Eighteen Systems, One Command Centre: How Parakkat Group Got a Single Source of Truth

What does a founder command centre actually do?

It puts every number a business runs on into one live view, and lets you ask questions of it in plain English. Instead of exporting from eight systems and reconciling them in a spreadsheet, you open one screen, or message an agent, and get the answer as it stands right now.

For Parakkat Group, that meant roughly eighteen applications - Shopify, Odoo, Google and Meta Ads, GA4, Search Console, Klaviyo, Shiprocket, TeleCRM and others - resolved into a single source of truth on infrastructure the group owns outright.

Why did eighteen systems need unifying in the first place?

Because the data existed but the answers did not. Every system had its own login, its own export button, and its own definition of a number. A question like which branches are below reorder level required a person to run a report, export it, and interpret it - by which time the situation had already moved.

The problem was never a shortage of dashboards. It was that no two systems agreed on what a number meant.

This is the ordinary condition of a group that grew by adding tools. It is also the condition most AI projects fail in. MIT's Project NANDA, reviewing more than 300 disclosed AI initiatives, found that 95% of generative AI pilots produced no measurable P&L impact - and concluded the cause was mishandled deployment rather than weak models. The same study found that buying from a specialist vendor succeeded roughly 67% of the time, against internal builds succeeding about a third as often.

How do you connect eighteen business systems without replacing them?

You put a connector in front of each one and normalise everything into a single shared schema, rather than mirroring each vendor's shape. Nothing gets ripped out. The systems keep running exactly as they did, and the command centre reads from them into one canonical store underneath.

We built it in four layers so any one of them can be replaced without disturbing the rest:

• Integration - one connector per source, normalised into a canonical schema.
• Context - a Postgres system of record in the Mumbai region for data residency, with vector storage alongside for semantic recall.
• Agent runtime - a self-hosted, open-source agent, governed and approval-gated, on the client's own server.
• Interface - a web command centre with dashboards, chat and an approval queue, plus the same agent on messaging.

The group's own AI provider key powers it. If they replaced us tomorrow, the system keeps running. That was a design requirement, not a courtesy.

What happened to the reports the team already relied on?

We reproduced all six of the group's ERP Excel reports as live data: SKU sales, category pivots, branch sales, stock limit and excess alerts, and production against sales. As spreadsheets they answered one fixed question on a schedule. As live data they answer any version of the question, for any date range, for any branch, on demand.

They surface two ways - as operational screens, and conversationally, so the agent can be asked directly which items in a given showroom have fallen below reorder. That is the difference between a report and an answer.

The first build was correct and too slow. Pages aggregating across 52 branches and several years of transactions stalled long enough that people stopped opening them, which makes a reporting system worthless however accurate it is. We moved the expensive aggregation to nightly rollups with a cache warmer, and pages now open in about a second.

Can an AI agent be trusted with live business systems?

Only if its limits are enforced outside itself. This agent does not just read - it can manage store content and adjust advertising. So reads are free and instant, while anything that changes the outside world drafts and waits for a human yes. The agent can be persuaded to attempt something; it cannot be persuaded past a permission it does not hold.

It also operates under a hard data rule: state no figure that did not come from a real query, and name the source. If a lookup fails, it says so and stops. Confident fabrication is the failure mode that makes most business AI unusable, and it has to be designed against explicitly rather than hoped away.

The same governance model runs our own agency, described in our AI operating system case study.

What changed for the business?

Questions that used to require a person and a next-day export now resolve on demand, from a browser or a phone. Stock and reorder positions across 52 branches, sales by category, and production against sales are visible as operational screens rather than assembled on request.

The command centre is live on the group's own domain and in daily use by the managing director, reachable from web, WhatsApp, Telegram and desktop - because a founder checking stock in a showroom is not sitting at a laptop. The engagement has since grown past reporting into active management, and the same foundation now carries a resort-side data feed.

If your numbers live in systems that do not talk to each other, that is reporting automation, and it starts with connecting AI to the systems you already run. Talk to us about your own build.