Skip to content

Plan with a worksheet

An AI agent in your company: a plan for the first 30 days

The first 30 days with an AI agent look like this: on day 0 measure today, in week 1 understand the work, in week 2 the agent drafts and a person approves, in week 3 measure and fix, and in week 4 decide whether to continue. Before that, check whether you need an agent at all and pick one process in the worksheet below. Most companies need an automation.

Downloadpdf

What you fill in and tick stays in this browser window only. We do not store or send anything. It is gone when you close the page, so print your result.

For owners, operations people and CTOs who want to hand boring work to software. Customer emails, orders or invoices, for example. The worksheet takes 30 minutes, before your first conversation with a vendor.

Since 2023 we have delivered 15+ production projects. We run agents and automations for clients and in our own company.

1. An agent, an automation, or nothing?

Most companies that want an AI agent need an automation. An agent makes sense where free text has to be read and decisions made within limits.

AssistantAutomationAgent
Who decidesA human at every stepA rule, a human for exceptionsThe agent, within limits
ExampleA draft text in an AI chatAn invoice from email into accountingA reply to a customer's question about an order
What you needAn accountA described process and examplesProcess, examples, rules and who takes the exceptions
Worst mistakeA bad text in one documentThe mistake repeats until someone stops itThe agent tells a customer something wrong in your name

An agent only works with an automation underneath. Without one it has nothing to read from.

2. When not to build an agent

If even one of these applies, start somewhere else.

Sometimes the best answer is to tidy a spreadsheet, or to do nothing. That's a result too.

3. Choosing the first process: a worksheet

Write down 2 or 3 candidates. Give one point for every "yes".

Question
Does it happen every day?
Does the input have a similar shape (email, invoice, order)?
Do you have examples of correct replies or results?
Can you name the exceptions and who gets them?
Can a mistake be fixed before the customer sees it?
Will the saved time go to sales or to customers?
Total (out of 6)000

Start with the highest score. On a tie, pick the one where a mistake is cheaper.

Three questions we ask before every project: Will it make money? Will it save money? Will it save time, for whom, and what will that person do with it? If you can't answer any of them, you don't need an agent yet.

4. The 30-day plan

Day 0: measure today

Without a baseline you won't know whether the agent helped. Time 10 typical cases with a stopwatch. How long one takes, who does it, how many there are each week.

Week 1: understand the work

Week 2: the agent drafts, a human approves

Week 3: measure and fix

Week 4: decide

What you seeDecision
You approve most drafts without changes and approving has become boringLet this type run on its own. Everything else still goes through a human.
You correct often, but less than in week 2Another month in draft mode. Keep fixing the material.
You correct more than the agent savedStop. Go back to choosing the process, or to an automation.

5. What to measure

NumberHow to get itWhy
Time per caseStopwatch on 10 cases, before and afterThe main proof that it helps
Share of drafts approved unchangedApproved / all draftsTells you when to let a type run on its own
Share of cases handed to a humanEscalations / all casesToo low is suspicious, not good
Mistakes the customer sawWrite down every oneThe only number that has to be zero
Time spent approvingEstimate from the person who approvesIf it's higher than the saving, something is wrong

6. What it looks like in practice

A client with an online shop: customer support. A reply to a customer used to take hours; today it arrives within 5 minutes. Support now handles only the complex cases. The first version, though, answered like a marketing flyer. What helped was rewriting the material in the founders' words and one rule: "when you don't know, hand it to a human".

Our own accounting: invoices. No agent, just automation. One invoice took 10 to 15 minutes; today it takes under a minute. 70 % go through without a human. The rest goes through a human on purpose: scans, new suppliers, several VAT rates.

7. Limits that apply from day one

Frequently asked questions

How do I introduce an AI agent in my company?
One process, one type of case, and for the first weeks the agent only drafts while a person approves. Every correction goes back into the agent’s material. Let a type run on its own only when you approve its drafts without changes.
When should I not build an AI agent?
When you need full accuracy without any human check, cannot describe the procedure in one paragraph, have no examples of correct results, or nobody has time to approve drafts. Then start with an automation or with cleaning up the data.
What should I measure?
Time per case before and after, the share of drafts approved unchanged, the share of cases passed to a person, errors a customer saw, and the time approval costs. Errors a customer saw must be zero.
What is the difference between an AI agent and an automation?
An automation follows a rule and a person handles the exceptions. An agent reads free text and decides within limits, for example answering a customer’s question about an order. An agent only works with an automation underneath it.

Not sure which process to start with?

Tell us which processes scored highest in your worksheet. We get back to you within two business days, and together we pick what to fix first.