AI for a business owner: where do you start? a plan for the first 30 days
You do not need an AI strategy for the whole company. You need one process that works after a month and has a number before and after. Below is a four-week plan: how to pick that process in an hour, what your team does without spending anything, where the work for someone outside begins, and how you will know after 30 days that it is worth going further.
- 30 dayslong enough to move from “we have heard about AI” to one working improvement, if you take a single process instead of the whole company
- 1 hourhow long the choice of the first process takes, once you have your team's tasks written down and six cut-off criteria
- EUR 0what the first two weeks cost: picking the process and measuring the “before” state is your own team's work
- EUR 0.005 to 0.07the cost of one query to an AI model, paid to the provider directly, with nobody's margin on top
Short answer
Where do you start with AI in the first 30 days?
Pick one process, not an area and not a strategy. It has to be repeatable, describable in rules, running on data that already sits in a system, and ending in a result you can count. The choice itself takes an hour if you sit down with a list of your team's tasks and the criteria in the table below.
Then four weeks: one to choose the process and get its owner to agree, one to measure the “before” state, one for a pilot running alongside normal work, one for a narrow production run and a decision. The first two weeks you run yourself and they cost nothing beyond your people's time.
Do not buy a tool in week one. Process first, tools second: a licence bought before the measurement sits unused, and after a month you still cannot say whether anything improved, because you have no “before” number.
Why start with one process instead of an AI strategy?
An AI strategy for the whole company looks good on a slide and changes nothing on Monday. The team is waiting for someone to say what they actually do differently from tomorrow, and that answer does not fit into a thirty-page document. One process gives it to them in four weeks.
There is a financial reason too. A single task has a countable cost today, so after a month you compare two numbers instead of trading impressions. The formula and a worked example are in the guide on automation ROI.
- One process has one owner, so the decision lands in a day rather than at a committee in three weeks
- Getting it wrong on one process costs a week of work; getting a strategy wrong costs a quarter and the credibility of the whole topic
- After the first working improvement the team stops debating AI in general and starts nominating tasks itself
- A result from one process goes to the board as a number, not as a plan for next year
How do you pick that one process in an hour?
Get two people who do the operational work and have them write down what they did last week. Not processes, tasks: “I retype orders from email into the system”, “I assemble a report out of three spreadsheets”, “I answer questions about where a delivery is”. In our workshops that list usually runs to 15 or 30 items and takes 20 minutes.
Then run the list through six criteria. A candidate that fails any one of them goes back on the list for later, not in the bin. One or two survive, and that is the whole decision of week one.
| Criterion | How you check it in 10 minutes | The candidate drops out when |
|---|---|---|
| Repeatable | Count from the calendar and the inbox how many times a week someone does it | it happens fewer than 5 times a week |
| Rules, not judgement | Try writing it as steps a new hire would follow | the decision depends on context that lives in no system |
| Data already in a system | Check that everything needed sits in the CRM, a spreadsheet or an inbox | someone has to retype paper or scans first |
| One owner | Name the person who can say “from tomorrow we do it this way” | three departments have to sign off |
| Countable result | State the number you will read after 30 days: hours, cases, response time | you cannot say what exactly you would compare |
| Cheap to get wrong | Answer what happens when the AI agent gets it wrong in week one | the error goes straight to a customer or into the accounts |
The criteria work as gates, not as a score. A process that passes five out of six is a candidate for your second rollout, not your first.
What does the 30-day plan look like, week by week?
This is the same order we use with clients: workshop, pilot on real data, production, measurement after 30 and 90 days. Cut down to one process and four weeks so you can walk it without us.
If you would rather not run weeks one and two yourself, that is what the process workshop is for, at EUR 3,500 net per week. You leave it with a process map, metrics and a 30-day plan, and the rollout gets quoted only after that.
Week 1: one process instead of a wish list
An hour with the team on the list of tasks, an hour running it through the six criteria, a short conversation with the process owner. You end the week with one sentence in writing: what we are automating and who decides about it.
Week 2: the “before” measurement and the edges
For five days you measure what happens by hand today: hours, number of cases, number of corrections, time from event to result. Alongside that you write down the exceptions, the situations where a person does it differently. Three metrics and the place each one is read from go into a spreadsheet.
Week 3: a pilot on real data, alongside the process
The AI agent gets the same cases a person does, but its output goes into a spreadsheet rather than to a customer. Take 20 real cases from last month and compare the answers. Fix the instructions, repeat, and only then start thinking about integration.
Week 4: narrow production and a decision
Turn it on for one team, one customer segment or one document type. Someone is on duty to fix things when they break. At the end of the 30 days you compare three metrics against week two and decide: extend, fix, or switch it off.
What do you do yourself for free, and where do you need help?
The line is sharper than it looks. Everything about your company (what happens, how long it takes, who decides) you do yourself and nobody will do it better. Everything that touches APIs, permissions and production is technical work, and that is where doing it yourself ends.
The first two rows of the table are your team's time, not an invoice. Only weeks three and four may need someone who writes the integration and picks up the alert on a Saturday morning.
| Task in the first 30 days | Do it yourself | When you call someone in |
|---|---|---|
| List of tasks and choice of process | Yes, an hour with the team | when after an hour you still have eight candidates and cannot choose |
| Measuring the “before” state for a week | Yes, a stopwatch and a spreadsheet | never, this is your people's work and nobody can stand in for them |
| Testing the instructions on 20 cases | Yes, in a browser, with no integration | when the output is good in 10 cases out of 20 and nobody knows why |
| Wiring it into the CRM, accounting or the shop | Rarely, and only with someone who knows APIs | when errors, rate limits and retries have to be handled |
| Go-live and being on duty | No | always, someone has to take the alerts and fix things in week four |
| Measuring after 30 and 90 days | Yes, on the metrics from week two | when the number disagrees with what the team feels |
Order matters: integration before the instructions are tested is the usual way to burn week three. First check whether the model does the job well at all on copied data.
What does the first month cost, and what are you paying for?
There are three line items, not ten: the AI model billed per query, the automation platform, and somebody's work if you are not writing the integration yourself. The rest is your people's time, which sits in the company's costs either way.
You pay the model provider directly, so nobody adds a margin along the way. A pilot on 20 cases closes in the region of a few euros, and that is the only spend you have to make before you see a first result.
- Weeks 1 and 2: EUR 0, because it is conversations, a list of tasks and a measurement done by your own team
- AI model: EUR 0.005 to 0.07 per query, paid to the provider directly, and the bill only grows with volume
- Automation platform: n8n self-hosted carries no licence fee, Make and Zapier per the vendor price list
- Process workshop: EUR 3,500 net per week, if you would rather not run weeks one and two yourself
- Rollout: quoted after the workshop, because the price follows the number of systems you have to plug into
How will you know after 30 days that it worked?
From the numbers written down in week two, not from how it feels. A metric you cannot read without asking a person does not belong on the list, because after a month nobody will read it. Three metrics are enough, and it is better if two of them come out of a system in one query.
The decision after 30 days has three versions and all of them are fine: extend it to the next team, fix one step and measure for another two weeks, or switch it off and go back to the candidate list. Switching off after a month is cheap. Dragging something that does not work through a quarter is not.
We measure our own rollouts the same way. Two AI agents have been running in production inside our founders' companies since March 2026, and at the performance marketing agency the margin in May, June and July 2026 was 5.55 percentage points higher year on year, with revenue up 1.5 percent. That is a coincidence in time, not proof of cause: the client mix changed over the same window. The full figures with the measurement method and the caveats are on the AI automation agency page.
- Time: hours on the task per week, counted the same way as in week two
- Volume: how many cases went through the AI agent and how many came back to a person
- Quality: how many answers had to be corrected, on the same 20 cases as in the pilot
- Adoption: how many people used it last week, because a rollout nobody touches still looks successful in a report
Common mistakes
Five mistakes of the first month with AI
Starting from an AI strategy for the whole company
After a quarter you have a document and zero working improvements. The team files the topic under initiatives that blow over, and engages less on the second attempt.
Buying a licence before choosing the process
You pay for a tool nobody can use, because nobody knows what it was for. Three months later the subscription is a cost line nobody questions.
Picking the first process because it hurts the most
What hurts most is usually the process full of exceptions with three departments inside it. The pilot bogs down in sign-offs and after a month there is nothing to show.
No measurement of the “before” state
After 30 days there is nothing to compare the result against. You are left with “it feels faster”, which does not survive a board meeting or the next budget round.
Running the pilot straight at customers
The first mistake goes outside and the topic goes back in the drawer for a year, no matter how good the other nineteen answers were.
Checklist
Checklist before the first 30 days start
If any line is a no, week one is not over yet.
- One process chosen, not an area and not a shortlist of three candidates
- The process owner knows the plan and agrees to change how the team works
- The task happens at least 5 times a week
- The data needed for the decision is already in a system, not on paper
- The “before” state measured over five days: hours, number of cases, number of corrections
- Three metrics written down together with the place each is read from
- The pilot runs alongside the process and a person checks the output before it reaches a customer
- The 30-day measurement date in the calendar, with the name of the person who will do it
Your first process
Let us pick that one process together
On a free 30-minute consultation we go through your list of tasks and point at the candidate for your first 30 days. If none of them is ready yet, we say so on the call. The process workshop costs EUR 3,500 net per week and ends with a process map, metrics and a 30-day plan you can execute without us. You write directly to Pawel, the founder, and we take at most two rollouts at a time.
Book a free consultationFrequently asked questions
Where do you start with AI in a company?
With one repeatable process, running on data you already hold in a system, ending in a result you can count. The choice takes an hour: your team's list of tasks run through six cut-off criteria. A strategy for the whole company can wait until one improvement is actually working.
How long does a first AI rollout take?
Four weeks for a single process: one week to choose it, one to measure the “before” state, one for a pilot alongside normal work, one for narrow production and a decision. It takes longer when the process crosses three departments or the data has to move off paper first.
Is ChatGPT enough to get started?
For testing the instructions in week three, yes, and that is its right place in this plan. For daily work it is not enough, because a generic tool cannot see your CRM, your invoices or your history with a customer, so you paste the context in by hand every time. An AI agent wired into company data reads it itself.
How do you choose the first process to automate?
Check six things: does it happen at least 5 times a week, can it be described in rules, is the data in a system, does it have one owner, is the result countable, and does a mistake stay away from the customer. A candidate that fails any one of them is good for your second rollout, not your first.
What does the first month with AI cost?
The first two weeks cost EUR 0, because they are your team's time. Then comes the AI model at EUR 0.005 to 0.07 per query paid to the provider directly, plus the automation platform: n8n self-hosted with no licence fee, Make or Zapier per the vendor price list. The process workshop is EUR 3,500 net per week, and the rollout is quoted after it.
Do I need an IT department to start?
Not for weeks one, two and three. Choosing the process, measuring it and testing the instructions on 20 cases all happen in a browser and a spreadsheet. A technical person is needed when you wire it into the CRM or accounting, and for the duty rota in week four when something breaks outside working hours.
What if the result is weak after 30 days?
Find which of the three metrics is out of line and pick one of three moves: fix a single step and measure for another two weeks, narrow the process to a simpler slice, or switch it off and go back to the candidate list. Switching off after a month costs a week of work, so it is cheaper than keeping something the team walks around.
Should I start with an agent for myself or for the whole team?
With yourself, or with one role, because then you decide alone and see the result daily. A personal AI agent works for one person; BOND answers the whole company from its data in 5 to 10 seconds. The first month is easier to walk on a narrow scope, and extending it to the team is the natural next step after the first measurement.