Services, 03 of 04
AI systems built to be operated, not just demonstrated.
Agentic applications, MCP and API integrations, and multi-step workflows. Built with the permissions, approvals, evaluation and logging a team needs to run them in production, and with every answer traceable to its source.
From $5,500
After a paid scoping phase.
01 Problem
The problem with most of these
AI pilots fail in a consistent pattern. The assistant sounds confident, gets something wrong in front of a customer or a regulator, and nobody trusts it again. Gartner reported in January 2026 that at least half of generative-AI projects were abandoned after proof of concept by the end of 2025. Scepticism is now the default position, and it is earned.
The failure is almost never the model. It is that the system was built to be demonstrated, not operated: no boundary on what it may touch, no record of what it did, and no way to tell an answer it can support from one it cannot.
02 Scope
What we build
Four kinds of system, usually combined. Each is scoped, priced and dated in writing before it starts.
Agentic applications
Custom AI applications and internal agents: systems that read your documents and data, reason over them, and act through tools you define, with every answer anchored to a source you can open.
MCP and API integrations
Models connected to your databases, documents, APIs and business applications through MCP servers and integrations: authenticated, authorised and scoped to what each tool may read or change, with explicit schemas and error handling.
Workflow engineering
Multi-step systems that combine agents, tools, structured outputs and human approvals, with the context architecture that keeps them reliable as the work grows.
Production controls
What a prototype skips: least-privilege access, approval gates before any change, structured validation, an evaluation suite run from your own questions, logging and observability, fallback behaviour, and a written record of what the system does not cover. We document the controls; we do not certify compliance with any regulation.
03 Method
How we build them
Retrieval first, generation second. Where a system answers from your records, it finds the passages before it writes, and every claim in the output is anchored to a source you can open and read. When retrieval comes back empty or thin, it reports the gap. It does not fill it.
Where a system acts, every action goes through a tool with a schema, a scope and an approval rule. The model proposes; the tool, its permissions and a person decide. Text the system reads, from a document or a web page, is given no authority. A model cannot be relied on to tell data from an instruction, so the limit sits outside it: nothing the system reads can widen what its tools are permitted to do.
We build mostly on Anthropic’s Claude, through the Claude API, Claude Code and MCP, and on other providers’ models where a project calls for them. An AI-assisted prototype is where a system starts, not where it ships: the work is in the architecture, the tests, the permissions and the handover that let your team run it without us.
04 Standard
What ships with every build
- Every answer the system gives linked to the record it came from.
- An evaluation set built from your own real questions, run before handover, with the results published rather than described.
- Where the system can call tools: a written map of what each tool may read and change, a person’s approval before any change is made, and a log of every call, kept in your account.
- A written record of what the system does not cover.
- Repository, hosting and API accounts in your name.
05 Price
From $5,500
Every engagement starts with a paid scoping phase, because the honest answer to "can AI do this" sometimes turns out to be no, and you should find that out for a small amount of money rather than a large one.
06 Fit
What we do not build
Customer-facing support chatbots. The evidence on those is poor, the failure is public when it comes, and we would rather turn down the work than sell you a rollback.
Tell us what it has to do, and what it must never do.
Kordal Systems is a design and engineering studio. We build websites, web applications and AI systems for e-commerce brands and B2B companies, and run software products of our own.
Kolkata, India · We reply within one business day.