How I work
One task. A prototype on your own examples. A separate account for AI. A person approves anything with consequences. Everything logged.
Six principles
- On what you have. I do not rebuild systems from scratch. I fix and extend what works.
- In small steps. Every change checked, with a backup and a way back.
- On a separate account. Named access, only to what is needed. Handed back at the end.
- In plain language. Options with pros and cons instead of one right answer. No jargon unless you want it.
- For time worked. Single jobs, blocks of hours or ongoing support. Quoted per case.
- With notes. I write down what I did and why, so the next person does not start from zero.
Three rules for AI
AI sees only what the account it runs on sees. So I start with permissions.
- A person approves consequences. AI does not send, delete or pay without a person confirming, until it has proven it does not get it wrong.
- Every result leads to its source. A figure in a report leads to a record; an answer leads to a document and a paragraph.
- Everything is logged. What AI received, what it returned, who approved it. Without that you can neither explain a mistake nor learn from it.
Five steps
A call
One task that hurts the most today. How often, who does it, where the data comes from, how we will know it is better. 30–60 minutes, online.
A prototype in 1–2 weeks
On your real examples: recent reports, a month of emails, a hundred invoices. You check the results next to what a person did. A fixed price, agreed after the call.
Permissions and data
A separate account for AI in each system, read-only where that is enough, sensitive fields excluded, business accounts with the model vendors, a log.
Go-live and a month of watching
First for a few people or part of the traffic. We measure time, errors and tool costs. After a month, a decision: extend, keep, or switch off.
Care
Models and tools change, data ages, people find new questions. A review of the log and costs once a month, fixes within the monthly hours.
Data rules
What I agree and write down on every implementation. Details in the article on what not to paste into AI.
Anonymisation
I mask personal data before it goes to the model, or leave it in your systems. What goes to the model is what the task needs, not the whole record.
Business accounts, not personal ones
With the model vendors I use business plans and the API, with settings that switch off training on your data. Staff’s personal accounts get disconnected.
Vendor terms
I check and record in the documentation: whether data is used for training, how long it is kept, where. When the terms change, I come back to it.
A log
What AI received, what it returned, who approved it. Kept with you, not with me. You set the retention period.
Data stays with you
I do not copy your data to my machines beyond what the prototype needs, and I delete that when it is done. An NDA before first access, on request.
Data processing agreement
Where GDPR requires one. You agree the template with your lawyer; I supply the list of systems, data and vendors it needs.
Ways of working together
Quoted after a call, depending on scope. AI tool costs are shown up front, separately from my work.
A fixed-price prototype
One task, 1–2 weeks, a working version on your data. Then a decision on whether to go further.
Implementation as a job
Go-live, permissions, documentation, a month of observation. Scope in writing after the prototype.
A block of hours
Hours paid up front, used over several months. For companies that want to add further tasks at their own pace.
Care
A monthly pool of hours: review of the log and costs, fixes after model and system changes, new questions from the team.
Alongside training
Your team learns to work with AI in a training session; I connect what they cannot connect themselves: systems, accounts, the log.
Alongside your CRM or IT provider
They look after their systems, I look after the AI step and the connection. We split the scope in writing.
Access and security
- A separate, named account for me in each system, and a separate account for AI, different from mine. You can switch either off at any time.
- Read-only for AI wherever reading is enough. Writing only after the prototype, and for a narrow scope.
- Two-factor authentication on every account, passwords and keys in a password manager, never in an email or a chat.
- Changes first on a copy or on past data, then in production.
- A list of connections: what is connected, on which account, with what scope, until when. Reviewed quarterly.
- At the end I hand back access and leave the documentation and the list of connections.
A developer by the hour, an AI agency, an off-the-shelf tool, or nothing?
For a small company that wants one task sorted, what matters is time to a first result, access to your data, and who looks after it once it is live.
| Developer by the hour (me) | AI agency | Off-the-shelf SaaS tool | Doing nothing | |
|---|---|---|---|---|
| Starting point | One task and your real data | Workshops and an “AI strategy” | What the tool can do, not what you need | A person keeps doing the task, every week |
| Connecting to your systems | I do it myself: CRM, email, databases, website | Often a separate integrator or subcontractor | Only what it has a ready connector for | — |
| Permissions and data | Separate account, read-only, logged. Set up and documented | Depends on the team | Usually asks for full access to your mailbox or drive | Staff paste data into personal chat accounts |
| Time to first result | A prototype in 1–2 weeks | After an analysis phase | Immediately, if it fits; never, if it does not | — |
| Cost | Fixed price for the prototype, then time worked | A project with a budget running for months | Per-user subscription that grows with the team | Hours every week, invisible on any invoice |
| After go-live | A month of observation, then care by the hour | Support under the contract | The vendor’s help desk | — |
Questions about working together
Where do we start if we do not know which task to pick?
With a call. I ask what takes the most time and repeats every week. Usually after half an hour there is one obvious candidate. If there is not, I say so.
How long does the whole thing take?
A prototype in 1–2 weeks from access to the data, go-live usually within the following two, then a month of observation. You see the first result after two weeks, not after a quarter.
Do we sign a contract?
For implementation and care, yes: scope, data, confidentiality, access. For the prototype, confirmation by email is enough and, on request, an NDA.
Do you work remotely?
Yes, fully remote, in English or Polish. On-site meetings for larger projects, by arrangement.
What happens if we stop working together?
I hand back access and leave the documentation, the list of connections and the instructions. The prototype and the code are yours. The next person does not start from zero.
Is there a task someone at your company does by hand every week? Describe it in one sentence.
I reply within one business day. The first call is free of obligation.