AI integration: assistants that work with your data, inside clear limits
We connect large language models to your ERP, CRM, documents and internal tools. The assistant reads only the data it is allowed to, shows where each answer came from, asks before it changes a record, and logs every step.
A written scope and a fixed budget before any work is billed.
From AI consulting to software that runs
Most AI efforts start with a demo and stay there. The demo is a general chatbot: it cannot see the company’s data, nobody owns a wrong answer, and nobody measures what it costs. The projects that work start narrow: one team, one job, one result you can measure. The questions a sales team asks the ERP every day, the answers a support team digs out of a document archive, or a report someone builds by hand every month.
So we begin with the consulting questions: which job, which data, and what a wrong answer would cost. Then we build the assistant or agent as ordinary software connected to your systems. Which data reaches the model, which tools it may call, which actions need a person to approve, and how much a single request may cost are all part of the written scope.
The layer the assistant reads through, the connection to your ERP and CRM, is our API and ERP integration service; most AI projects need it first. If the AI feature will be part of a product you sell, see SaaS development too.
Sound familiar?
The demo impressed, then nobody used it
A chatbot was tried; without access to the company’s data it gave generic answers, and the team stopped opening it within a week.
Sending data to a model makes people nervous
Nobody can say where customer, pricing or staff data goes, where it is stored or who can see it, and legal and IT are right to object.
The same question is answered by hand every day
Order status, stock, a contract clause or a procedure: the answer is in your systems, but someone has to open them and read.
An agent was requested, its limits never drawn
There is talk of a tool that sends email or edits records, but nobody has said who approves which step.
What we build
Each one is defined by its own data, permissions and measure of success.
Assistants over your data
An assistant that queries your ERP, CRM or database with read-only access and cites the source of every answer, for sales, support or operations.
- ERP
- Read-only
- Citations
Document and knowledge search
Search by meaning across contracts, procedures and technical documents; every answer arrives with the document and section it quotes.
- Documents
- Semantic search
Agents that do work
Agents with a defined set of tools: draft quotes, enter orders, build reports. Every step that changes a record goes through a person.
- Tool calls
- Human approval
AI features in existing software
A narrow, measurable feature added to your product: summaries, classification, field extraction or suggestions.
- Classification
- Summaries
Evaluation and monitoring
A test set of real questions before launch; answer quality, latency and cost tracked after it.
- Test set
- Spending cap
What you keep at the end
Everything below is yours from the first commit, under your own accounts wherever possible. Nothing in it is licensed from us, and nothing can only be maintained by us.
your-company/
- docs/requirements.mdThe job, the data, the permissions, the approval steps and the fixed budget.
- docs/data-access.mdWhich data reaches the model, which fields are masked, and where it is processed.
- src/The assistant or agent with its tools and interface, in a repository you own.
- evals/The test set of real questions and each release’s score against it.
- docs/runbook.mdWhat to do when the model provider changes, the spending cap is hit or answer quality drops.
Four stages, and what
each one hands you
Every stage ends in something you can review and keep, including the stages that come before any code is written.
Discovery and requirements analysis
We map the real problem before any code is written: stakeholders, constraints, existing systems, and what success means in measurable terms.
- requirements.md
- scope + fixed budget
- risk register
Design and architecture
Interface flows and system architecture designed together, so what we specify is what we can build and scale.
- clickable prototype
- system diagram
- data model
Build, with a weekly demo
Short delivery cycles with visible progress. You review working software throughout, not a finished system revealed on launch day.
- staging deploy / cycle
- test suite + CI
- weekly demo
Launch, then operations
We deploy the system, monitor it, and hand over the runbook. We then continue to operate it under a monthly agreement you can end with 30 days’ notice.
- production infra
- runbook + handover docs
- monitoring + alerts
Terms we put in writing
before work begins
The figures we commit to on every engagement. Your project gets its own dated plan on top of them.
- Inquiry
- 1 business dayA reply from someone who has read the brief.
- Discovery
- 1 to 3 weeksEnds in a written scope and a fixed budget.
- Weekly review
- Every weekThe work as it stands, on a link you can open yourself.
- First release
- 8 to 16 weeksTypical time from the end of discovery to production.
- Operations
- 30 days’ noticeMonitoring and iteration under a monthly agreement.
Illustrative. Every project receives its own dated plan before work begins.
What does an AI project cost?
The build is priced as one fixed amount at the end of discovery. Model usage is a separate, variable line: you pay the provider from your own account according to the number of requests and the length of the text sent. The scope includes a monthly spending cap and an alert that watches it.
- 01How many systems and data sources the assistant connects to
- 02Whether it only reads or also changes records
- 03How large and how orderly the document archive is
- 04Personal data, masking and where data may be processed
- 05The size of the test set and the acceptance bar
- 06Who monitors it and handles model updates after launch
What we build it with
Tools with long support horizons and large talent pools, so the system stays maintainable by any capable team, not only ours.
TypeScript
Typed schemas for tool definitions and the model’s structured output, so a malformed answer is caught before it enters the system.
Node.js
Backend services for model calls, tools and queues.
PostgreSQL
Vector search over document chunks and a log of every request, in the same database.
Next.js
The interface your team uses and the admin screens behind it.
Questions buyers ask us
The ones that come up most in first conversations about this service.
Something not covered here?
Email us at hello@kodfabrika.com.
Related services
Built to last.
Started this week.
Tell us what you’re building. You’ll receive a written scope, a defined budget, and a delivery plan. Not a proposal deck.