AI inside the systems your business runs on, with control and responsibility

SLAtech builds AI into a company's existing systems: CRM, ERP, portals and document archives. Permissions, data and regulatory requirements stay under control, and the SLAtech team is responsible for what we build.

In short

AI in an organization is not a chatbot on the website. Without architecture, data controls and permissions it creates operational, regulatory and security risk. So we design AI as part of the enterprise system: it works with your data, respects what each user is allowed to see, and answers with a reference to the source.

What we build

Search and answers over documents

Staff lose time searching archives, folders and old email threads. The system finds the answer in the company's documents and shows which document and section it came from.

Assistants inside CRM and ERP

Customers and staff keep asking the same questions about orders, statuses and rules. The assistant answers from CRM or ERP data, and when a question goes beyond its scope, it hands the conversation to a person, history included.

Document processing and automation

Invoices, contracts and forms are typed in by hand, with errors and delays. AI extracts the data from the document, checks it against the rules and passes it to the system, leaving disputed cases to a person.

AI agent access to your systems through MCP

The AI assistants your staff and customers use cannot see your systems. Through an MCP server they get the data and actions they need, only within the permissions granted and with a log of every access. MCP integration

How search over documents works (RAG)

RAG (Retrieval-Augmented Generation) answers from your documents, not from the model's general memory. It works in four steps:

  1. Indexing. Documents are split into meaningful passages, each gets a vector representation, and everything is stored in a search index.
  2. Retrieval. The user's question is also turned into a vector, and the system finds the passages closest in meaning. Only what this user is allowed to see is returned.
  3. Answer. The model receives the question and the passages found, and writes an answer based on them alone. If the documents do not contain the answer, the system says so.
  4. Source reference. The answer links to the document and section, so it can be checked against the original.

If data must not leave the organization, the system can be deployed on your servers or in a private cloud.

Control: accuracy, security, regulation

Accuracy. Before launch, we build a set of real questions with verified answers and run the system against it, then repeat the check after every change to the model, the instructions or the data. An answer that is not grounded in the documents counts as an error.

AI-specific threats:

  • Prompt injection: text in a document or an email that tries to rewrite the model's instructions. Instructions and data are kept separate, and actions with consequences need a person's confirmation.
  • Leakage between customers and departments: the index and the search know whose document it is, so a user never receives a passage from someone else's documents.
  • Log: every query and answer is recorded, so a disputed case can be investigated and an auditor can see what the system saw.
  • For agents already in production, there is a separate AI agent security audit.

Regulation. Systems that hold personal data in Israel fall under the Privacy Protection Law and Amendment 13. Storage, access and logs are designed with these requirements in mind: Amendment 13 for information systems. Rules for AI use, the split of responsibility and documentation (AI governance) are prepared together with you.

Audits of AI systems

If the system was already built with AI, or an agent is already connected to your data, start with a review.

Audit of systems built with AI

Security, data, load, cost and maintainability. You get a written report, a fix plan and a reasoned recommendation: stabilize or rebuild.

Security audit for AI-generated code

Secrets in browser code, database access rules, server-side authorization, hallucinated packages.

AI agent security audit

Permissions of the agent and its MCP connections, prompt injection, data leakage, irreversible actions, monitoring.

Technical due diligence of AI-built products

For investors and buyers: code ownership, licenses, data access, platform lock-in, AI cost per user.

Where AI helps and where it does not

It helps when:

  • there are many repetitive questions, and the answers already exist in documents or in a system;
  • data arrives in documents (invoices, contracts, forms) and the result can be checked against rules;
  • you need to search a large archive that nobody remembers in full.

It does not help when:

  • a mistake is costly and only becomes apparent afterwards, with no way to check the answer in advance;
  • the data needed does not exist, or lives only in people's heads;
  • a simple rule, form or report solves the task. Then AI adds cost and risk without adding value.

How we work

  1. Discovery. We look at the data and the processes: where AI will pay off and where it will not, whether there is enough data, and what state it is in.
  2. Prototype on your data (PoC). A working version on real documents, checked against a set of questions with verified answers, and a decision: continue, change the approach or stop. Usually two to four weeks.
  3. Production. We build it into your systems, set up permissions, logs and monitoring, and train your team.
  4. Support 24/7. We track answer quality against the same question set and the spend on models, handle incidents and add new capabilities.

Ownership and switching providers

  • The code, the document index and the model instructions belong to you. Servers, domains and access keys sit in your accounts from day one.
  • The model is chosen for each task: a cheaper one for simple requests, a stronger one for complex ones. If a provider changes its prices or terms, the model can be replaced without rewriting the system.
  • A team at SLAtech knows the system, not just one person. The documentation is written so that your own team can work with it too.

Ready-made product: SLAtech AI

If you need a ready assistant for your website rather than a project, there is SLAtech AI, a product of SLAtech LTD. It answers visitors in Hebrew, Russian and English from your website's content and links the source page. Product website · How we built it

Tools

Frequently asked questions

How is RAG different from fine-tuning a model?

RAG finds the answer in your documents at the moment of the question, so new documents work straight away and every answer has a source. Fine-tuning changes the model itself. It is rarely needed, for a particular style or format, and it does not suit facts that change often.

How long does a launch take?

A prototype on your data usually takes two to four weeks. The time to production depends on how many systems need to be integrated and is set after discovery.

What does it cost?

The cost depends on three things: the volume and state of the data, the number of integrations, and the security requirements. Running costs for the models are counted separately. We give a firm estimate after discovery, and you can get the order of magnitude right away in the project estimator.

Which models do you use?

Whichever suits the task in terms of quality, price and data requirements, often different models for different steps. There is no lock-in to one provider: the model can be replaced without rewriting the system.

How does AI work in Hebrew?

We test quality in Hebrew separately, using a set of questions in Hebrew. Hebrew has its own difficulties: prefixes, for example, can make word search miss matches. The Hebrew interface runs right to left.

Can AI be used with personal data under Israeli law?

Yes, if the system is designed for the Privacy Protection Law and Amendment 13: least-privilege access, logs, and storage and deletion in line with the rules. More on the Amendment 13 page.

Let's discuss your task

Tell us what data and systems you have, and what AI should do. We will tell you where it will pay off and where it will not.

Talk to us

Related: MCP integration, AI access to your website, Vibe-coded projects, SLAtech AI case study, MCP and privacy: what AI sees and what it doesn't, What MCP gives your business, Why now, not later: the MCP early-mover advantage, MCP: the next website your business needs

SLAtech LTD