AI in small and mid-sized firms: what works, what fails

Where does AI actually save time in service and manufacturing companies?

In four places: documents, enquiries, offers and company knowledge. They have one thing in common: lots of similar text that someone reads today, retypes and files in the right place. That's where AI is fast, cheap to run and easy to check.

Here's what it looks like at Reefa, my own company:

• Contracts. Before a contract goes out for signature, AI reads the parties, tax ID, company register number and signatories from the PDF. It takes 13 seconds.
• Staff documents. I send the bot a photo of a signed staff document, and AI works out whose it is, what type it is and which month it covers. Then it asks me to confirm with a button and asks about any missing months.
• Requests from the team. Staff ask for supplies by text or photo, in Polish, Russian or Ukrainian. AI splits the request into items from a 52-⁠product catalogue, and when something is unclear it asks a follow-up question with buttons.
• Tenders. An agent reads new public tender notices from BZP and TED every day along with the tender documents, judges whether it's worth bidding, and with one click drafts a package of five documents for the bid.

In a manufacturing company the pattern is the same; only the documents change: orders and specifications in PDFs, delivery notes and CMR consignment notes, complaints, questions about machine manuals. I don't run a factory, so I won't quote production numbers here. The principle works the same way, though: AI reads, the system records, a person approves the exceptions.

Where does AI fail?

AI fails where it lacks data, where a mistake is expensive or irreversible, and where the task is too broad. A model that doesn't know the answer will often invent one, in exactly the same confident tone as always.

Facts that aren't in the data

Ask AI to write about your company without hard data and it fills the gaps itself, for example with a client count or a rating nobody checked. That's why, since 4 October 2026, Reefa's automated blog has the company's facts written into the model's instructions. The content routine, which may edit pages and articles, has a guard: it blocks competitor names, unverified ratings and site counts.

Money

A language model isn't a calculator and shouldn't be the source of any amount. My morning brief on money, sales and operations deliberately runs on a fixed template without a language model, because it has to calculate the same way every time. Where an agent touches financial data, it may change it only on an exact match with a bank record or a document.

Decisions you can't undo

Sending an offer, approving a discount, answering a complaint: here AI drafts and a person sends. My outreach agent has a simple rule. If a reply contains a phone number, it stops writing and I make the call.

Jobs described as “everything”

An assistant that's meant to help with everything in practice helps with nothing specific, and you can't measure it. A rollout that still works six months later has one task and one measure: how many documents it read correctly, how many enquiries reached the CRM, how many minutes a draft offer takes.

Data that isn't in any system

If orders live in a notebook and prices in a salesperson's head, AI has nothing to read. Then the first step isn't AI but an automation or a system that collects the data in one place. I explain how to pick that first process in the guide to starting with automation.

How do you keep AI under human control?

A person approves everything that goes to a client or moves money, and the AI leaves a trace for every decision. In practice I use a handful of safeguards that carry over to any company.

• A confidence threshold. The bot that reads messages from the apartment cleaning crews reads the door number from a photo only when its confidence is at least 0.85. Below that, it asks the crew in the chat.
• Confirmation with a button. AI proposes, a person taps “yes”. That's how staff documents are filed.
• A second opinion. In content moderation on KtoSprząta.pl, editing or rejecting a submission needs a second agent to agree, and each run is capped at 30 decisions and 10 rejections.
• A decision log. Since 16 October 2026 my routines record every decision: date, question, options, choice, reasoning and how to roll it back. The log is append-only.
• A gate the agent can't get around. The outreach agent sends an offer only after the recipient agrees. That's enforced by a server-side gate with a suppression list, no repeats and a daily cap, not by a sentence in the prompt.

Which leads to the most important rule: I keep limits in code and in the database, not in a request to the model. A model can interpret a request its own way; it can't get past a gate on the server. The rules for my routines themselves sit in locked files with a SHA-256 checksum, so an agent can't change them without my approval.

An example: the AI routines at Reefa and KtoSprząta.pl and their limits

At Reefa and KtoSprząta.pl several AI routines run every day, checking data, fixing code and shipping changes on their own. Each has written rules about what it may not do, and leaves me only what it physically can't do: logging in, paying and signing.

I measured rolling back a change on the site at 3 seconds. That's what lets me allow a routine to deploy on its own: if the post-deploy check fails, the previous version comes back automatically.

To be straight about it: I haven't measured the effect of these routines on the number of enquiries yet, so I don't quote one. The oldest have run since June 2026, the newest since October. I measure what can already be measured: how many checks passed, how many decisions were logged, how long a rollback took. The full write-up is in the case study on AI agents that run part of the company.

GDPR and data security: the minimum I insist on

Send personal data only to business AI services you have a data processing agreement with, and only what the task needs. AI should see as much as the person it acts for, and keys and passwords never reach the model. This isn't legal advice, just the checklist I use on every rollout.

• A business account, not a personal one. I use models through the API or in business plans where the provider processes data on your behalf and you can sign a data processing agreement (Article 28 GDPR). Whether submitted data is used to train models, and where it's processed, I check in each provider's terms.
• Minimum data. To read a cost invoice, the model doesn't need an employee's national ID number. I send only the fields the task needs.
• The same permissions as a person. An assistant sees only what the person using it has access to. At Reefa a coordinator sees only their own sites and their own staff, and the database enforces that, not the screen.
• Secrets stay away from the model. API keys and passwords live only on the server. The model gets data for the task, never access to systems.
• A log. I record what went to the model and what came back, so any decision can be reconstructed.
• People are told. If AI processes client or employee data, that belongs in your privacy notice. Decisions that significantly affect people, such as rejecting a job applicant, I never leave to AI alone. The GDPR has a separate article on exactly that, Article 22.

Then there's the AI Act, Regulation (EU) 2024/1689. The obligation that applies to almost every company using AI comes from Article 4: a company has to take measures to build AI literacy among the people who work with these systems. In practice that means written rules for using AI and going through them with the team on examples from your own company. The European Commission's AI Act Service Desk publishes the current text, including later amendments.

If AI talks to clients directly, for example in a chat on your site, people should know they're talking to AI. Article 50 of the AI Act puts that obligation on the providers of such systems, and I treat it as a rule on every rollout.

What does an AI rollout cost a small or medium business?

Most of the cost is usually the rollout itself: connecting AI to your data, testing on real documents and setting up the controls. You pay the model provider by usage, and what that comes to is visible at the prototype stage, on your own data. My AI rollouts start from PLN 15,000 excl. VAT.

If you don't yet know where AI would help you most, a cheaper first step is the AI and process audit (from PLN 6,000 excl. VAT): two weeks of interviews with your team, a process map and a rollout plan with order and budget. For a quick picture at no cost, there's the AI readiness assessment: 12 questions and a score in five areas. All prices are on the pricing page.

How do you start so the rollout still works six months later?

Start with one narrow task, test the AI on real examples from your company, and give it autonomy only where you know how it gets things wrong. That's the order I follow on every rollout.

1. Pick one task with lots of similar cases, such as reading cost invoices or sorting emails from a shared inbox.
2. Collect a few dozen real examples together with the correct result. That's your test set and your baseline.
3. Start in “AI proposes, a person approves” mode and measure how many proposals go through without corrections.
4. Decide in which cases, and above what accuracy, AI may act without approval. Money and clients stay behind approval.
5. Name an owner who reads the log and gets the alerts, and schedule a review after one month and after six.

The first agent should check, not change. At Reefa, the morning brief and the tender agent, which only read and report, have run since June 2026. Routines that ship code changes on their own got that right only in October, once they had locked rules and automatic rollback.

If you want to find out whether AI can handle the documents or enquiries in your company, book a 30-⁠minute call. Scope, process and price are on the AI implementation page.