AI chatbot for small business: should it answer your customers?
An AI chatbot for small business can answer messages while you sleep, but it can also lose a lead politely. When it helps, when it hurts, and how to test it.
AI
AI implementation for business — from ready-made assistants to a private LLM
We start from the process, not the model: ChatGPT, Copilot or Gemini set up for your team, AI agents built into your workflows, or an open-weight model deployed and trained on your own infrastructure.
When this is about you
The team spends hours on copy-paste: emails, documents, reports, data entry.
Employees already use ChatGPT on their own — with company data and no rules.
You tried an AI tool, it gave confident wrong answers, and the pilot quietly died.
Your data can't leave the company, so cloud AI seems off the table.
What we implement
Entry point
AI process audit
A map of where AI saves hours and where it adds risk: tasks, volumes, data, owners.
ChatGPT · Copilot · Gemini · Claude
Ready-made AI for the team
Business plans set up for the company: accounts, data rules, usage policy, training.
n8n · Make · APIs
AI agents & automation
Agents that sort requests, draft replies, pull data from documents and update the CRM — with a person approving where it matters.
RAG
Knowledge assistant
An internal assistant that answers from your contracts, manuals and wiki — with links to the source.
Llama · Mistral · Qwen
Private LLM deployment
An open-weight model on your servers or a dedicated EU cloud: prompts and documents never leave your perimeter.
LoRA · test sets
Fine-tuning & evaluation
Training a model on your examples for tone, format and domain terms — measured against a test set before and after.
How an AI implementation runs
Map the process
Where the hours go today: tasks, volumes, who does what and with which data.
Pick the use cases
Ranked by hours saved and by risk; we start with the one that pays back fastest and is easiest to measure.
Data & compliance
What data the task touches, where it may be processed, what GDPR and the EU AI Act require.
Choose the approach
Ready-made tool, automation, RAG or a private model — by the matrix below, with written reasoning.
Pilot on real work
A small group, real tasks and a test set of your own examples to check answers against.
Evaluate
Accuracy, time saved and errors compared with the manual process — not impressions.
Roll out & train
Access, rules of use and training for the team. AI without rules of use is a data leak waiting to happen.
Monitor & improve
Regular quality checks, model updates and the next use cases.
Which approach — the honest matrix
From a ready-made assistant to your own model — we choose by the task, the data and the volume, not by the hype:
| ✓ | ✕ | |
|---|---|---|
| Ready-made assistants (ChatGPT, Copilot, Gemini, Claude) | A fast start for the whole team; strong models; admin controls and business terms for data | Data is processed in the vendor's cloud under contract; limited depth of integration |
| AI inside your current tools | CRM, helpdesk and office suites already ship AI features — nothing new to adopt | You get what the vendor built, not what your process needs |
| Automation + model APIs (n8n, Make, code) | AI in specific steps: sorting mail, drafting replies, moving data from documents into the CRM | Needs monitoring and error handling; doesn't fix a broken process |
| RAG — answers from your documents | An assistant that cites your own contracts, manuals and wiki; knowledge updates without retraining | Only as good as the documents behind it — a messy archive gives messy answers |
| Private LLM (open-weight model on your servers) | Data never leaves your perimeter; predictable cost at high volume; full control over versions | Hardware or GPU cloud and someone to run it; smaller models than the top cloud ones |
| Fine-tuning | A fixed tone and format, domain vocabulary, a small model doing one narrow job well | Not a way to teach facts — that's what RAG is for; needs a solid set of good examples |
Our principle: the simplest approach that meets your data requirements wins. A private model when the data demands it — not for prestige.
Honest expectations
- Not every task needs AI. If a checklist or a simple automation does the job, we'll say so — and build that instead.
- AI makes mistakes. For anything customer-facing or legally binding, a person approves — we design that step in from the start.
- We measure before rollout: a test set of your real examples and a clear metric. "It looks smart" is not a result.
- Your data, your assets: prompts, knowledge bases, fine-tuned models and accounts belong to your company.
First case studies are being prepared
We publish cases only with client consent and verified numbers. Until then — our method and checklists above are the most honest thing we can show.
From the blog
Claude Opus 5.5: Fable-level work at 40% lower cost than Opus 5
Claude Opus 5.5 launched on 22 September: $4/$20 per million tokens, Fable 5.1-level results, 30% faster output. What changed and what it means for teams.
ChatGPT Sponsored Agents: OpenAI tests ads that talk back
OpenAI is testing ChatGPT Sponsored Agents: click an ad, then chat with the brand's agent. Plus AI campaign tools and HubSpot and Shopify integrations.
Frequently asked
Where do we start with AI in our company?
With one process, not with a tool. The audit shows where the hours go and which tasks AI can take over safely; we start with the use case that pays back fastest and is easiest to measure.
Is our data safe with ChatGPT, Copilot or Gemini?
Business plans of these assistants come with contractual data terms and admin controls, and the terms differ from vendor to vendor. Whether they are enough depends on your data and your industry — we check them against your requirements before rollout.
What is a private LLM and when do we need one?
An open-weight language model running on your own servers or a dedicated cloud, so prompts and documents never leave your perimeter. It makes sense for sensitive data — health, legal, finance, trade secrets — strict client contracts, or high volumes where per-request pricing adds up.
Do we need to fine-tune a model?
Usually not at first. Most business tasks are solved by a good model, your documents (RAG) and clear instructions. Fine-tuning pays off when you need a fixed tone or format, domain vocabulary, or a smaller, cheaper model doing one narrow job — and when you have enough good examples.
How much does AI implementation cost?
It depends on the approach: ready-made tools are mostly licences and setup; a private LLM adds hardware or GPU cloud and operations. We quote after the audit — with the running costs calculated, not only the build.
What about GDPR and the EU AI Act?
We map which rules apply to your use case: telling customers when they talk to AI, data processing agreements, where data is stored and for how long. We are not your lawyers, but we build so your lawyers can sign off.
Start with an AI audit
Tell us which process eats the most hours — we'll answer with the use cases worth piloting and what they need from your data.
Scope runs from a single task to running the discipline end to end — no fixed packages, the range comes after the brief.
Related services:
Thank you — what happens next
- We reply within one business day.
- A short call to understand the task.
- We propose an entry point — usually an audit.
Want a sharper estimate? Fill in the brief — it arrives as an addition to your enquiry.
Fill in the brief