Your CEO told you to build agents. What to do in the first two weeks

By Izzy Aly, Co-Founder, Systems and Architecture


Your CEO or founder wants AI agents and has asked you to build them, alongside your current work, and you want the first one to go well even though you haven't built one before.

We think what your CEO wants is to see the company make a start with AI agents, the way they see other companies doing. Over the next two weeks you can give them that start by picking one job to automate, using a connection or a product wherever one already does that job, and putting a price on building the rest. By our estimate it takes 5 to 10 hours of your time in total, which is about an hour a day on four or five of those days, or one half day in each week.

This post walks you through those two weeks and gives you two prompts to paste into Claude or ChatGPT: one writes the spec with you, and one drafts your plan for the CEO. If you hold a rollout mandate for the whole team, Why AI Rollouts Stall by Month Three is written for you.

Most of your hours go into writing the spec in week two Our estimate of your own time per step, for one job. Week 1, find the job and check what already exists: ask the CEO the four questions, 30 minutes to 1 hour; check connections and try one product, 1 to 2 hours. Week 2, write it down, count the upkeep and send the plan: write the spec with the prompt, 2 to 4 hours with the person doing the job; count the upkeep, 30 minutes to 1 hour; draft the plan with the prompt, 30 minutes to 1 hour; walk the CEO through the spec, 15 to 30 minutes. OUR ESTIMATE OF YOUR OWN TIME PER STEP, FOR ONE JOB Week 1 Find the job and check what already exists Week 2 Write it down, count the upkeep, send the plan Ask the CEO the four questions 30 minutes to 1 hour Check connections, try one product 1 to 2 hours Write the spec with the prompt 2 to 4 hours, with the person doing the job Count the upkeep 30 minutes to 1 hour Draft the plan with the prompt 30 minutes to 1 hour Walk the CEO through the spec 15 to 30 minutes
Most of your hours go into writing the spec in week two Our estimate of your own time per step, for one job. Week 1, find the job and check what already exists: ask the CEO the four questions, 30 minutes to 1 hour; check connections and try one product, 1 to 2 hours. Week 2, write it down, count the upkeep and send the plan: write the spec with the prompt, 2 to 4 hours with the person doing the job; count the upkeep, 30 minutes to 1 hour; draft the plan with the prompt, 30 minutes to 1 hour; walk the CEO through the spec, 15 to 30 minutes. OUR ESTIMATE OF YOUR OWN TIME PER STEP Week 1 Find the job and check what already exists 01 Ask the CEO the four questions 30 minutes to 1 hour 02 Check connections, try one product 1 to 2 hours Week 2 Write it down, count the upkeep, send the plan 03 Write the spec with the prompt 2 to 4 hours, with the person doing the job 04 Count the upkeep 30 minutes to 1 hour 05 Draft the plan with the prompt 30 minutes to 1 hour 06 Walk the CEO through the spec 15 to 30 minutes
Fig. 1 — The two weeks · 6 steps and our estimate of your hours for each

Key takeaways

  • Asking the CEO what should change and how you'll know it worked gives you a job with an owner and a finish line in place of an open-ended project.
  • Check what the company already pays for first, including Claude or ChatGPT connected to your tools, and try a new product only when you can test it on a free version, a trial or a monthly plan you can cancel, so walking away costs nothing.
  • Only the job that crosses your own tools and depends on your own templates and rules needs a build, because nobody sells a product that knows those, and you can build it yourself or pay an outside team to do it.
  • Writing the job down as a spec needs no code and is never wasted, because whoever builds the job later starts from your page and not from a guess.
  • Send the CEO a short plan by email with your spec attached, and when it names one job, what you can use now, what each option costs including upkeep and what you need from them, it reads as progress, which is what the request was about in the first place.

What is your CEO asking for?

Agents can mean anything from a chatbot on the website to a system that sets up every new client across the CRM, the project tool and billing, so start by asking the CEO which job they have in mind. These four questions get you there:

  1. What should be different once this works, and how will we know it worked?
  2. Which job did you have in mind, who does it today and how often does it happen?
  3. What should it do on its own, and what should a person still check or approve?
  4. By when do you want to see it working, and who else needs to agree?

The answers to the first and third questions become the done criteria and approval points in your spec, and you know which job matters to the CEO, who does it today and what done looks like, so you can check whether a connection or a product already does it.

Does a connector or a product already do this?

Often one does, and switching it on is the quickest progress you can show. If your team already uses Claude or ChatGPT, connecting it to your CRM and your other tools through its built-in connectors or a third-party integration tool can cover the job without anyone building an agent.

Check in this order:

  1. What the company already pays for, such as those connections, meeting summaries in your video call tool, or your CRM's built-in AI (Breeze in HubSpot, Agentforce in Salesforce), which may already draft the follow-up or update the record.
  2. A product that does the job as it comes, if you can try it on a free version, a trial or a monthly plan you can cancel. When it needs an annual contract, list it in your plan as an option with its price and let the CEO decide.
  3. A build, by you or an outside team, for what the first two don't cover.

When a connection or a product does the job, use it. Try it on a real task this week and write down what it missed, since that is the list you take into week two.

Which one job is worth building?

The job worth building is the one no connection or product covers, usually because it runs across your own tools and follows your own rules. If you have more than one candidate, the Toil Test helps you pick the first.

Then write it down as a spec, one page that says what starts the job, what it reads, what it does, where it writes, what can go wrong and who signs off, so whoever builds it works from the same description.

The spec takes the most time of anything in these two weeks, because you sit with the person who does the job, collect a few real examples and list the exceptions, and none of it needs code.

The prompt below does the writing. Paste it into Claude or ChatGPT with the CEO's answers to the four questions to hand, answer its questions with the person who does the job, and it turns everything into a one-page spec with the same fields as our AI Workflow Lab spec cards.

You're helping me write a one-page spec for one job I want to automate with AI. Interview me first, then write the spec.

Ask me one question at a time and wait for my answer before the next one. Ask a follow-up when my answer is vague or I say it depends. Don't invent steps, tools, numbers or names. If I don't know an answer, mark it as open.

Get an answer for each of these fields:
1. Objective: what should be different once this works, and how we'll measure it.
2. Trigger: what starts the job.
3. Inputs: what the job reads, and where each piece lives.
4. Steps: what a person does today, in order.
5. Systems touched: every tool it reads from or writes to.
6. Outputs: what exists when the job is done.
7. Exceptions: what can go wrong or look unusual, and what should happen then.
8. Approval points: what it can do on its own and what a person must check first.
9. Done criteria: how we'll know each run worked.
10. Volume: how often it happens and how long it takes a person today.

When you have everything, write the spec as one page with a heading per field, in plain sentences. End with a list of open questions and anything you assumed, so I can check them with the person who does the job.

Start by asking me which job this is and who does it today.

An agent can only follow a spec if it knows what your fields mean, for example which stage counts as closed and which field holds the signed scope. If your team hasn't agreed on that yet, How to tell AI agents what your CRM fields mean shows how to write it down in a day.

Who keeps it running after it's built?

Once a job is live, someone has to check that it works and update it when your tools or the model change. That costs your hours if you do it yourself, or a monthly fee if an outside team does it, and either way it belongs in your plan next to the build.

These four questions put a number on it:

  1. Who checks that it's working, and how often?
  2. Who updates it when one of the tools changes, or a new model changes what the agent writes?
  3. Who covers it while you're on holiday?
  4. How many hours a month will that take, and whose hours are they?

Write the answers down with real names and a real number, because that number goes into your plan. If an outside team keeps it running for you, you still approve the spec, test the output and own the result.

If you build it yourself, Claude Code for RevOps shows how to start. Build one small job yourself when the time is agreed and the upkeep is counted, and bring in an outside team when the list is longer.

What do you send your CEO?

At the end of the second week you send the CEO a short plan by email, with your spec attached, and it has five parts:

  1. The job you'd start with, and how you'll know it worked.
  2. What you can use now: a connection, a product on trial or something the company already pays for.
  3. What still needs to be built, with your spec attached.
  4. What each option costs, including the upkeep.
  5. What you need from the CEO, either protected time or a budget.

The prompt below drafts that email in Claude or ChatGPT from your notes and your spec. It asks you for anything missing instead of filling the gap itself, so every hour and price in the email is one you gave it. Use the account your company has approved, and leave client names and account data out.

You're helping me write a short email to my CEO. They asked me to build AI agents. I've spent two weeks working out what to do first, and I want to reply with a plan, not a yes or a no.

Use only the facts in my notes below. Don't invent hours, prices, product names or dates. If anything on the checklist is missing from my notes, ask me for it before you write, one question at a time.

Checklist:
1. The CEO's original request, in their words if I have them.
2. The one job I'd start with: what it is, who does it now, how often, how many hours a week it takes, what should be different once it works and how we'll measure it.
3. What we can use instead of building: a connection between our AI assistant and our tools, a feature we already pay for, or a product on a free version, trial or monthly plan. Say what it covers, what it costs and when I can set it up. If a product needs an annual contract, list it as an option with its price. Or say that nothing we found fits.
4. What needs to be built: the job, the tools it runs across and why no product does it. I'll attach a one-page spec. If nothing needs building, say so.
5. What building it costs: my hours to build it, my hours each month to keep it running, and what I'd stop doing to make room. If we bring in an outside team, their price and monthly fee, or say the price isn't known yet.
6. What I need from the CEO: a decision on the purchase, and either protected hours for the build or a budget to get a fixed price.

How to write it:
- Open on the CEO's request, not on my workload, so it reads as progress.
- Keep it under 200 words in plain sentences. Number the parts of the plan.
- Always give the upkeep cost, not just the build cost.
- Give the CEO two choices in the ask, time or budget, and make both sound like fine answers.
- No hype, no buzzwords, no exclamation marks.
- End by offering a 15-minute walk-through of the spec this week.

Give me a subject line and the email. Then list anything you assumed, so I can check it before I send.

My notes:
[paste your notes, your spec and your estimates here]

Read the list of assumptions at the end before you send anything, and fix any number you didn't give it. If you'd rather not paste company notes into an assistant, the five parts above are enough to write the email yourself.

When does an outside team make sense?

An outside team makes sense in two situations. The first is a list of several jobs that nobody can put a price on, which is what our AI Workflow Lab is for: in two to three weeks it turns the list into a ranked backlog of spec cards with a fixed build price on each, a training plan for your team, a map of the data and approvals each job needs, and one working prototype.

The second is one clear job you don't want to build yourself. You can skip the Lab for that and get a fixed-price proposal for the build and, if you want, a monthly support plan to keep it running.

The cost part of your plan needs a price if you're asking for a budget, and that's the one number you can't work out alone. We build AI workflows that run across the tools your team already uses, with coding agents like Claude doing the work. Book a free 30-minute workflow scoping call and bring your spec, and we'll tell you whether that job can go straight to a fixed-price proposal or whether a longer list needs the Lab first.

FAQ

What if my CEO wanted an agent and I come back with a connection or a product?

Show the CEO it working on one of their own tasks, next to the spec for the job nothing covers yet. Something that runs this week and a priced plan for the rest is the start they asked for.

What if the CEO says no to both the time and the budget?

You still have what you switched on, and the spec is ready for when the answer changes. Ask the CEO which of your current work they'd move to make room, so the conversation becomes a choice between priorities.

We already pay for HubSpot's Breeze or Salesforce's Agentforce. Do we still need any of this?

Start with it, because it already sits in your CRM and the company already pays for it. What's left to build is the steps it doesn't reach in your other tools, such as your project tool and team chat, so the two work side by side.

Author
Izzy Aly · Co-Founder, Systems and Architecture
Share

Work with us

Got more workflows than time to fix them?

The AI Workflow Lab maps where AI fits across your CRM and the tools around it, and tells you what to build first and what each build will cost.

or book a scoping call (opens Cal.com)