This text is written as Marcus Sheridan advises in the book „They Ask, You Answer” – first, the questions you already have in mind, then concrete answers. Without hiding drawbacks and without making empty promises.
If you are looking for answers to any of the following questions, you've come to the right place:
- What exactly is this „AI” since everyone is using the word too much?
- How do ChatGPT, Claude, and Copilot differ – and which one to choose for your team?
- What is the cost per year – not „from,” but the order of magnitude?
- What could go wrong (hallucinations, data, lack of source attribution)?
- Where to start so as not to burn through the budget at the beginning?
Below you will find it directly – with examples from service companies and offices.
What is artificial intelligence – really, not marketing hype?
One email with a concrete point each week.
AI, B2B sales, and implementations. No spam, unsubscribe with one click.
Artificial intelligence (AI) This is a broad concept: programs that imitate elements of human „intelligence” – they recognize patterns, understand text, and support decisions. It’s as broad as the word „vehicle,” encompassing both a bicycle and an airplane.
The problem is that today you see the inscription „with AI” on a washing machine, in spam filters, and in ChatGPT. It's fundamentally different Technology. To make a sensible comparison of implementation offers, you need to know hierarchy of concepts.
Hierarchy: AI → Machine Learning → Deep Learning → LLMs
Imagine it like a matryoshka doll – each subsequent level is a subset of the previous one:
Artificial Intelligence (AI)
└── Machine Learning (ML)
└── Deep learning
└── Large Language Models (LLM)
└── ChatGPT, Claude, Gemini…
AI – broadest category: from simple rules to complex models. The term has been in use since the 1950s.
Machine Learning (ML) – systems that they learn from data, instead of having every rule written by hand. For example, the model analyzes thousands of transactions and detects fraud patterns on its own – no one explicitly told it: „if the amount > 10,000, it's fraud.”.
Deep learning – a subset of ML based on neural networks. This is where facial recognition, translations, and autonomous assistance, among other things, come from.
Large Language Models – models trained on huge amount of text. They understand natural language, write, summarize, and help analyze documents. LLMs are behind ChatGPT, Claude, and Gemini.
Why does this matter when talking to a supplier?
Because when someone says, „we will implement AI in your company,” you have the right to clarify: what / which AI. Invoice scan recognition is often classic ML. A chatbot for customer questions - LLM. Demand forecasting - usually ML, not chat. Each variant represents a different budget, time, and risk.
This article focuses on Large Language Model, as they are the loudest in the context of „AI for business” and offices (law firms, consulting, HR, finance).
What are ChatGPT, Claude, Gemini, and Copilot?
It's commercial products built on language models. They differ in company, model, and ecosystem.
| Product | Company | Model (example) | Why will you find him |
|---|---|---|---|
| ChatGPT | OpenAI | GPT-4o, o1 | Widest recognition, fast start |
| Claude | Anthropic | Claude 4 | Long documents, strong emphasis on security |
| Gemini | Gemini 2.5 | Integration with Google Workspace | |
| Copilot | Microsoft | m.in. GPT-4o | Built into Word, Excel, Outlook, Teams |
| Llama | Meta | Llama 3 | Open source – self-hosting capability |
Analogy LLM is an engine; ChatGPT, Claude, or Gemini are cars of various brands - they all drive, but they have different handling, different prices, and different equipment.
What's the difference in practice (honestly)?
| Feature | ChatGPT | Claude | Gemini | Copilot |
|---|---|---|---|---|
| Document analysis | Good | Very good (long files) | Good | Very good (in Office) |
| Writing texts | Very good | Very good | Good | Good |
| Less „making things up out of thin air.” | Good | Very good | Average | Good |
| Integration with work | Weaker | Weaker | Google Workspace | Microsoft 365 |
| Price (business, order of magnitude) | ~20–30 EUR/person/month. | ~20–25 EUR/person/month. | ~20–30 EUR/person/month. | ~25–30 EUR/person/month. |
There is no single „best.” It depends on, What is the team working on i what questions must it ask the tool most often.
Glossary – so you know what is being discussed in the agreement
Prompt
Prompt = what you input into the model – question, command, context. The quality of the response depends on the quality of the prompt: the clearer the goal, the fewer corrections.
Token
A token is a billing unit. – text fragment (often ~1-2 tokens per word). You pay for tokens – longer questions, longer answers, large files = higher bill. A single conversation is usually hundreds to thousands of tokens; analyzing a thick PDF can reach tens of thousands.
Hallucination
Hallucination = the model „makes up” facts that are not there. It sounds credible, looks professional – and is untrue. For example, it provides a ruling citation that doesn't exist in databases. The model doesn't „lie” like a human – generates statistically probable text, not necessarily true. In law, medicine, and finance there is always human verification.
SaaS
Software as a Service – you pay a subscription, you log in through a browser or an app. ChatGPT Teams or Microsoft 365 are SaaS. The opposite is on-premise – your software on your server.
RAG (Retrieval-Augmented Generation)
RAG = the model first searches your documents, then answers based on that, with the ability to point to sources. This is a typical path when the question is „does the model know ours agreements, not just the internet from 2023?”.
Agent AI
An agent is not just a chat response, but a chain of actions – e.g., find a contract, calculate interest, prepare a letter template, prepare a file. The difference is like between a call center and an assistant who actually gets things done.
Three ways to implement AI in a company (from the simplest)
Path 1: Ready-made tools (SaaS)
What it consists of: You buy licenses and you start. Without an implementation plan and without a development team „underneath.”.
| Tool | What are you doing with him at work | Government price | How to start |
|---|---|---|---|
| ChatGPT Teams | Questions, pasting parts of documents, writing | ~25–30 EUR/person/month. | openai.com |
| Microsoft Copilot | Help with Word, Excel, Outlook, Teams | ~25–30 EUR/person/month. | Microsoft 365 Licenses |
| Claude Pro | Long documents, analysis | ~20–25 EUR/person/month. | claude.ai |
| Google Gemini | Gmail, Docs, Sheets | ~20–25 EUR/person/month. | Google Workspace |
Example: you paste a contract excerpt and ask: „List the penalty clauses in bullet points.” You get a draft response in a few dozen seconds.
Pros: Start immediately, low barrier to entry, predictable monthly cost.
The downsides (let's be blunt): model doesn't know the entire archive – you only get what you paste. Lack of a central database for „all our documents” with annotations. The data still goes to the provider's cloud (in business plans usually are not used for training model – but you must have this in the contract / DPA).
For whom: small teams, first few months, learning to work with LLMs, simple text tasks, and single file analysis.
From the perspective of „what the customer wants to know”: before you tell the client „we have AI”, establish internally, What questions does the team ask most often? And is chat enough? It's the same mechanism that underlies good digital bidding and calculations – organized answers to frequently asked questions, instead of chaos in emails.
Road 2: AI in Company Documents (RAG)
What it consists of: The system is indexing your files and responding. based on it, with references to source fragments.
Example question: „What penalty clauses do we have in our agreements with client ABC?”
Example Response (Schema):
„In the agreement dated 01/15/2024, the penalty is 10% of the order value [1]. In the addendum dated March 2024, it was raised to 15% [2].”
[1] Contract_ABC_2024.pdf, p. 8
[2] Annex_ABC_03_2024.pdf, p. 2
Difference from Road 1: You don't have to remember, in which file It was „that sentence about penalties” – the system searches the repository. You can also separate access (roles, document „areas”).
| Way | What is it | Cost (order of magnitude) | Time | For whom |
|---|---|---|---|---|
| Platform (e.g., Azure AI Studio, AWS Bedrock) | Configuring a ready-made service for your data | 1,000–5,000 EUR start + 100–500 EUR/month. | 1–4 wks. | IT company |
| Custom system | Application aligned with your processes and security policy | 15,000–60,000+ EUR + 500–2,000 EUR/month. | 6–16 weeks. | Confidential data, multiple roles, audit |
Pros: „from your documents” answers, footnotes, knowledge scaling with archives.
Cons: Cost and time, quality = quality of input documents, errors still possible – verification remains.
For whom: larger teams, many contracts/procedures, regulated industries.
Path 3: AI Agents (Multi-step Automation)
What it consists of: one command runs sequence - searching for contracts, calculating interest, selecting a template, generating a DOCX file, etc.
Difference vs RAG: RAG mainly responds. Agent executes a chain of tasks On your behalf (within the limits you define).
Pros: Time savings on repetitive scenarios.
Cons: highest cost and complexity, an agent error can have greater consequences than a chat error itself – it must be clearly defined, What is he not allowed to do.
For whom: usually step after RAG, not the first day with AI.
Here the subject intersects process At the company: if sales or customer service still rely on „manually pushing” tasks, the agent alone can't fix much. It's worth looking at it like a B2B sales process bottlenecks – First, order in the workflow, then automation.
Comparison Table – Three Roads Side-by-Side
| Feature | Ready-made tools | RAG | AI Agency |
|---|---|---|---|
| Knows company documents | No (only what you paste) | Tak (repository) | Yes |
| Indicates sources | No | Yes | Yes |
| Performs multi-step actions | No | No | Yes |
| Access control | Limited | Yes (role, areas) | Yes |
| Data security | „Cloud standard” | To configure | To configure |
| Startup cost | 0 EUR (excluding licenses) | 1,000–60,000+ EUR | Ok. 30,000–80,000+ EUR |
| Monthly cost per person (approximate) | $20–30 | 10–70 EUR* | 20–100 EUR |
| Time to launch | 1 day | 1–16 weeks | 3–6 months |
| Complexity | Low | Average | Tall |
It depends on the number of users and the volume of requests.
Which path to choose – without planning three years in advance
Most organizations sensibly go like this:
Stage 1: Ready-made tools (from now on)
- Licenses for the team (ChatGPT Teams or Copilot – depending on the Office/Google ecosystem).
- Observation: What do people actually use AI for? What are they missing?
- Cost: Government 20–30 EUR/person/month. Start: one day.
This is exactly the logic small, verifiable step before you make a big purchase – similar to the approach MVP in IT projects: First you test the hypothesis at a low cost, then you scale.
Stage 2: RAG (when the signal appears)
It's indicated by a sentence in the style of: „Cool, but it's a shame it doesn't search all our contracts.”
Then: one department, one document area, Pilot program, then full implementation.
Stage 3: Agents (optional, later)
When you hear: „I want to get ready-made documents with numbers from our systems with one click.” – that's already a topic for agents, often on the foundation of RAG.
You don't have to plan all the stages in advance. Start with Stage 1 – you will see real needs within weeks, not years.
How much does this cost? Estimate for 12 months (company of ~10 people)
| Solution | What do you get | Year 1 (Respect) | 2+ year |
|---|---|---|---|
| ChatGPT Teams | Chat, work on individual documents | ~3,000 EUR | ~3,000 EUR |
| Copilot M365 | AI in Microsoft Office | ~3,500 EUR | ~3,500 EUR |
| RAG (platform) | Document index, footnotes | 5,000–15,000 EUR | 2,000–8,000 EUR |
| RAG (custom) | As above + roles, audit, full control | 25,000–80,000+ EUR | 6,000–25,000 EUR |
| RAG + agent | Document generation in process | 40,000–120,000+ EUR | 8,000–30,000 EUR |
Sheridan-style question: how many hours per week does the team wasting time searching for answers in old emails and files? Multiply by the hourly labor cost. If it's more than the tool maintenance cost, you have a business argument, not an „AI trend.”.
Ten Questions for an AI Implementation Provider (to Differentiate Marketing from Substance)
If you are having conversations with implementation partners, ask them directly:
- Where is my data physically processed? Good answer: specific region, e.g. EU / Frankfurt – not „somewhere in the cloud”.
- Will the data be used for model training? Good answer: No, plus the agreement/DPA.
- What if the model provides incorrect information? Good answer: sources, logs, verification procedure – not „it doesn't happen.”
- What does access control look like? (Role, areas, possibly audit.)
- What model do you use, and why that particular one?
- Will you do a PoC on our documents? (Under a closed NDA.)
- What does the output from you look like – data export, formats?
- What is the fixed cost after implementation – a number, not just „it depends”?
- Who maintains the system after launch - SLA, contact, escalations?
- Can we start with a small scope and expand? (Yes = healthy approach; „only the full package right away” = red flag.)
Summary - what to choose if you don't want to get lost
| If... | Consider... |
|---|---|
| Want to start quickly and cheaply | ChatGPT Teams or Copilot (€20–30/user/month) |
| Do you want the model to „know your” agreements and procedures | RAG – dedicated platform or solution |
| You have high requirements for confidentiality and auditing. | Custom RAG + policy-compliant hosting (e.g., EU) |
| Do you want to automatically assemble documents in the process | RAG + agent – usually a later stage |
| You don't know what you need yet | Phase 1 + observation for 4–8 weeks |
Technology is mature enough that The question is not „if”, only „Where to start to avoid overpaying and exposing the company to incorrect answers without control”.
And what about the end customer who wants the answer themselves?
Some questions – pricing, configuration, option availability – He doesn't have to go to the salesman.. Well designed online configurators and presentations or a coherent catalog with offer logic do the same as a good article: respond early, before someone goes to the competition.
Would you like to find a path that fits your company? Write or schedule a call – we'll go through your situation without technical jargon and without imposing ready-made solutions.
