants using GPT-4: 40% better performance, 25% faster on research, ideation, and writing. First tasks were research synthesis and first drafts.
• Small law firms: Case law research saved junior associates 4–6 hours per case.
• Shopify merchants: Product description writing was the gateway—expanded to chatbots and inventory forecasting.
• CPA firms: 80%+ accuracy on invoice processing from day one.
• Marketing agencies: Competitive analysis automation was the most common first task—weekly briefings from competitor social, ad libraries, and websites.

• Hijacked instructions: A malicious website can embed hidden text designed to trick the agent into doing something you didn't ask for—like logging into a site on your behalf, or stealing your data. This has already been demonstrated in the real world.
• Your passwords on screen: The agent sees everything visible on your screen. If your password manager auto-fills, your bank tab is open, or your email is up—the agent captures all of it. Some setups save screenshots for debugging. That's your personal information sitting in a log file.
• Mistakes you can't undo: Agents can click wrong buttons when something unexpected pops up on screen. A sent email, a deleted file, a confirmed purchase—one misread screen can cause real damage.
• Your screen is being recorded: Screenshots are sometimes sent to third-party servers for processing. Anything visible on your screen—medical info, financial records, private messages—could end up in those logs.
• Being sent somewhere dangerous: An agent browsing the web could be redirected to malicious sites, or tricked into uploading your files somewhere they shouldn't go.

A safe zone means that if something goes wrong, it stays contained. Your photos, documents, passwords, and accounts are never touched. Here's how to get there—pick your path.

Give the agent only access to what it needs for the specific task. Nothing more.

1. Don't run it as an administrator on your computer.
2. Create a separate browser profile for the agent—no saved passwords, no accounts logged in except the ones it needs.
3. Create a specific folder called AI_Agent_Workspace. Only put files there that you're comfortable with it touching.
4. Set files to read-only wherever possible—so it can look but not change.
5. Only run the agent when you're actually using it. Don't leave it running in the background.
6. Give it specific, narrow instructions rather than broad open-ended ones.

Some things should always need your sign-off before the agent does them:

• Sending any email or message
• Making any purchase or payment
• Deleting any file or data
• Logging into any account
• Installing any software
• Uploading anything to the internet

Start by approving everything. As it proves itself, you can gradually let it handle the safe, reversible stuff on its own.

NEVER LET IT SEE THESE

• Social Security Numbers or government IDs
• Credit card numbers or banking credentials
• Medical records or health information
• Legal documents in progress
• Encryption keys or 2FA backup codes
• Tax documents

PRE-SESSION HYGIENE ROUTINE

• Close all sensitive browser tabs (banking, email, medical portals)
• Close and lock your password manager
• Clear clipboard
• Log out of accounts the agent doesn't need
• Close communication apps (Slack, Teams)
• Disable desktop notifications from sensitive apps

• Complete beginners, just curious → Rung 1–2
• Business owners wanting structured automation → Rung 2–3
• Non-technical but wanting autonomous agents → Rung 4
• Tech-curious and willing to learn → Rung 5
• Developers, power users, or anyone handling sensitive data → Rung 6

A system prompt is the instruction manual you give an AI before it starts working. Think of it as sitting down with a new employee on Day 1 and saying: "Here's who you are, here's what you do, here's how you do it, here's what you never do, and here's what good work looks like."

THE ROLE FRAMEWORK

• R — Role: "You are a [role] at [company] who specializes in [specialty]."
  Example: "You are a customer service agent at Greenfield Solar who specializes in residential installation inquiries."
• O — Objective: "Your job is to [primary task]. Success looks like [measurable outcome]."
• L — Limits: "You should never [boundaries]. If you encounter [edge case], do [specific action]."
• E — Examples: "Here's an example of a great response: [paste]. Here's a bad response: [paste]."

WRITING TIPS

• Write it like you're training a smart but literal-minded new hire.
• Be specific about what NOT to do—constraints matter more than capabilities.
• Include real examples. They're worth more than paragraphs of description.
• Start simple, add rules as gaps appear.
• Test it by asking: "If I handed this to a stranger, could they do the job correctly?"

• Adjust your instructions: Run the agent, see what it does, rewrite the instructions to fix what went wrong, run it again. Repeat until it consistently gets it right.
• Refine the steps: Sometimes the problem isn't the instructions—it's the order of things. Adjust the sequence until the output is clean.
• Build up examples: The more "here's what good looks like" examples you give it, the better it performs. Save the ones that work.
• Handle the weird edge cases: When it fails on something unusual, add a specific rule for that situation. Over time your instructions get smarter.
• Tighten the guardrails: If it keeps going too far or not far enough, add or loosen specific rules based on what you're seeing.

TIME TO VALUE

• THIS WEEK1–5 days: Moving data between systems, renaming and sorting files, sorting expenses into categories
• THIS MONTH1–4 weeks: Building reports, drafting emails, scheduling social media posts, researching contacts
• THIS QUARTER1–3 months: Multi-step tasks that touch several tools, generating proposals, tracking competitors, organizing large document libraries

Most people who start with a simple task see real time savings within two weeks. The more you use it, the faster it gets.

Before you hand CoWork the keys to your desktop, a few things worth doing first.

• Set a spending limit. Go into your account settings and cap your daily API spend—$5 is a reasonable starting range while you're still getting a feel for how tokens are used.
• Turn on action logging. CoWork can touch files, move things around, and automate tasks on your behalf. Logging gives you a paper trail. If something goes sideways, you'll know exactly what happened and when.
• Close what you don't want touched. Sensitive documents, financial software, password managers—close them before spinning up a session, especially in the early days.
• Know the kill switch. If a task is running and something looks wrong, close the CoWork window or force-quit the app. That's it—no drama. Just stop the process and assess.
• Start small. Point CoWork at low-stakes folders and simple workflows first. Build trust through small wins before handing it anything mission-critical.

IDEAL LOW-STAKES FIRST TASKS (IN ORDER OF RISK)

1. "Organize my Downloads folder" — Sort files into subfolders by type. Low risk, easy to verify, easy to undo.
2. "Research 5 competitors and compile into a spreadsheet" — Tests multi-step web browsing, data extraction, structured output.
3. "Summarize these 10 articles" — Tests reading comprehension and structured output.
4. "Rename these 30 files following this naming convention" — Tests precision and instruction-following.

WHAT TO WATCH FOR

• Is it clicking the right things? Watch it navigate and make sure it's going where you intended.
• Is it stuck in a loop? If it keeps repeating the same action, something went wrong—stop it and check.
• Is it making things up? Verify a sample of its outputs against the real source.
• Is it going places you didn't ask it to? It should stay within the task you gave it.
• Does it handle surprises? Pop-ups, login screens, unexpected pages—does it adapt or freeze?
• Is it cheap to run? A simple task shouldn't cost more than $1–3 in API fees.

• At least one task running reliably—not perfectly, but well enough that you trust it with light supervision
• Notes on what works—the exact instructions and settings that produced good results
• You know the kill switch—and you're comfortable using it if something looks wrong
• A rough sense of cost—you can predict roughly what each task will cost to run
• A short list of ideas—3 to 5 other things you'd like to try next
• Realistic expectations—you've seen what it's good at and where it needs help

Builds directly on Week 2's price data. The agent compiles the week's price changes, searches for industry news, competitor announcements, and market trends—then produces a formatted 2–3 page weekly report.

• Uses Week 2's spreadsheet as input, adds contextual analysis
• Human oversight: 30–45 min/week for editorial review
• Accuracy: 85–90% factual. Needs human polish for tone.

The agent takes exported leads from your CRM, researches each online, and enriches records with updated job titles, company size, recent news, LinkedIn presence, and tech stack.

• Human oversight: 20–30 min per batch of 25 leads
• Accuracy: 80–88% (job titles change frequently)
• Builds: An enriched contact database that feeds directly into Week 5

Given a template and a brief (client name, scope, requirements), the agent fills in the template with relevant info, pulls in case studies, and customizes messaging. Can use enriched CRM data from Week 4 to personalize.

• Human oversight: 45–60 min per proposal for quality review
• Accuracy: 75–85% content; formatting may need adjustment

The agent: identifies new leads → researches and enriches each → checks competitor pricing data for messaging angles → generates personalized outreach email drafts from templates.

• Human oversight: 30–45 min to review pipeline output and approve outreach
• End-to-end accuracy: 70–80% (error compounds across steps)
• Individual step accuracy: 85%+

• Changing your instructions without keeping notes: You tweak something to fix one problem and break something else. Fix: keep a running record of every version of your instructions so you can go back.
• Situations you didn't plan for: A website redesigns its layout. A PDF has an unusual format. The agent gets confused. Fix: build in a fallback—"if you can't find this, stop and flag it for me."
• Costs creeping up: Complex tasks visiting multiple sites can rack up surprising fees. Fix: set a cost alert so you know if a task is running over budget.
• Stopping checking because it's been fine: After 20 good runs, you stop looking. Run 21 has a real mistake. Fix: always spot-check at least 1 in 10 outputs, even on tasks that feel routine.
• Websites change and break your task: A site updates its layout and the agent can't navigate it anymore. Fix: write in a graceful failure—"if you can't find X, let me know instead of guessing."
• Trying to add too much too fast: Adding three new tasks a week instead of one. Fix: one task per week, done properly. Getting one thing working well is worth more than five things working badly.

WHERE THAT TIME COMES FROM

• Keeping an eye on competitors — 3–5 hrs/week
• Looking up and researching people or companies — 3–5 hrs/week
• Building and sending reports — 2–3 hrs/week
• Filing, organizing, and entering data — 2–4 hrs/week
• General research and fact-finding — 1–3 hrs/week

Running costs: roughly $150–400/month in fees, offset by 40–80 hours of recovered time. Most people break even in the first two weeks.

WHAT 90 DAYS GIVES YOU THAT YOU CAN'T SKIP

• Instructions that actually work: After 90 days of trial and error, your instructions are dramatically better than what you started with. There's no shortcut—you have to earn this through repetition.
• A feel for what it's good at: You'll know which tasks it handles confidently and which ones need more oversight. That instinct takes weeks to develop and can't be downloaded.
• Real accumulated data: 90 days of competitive research, organized contacts, and refined templates. Someone starting fresh three months later begins at zero.
• A new habit: The shift from "I do everything myself" to "I hand things off and review the results" takes 4–8 weeks to feel natural. There's no faster path.