// Service · AI in your product, Agents & MCP
AI in your product
Integration, Agents, MCP
Everyone says they're "now powered by AI." Few actually understand what that means.
In the past year, every software product started "offering AI." In most cases that means a ChatGPT wrapper bolted onto an existing UI. A chatbot that gives generic answers. A "summarize this" button that doesn't really understand the product. That's not AI in a product. That's set dressing.
Real AI in a product changes what the product can do. It understands context, accesses the right data, takes real actions, and serves the user in ways that weren't possible without it.
We build that. Not a wrapper — real integration.
// 01 · What we do
What we do
From integrating AI into existing products to autonomous agents and MCP Servers that connect AI to your world.
- 01
AI integration into existing products
We take a working product and add AI capabilities that actually improve it. Summaries, semantic search, task automation, content generation.
- 02
AI Agents
Not chatbots. Agents that execute multi-step tasks autonomously: send emails, analyze data, update systems, make decisions.
- 03
MCP Servers
Building MCP Servers that connect your AI to your systems. Google Drive, Salesforce, internal databases, enterprise systems. AI that acts in the real world — doesn't just answer.
- 04
RAG & knowledge bases
Connecting AI to your specific knowledge. Internal docs, product specs, contracts, data. AI that answers from your information — not Google's.
- 05
AI assistants in SaaS
A Copilot for your product. Guides users, answers questions, takes action mid-flow.
- 06
AI in the backend
Classification, text analysis, smart OCR, image analysis. AI that works behind the scenes.
- 07
Language & document processing
Contracts, invoices, medical records, emails. Extraction, classification, semantic search.
// 02 · MCP
What MCP is and why it changes the game
MCP — Model Context Protocol — is the standard Anthropic released in 2024, which became the industry de-facto standard in 2025.
The idea: AI is smart, but without access to your data and systems it's limited. MCP is the protocol that connects AI to your systems in a secure, structured way.
That turns it from "a chatbot that gives advice" into "an assistant that gets things done." We build dedicated MCP Servers that connect AI to your specific systems.
Done right, MCP lets the AI:
- 01
Read documents from Google Drive
- 02
Update records in your CRM
- 03
Send emails
- 04
Run database queries
- 05
Move tasks in Jira or Asana
// 03 · Technology
Our stack
Models, frameworks, vector DBs, MCP, embeddings, orchestration and fine-tuning. We pick the right tool for every integration.
- AI Models
- Claude (Anthropic)
- GPT (OpenAI)
- Gemini (Google)
- Open Models
- Frameworks
- LangChain
- LangGraph
- LlamaIndex
- Vector DBs
- Pinecone
- Weaviate
- Qdrant
- pgvector
- MCP
- Anthropic MCP SDK
- Custom MCP Servers
- Embeddings
- OpenAI
- Cohere
- Voyage AI
- Orchestration
- Temporal
- n8n
- On-Device AI
- Core ML
- Ollama
- Fine-tuning
- OpenAI Fine-Tuning
- Hugging Face
// 04 · How we work
How we work
- 01STEP 1
Understand the business case
Not every problem needs AI. And not every feature gets better with AI. We start by understanding the problem — then decide whether AI is the right answer.
- 02STEP 2
Pick the right model
Claude, GPT, Gemini, an open model, or fine-tuning? Each has its strengths. We pick by need — not by hype.
- 03STEP 3
Design the architecture
How does AI connect to the product? When does it run? What context does it get? What actions is it allowed to take? Plan before build.
- 04STEP 4
Development & testing
We build the integration and test on real cases. An AI giving inaccurate answers is worse than no AI at all.
- 05STEP 5
Cost optimization
AI API calls cost money. We optimize: caching common answers, smaller models where they fit, efficient prompts.
- 06STEP 6
Monitor & improve
AI isn't perfect. We monitor outputs, gather feedback, and improve prompts and architecture over time.
// A pilot before commitment
AI in your product, with a clear test process
Before AI goes into a product you need to know it performs the task at the level you need — on your data, not in a demo. This is what a pilot you can decide on looks like, and what it takes to turn it into a stable product.
Is AI right for the need? Four questions before a pilot
- 01
Which action are we improving?
One defined action: answering questions from existing knowledge, helping a user fill in or phrase something, processing documents, or a defined decision with clear rules. Not "adding AI to the product".
- 02
What quality level is required?
Is an error an inconvenience (a suggestion the user rejects) or harm (an action that was taken)? The higher the price of an error, the more human oversight and approval before action are needed.
- 03
What data exists?
Do the information sources the AI should rely on — documents, a knowledge base, request history — exist, stay current and may be used? Without them there is nothing to test.
- 04
What is the alternative without AI?
An "if-then" rule, better search, a structured form. If the alternative solves the problem — it is cheaper, more predictable, and we will say so.
What the pilot includes
- 1
Defined information sources
A closed list of documents, tables or systems the AI may rely on, with access permissions matching what the user it serves is allowed to see.
- 2
An example user and a trial group
Who uses the pilot, in what context, and at what volume — usually a small group that knows the domain and can judge answers.
- 3
Success conditions set in advance
Quality (how many answers in the scenario set a person would approve), time (versus the manual process), adoption (do users come back) and cost per action. The target is set with the process owner, not with us.
How quality and failures are tested
| Test | What is done | What must be defined |
|---|---|---|
| A scenario set | A collection of real questions or cases (approved for use) with the answer or outcome a person would accept, including edge cases and questions the sources cannot answer | Who builds the set, who judges, and which part of it is held back for the final check |
| Behaviour when there is no answer | The system says it does not know, points to a source or a person — instead of inventing | The wording, when to escalate to a person, and how it is measured in the scenario set |
| Approval of sensitive actions | Every action that changes data, sends a message or charges money passes human approval in the pilot | The list of sensitive actions and who approves |
| Tracking and logging | Every question, retrieved source, answer and judgment is stored — so failure patterns are found, not single cases | What is stored, for how long, and who sees it |
A scenario set
What is done: A collection of real questions or cases (approved for use) with the answer or outcome a person would accept, including edge cases and questions the sources cannot answer
What must be defined: Who builds the set, who judges, and which part of it is held back for the final check
Behaviour when there is no answer
What is done: The system says it does not know, points to a source or a person — instead of inventing
What must be defined: The wording, when to escalate to a person, and how it is measured in the scenario set
Approval of sensitive actions
What is done: Every action that changes data, sends a message or charges money passes human approval in the pilot
What must be defined: The list of sensitive actions and who approves
Tracking and logging
What is done: Every question, retrieved source, answer and judgment is stored — so failure patterns are found, not single cases
What must be defined: What is stored, for how long, and who sees it
What production requires
- 01
Permissions
The AI sees only what the user it serves may see; access to sources goes through the same permissions as the rest of the system.
- 02
Monitoring
Answer quality, escalation rate to a person, response times and model-provider errors — with an alert when something drifts.
- 03
Usage budget
Every model call costs money. Set a ceiling, measure cost per action, and decide what happens when the ceiling is near.
- 04
Support and updates
Who handles a reported wrong answer, how information sources are updated, and what happens when the model provider changes a version.
Relevant experience — exactly what was built
At OneAI, a conversational AI platform, we worked for five years as technology partner — from language AI APIs, through chat agents, to a voice AI system running business conversations, including a conversation scenario builder. At JUSK, an employee training system, we built AI-assisted creation of training content and tests from existing company knowledge, and an AI chat that answers employees from the organisation's procedures and information. Exactly what was built in each case is described on the case page — and nothing beyond it.
// 05 · Who it's for
Who this is for
- 01
SaaS products
That want to add a Copilot, a smart assistant, or AI capabilities to existing features.
- 02
Service companies
Handling lots of repeat inquiries — AI can take care of most of them.
- 03
Organizations with internal docs
Knowledge bases, procedures, documents. AI can help employees navigate them.
- 04
Analytics products
That need to process text, surface insights, or summarize data.
- 05
Businesses that want to automate
Combining MCP and AI for real workflows — not just chat.
// 06 · Why us
Why us
- 01
We build — we don't just talk
Many companies talk AI. We work with it every day.
- 02
We track the market
Our CTO stays updated daily. New models, new MCP, new capabilities — we're on it.
- 03
We understand the whole stack
Good AI needs good infrastructure. Fast queries, caching, stable integrations.
- 04
Not in love with the tech
If AI isn't the right answer — we'll say so. We won't push a feature just because it's trendy.
// 07 · FAQ
Frequently asked questions
01.What's the difference between AI I can wire up myself (ChatGPT API) and a professional integration?
The ChatGPT API gives you a model. A professional integration builds everything around it: architecture, efficient prompts, caching, context management, integration with your data, cost monitoring. It's the difference between a toy and a product.
02.Which models do you use?
Claude, GPT, Gemini, and open models (Llama, Mistral) when they fit. We pick by task.
03.What's an Agent and how is it different from a chatbot?
A chatbot answers a question. An Agent acts. It can execute multi-step tasks: "Check what this customer ordered, send them a summary email, and open a ticket if there's an issue."
04.So what is MCP, really?
A protocol that connects AI to external systems securely. Instead of the AI knowing only what it was trained on — it can read documents, take actions, and update systems in real time.
05.How much does an AI integration cost?
Depends. A simple integration (chatbot with RAG) — $8K–$22K. An advanced in-product AI assistant — $22K–$85K. A complex AI agent with MCP — $40K–$140K.
06.There are ongoing costs too, right?
Yes. Every API call costs money. We optimize to bring it down (caching, efficient models, precise prompts), but there's still a running cost. Usually a small percentage of the value the product creates.
07.What about privacy? Does our data go to OpenAI?
Depends on the model and the configuration. OpenAI and Anthropic offer Enterprise tiers that don't use your data for training. And there are open models you can run on your own infrastructure. We help choose based on data sensitivity.
08.What is RAG?
Retrieval Augmented Generation. The AI doesn't rely only on what it knows — it searches your knowledge base in real time and answers from there. That's how it can address questions specific to your product.
09.We want AI that understands Hebrew — is that a problem?
Not in 2026. Claude, GPT, and Gemini work great in Hebrew — understanding, writing, and answering at a high level.
10.What about AI on mobile?
Two approaches: an external API (fast to build, needs internet) or on-device (maximum privacy, works offline). It depends on the case.
11.When is adding AI not worth it?
When the problem doesn't need it. If you can solve something with a simple "if-then" rule — AI is overkill. We'll say so straight.
12.Do you also do fine-tuning?
Yes. When there's a real need — a very specific domain, a precise response format, a unique style — fine-tuning can give a better result at a lower cost.
13.We want our own MCP Server to connect our internal tools. Do you build that?
Yes. It's one of the areas we work on most. We'll build an MCP Server that connects your specific tools to your AI.
14.What is the difference between an AI pilot and a stable product?
A pilot tests one question — does the AI perform the defined task at the level you set, on your scenarios — with a small user group, defined information sources and human oversight. A stable product adds what a pilot does not need: permissions for every user, monitoring, a usage budget, a process for updating the model and the content, and support. A successful pilot is a condition for continuing — not the product itself.
15.How do you define success for an AI pilot without promising accuracy percentages?
The target is set with the process owner before the pilot, against the existing alternative: how many answers in the scenario set must be at a level a person would approve, how much time it saves compared with the manual process, and what the usage cost per action is. We do not commit to an accuracy percentage before measuring it on your data — the pilot's result is the number, and the decision is made on it.
Last reviewed:
Related reading
Service
AI automation for business
The process side of the same capability: specifying and building automation for repetitive manual work, wired into the CRM, ERP and systems you already run.
Case study
Case study: OneAI — five years of AI, acquired by monday.com
A long-term technology partnership: from a language AI and API platform, through chat agents, to a voice AI system — and the 2026 acquisition by monday.com.
Case study
Case study: Mystorya — an AI companion that remembers your story
An end-to-end personal-guidance app: iOS, Android and web, an AI chat fluent in astrology, numerology, tarot and palmistry, a daily reading, a multi-layer memory system and an admin console.
Case study
Case study: Onit — a doctor's appointment in days, with AI and data
An end-to-end health-tech app: iOS and Android, an appointment matching engine, insurance policy analysis, claims and refunds, a conversational AI assistant and user-data analytics.
Comparison
Vibe Coding vs production engineering — when each one fits
MVPs from Cursor/Lovable/Bolt vs production-engineered code. What AI tools handle, where they break, and how to migrate between the two.
Case study
Case study: JUSK — an employee onboarding platform
An HR tech platform consolidating forms, digital signatures, training and assessments into one onboarding journey: gated access, reminders, AI and a full admin platform.
// 08 · Let's talk
Real AI, not set dressing.
A 30-minute intro call. No strings attached. We'll figure out together if AI is the right answer — and if so, how to do it right.
Straight. To the point. On time.