For every five people you employ, one spends the whole day hunting for information. Lucky Core Industries and Servis Group stopped paying that tax. This is how they did it.
Introduction
Ask a senior executive at almost any industrial group how long it takes to see consolidated sales against budget. Most of them describe the same routine: email three departments, wait two days, reconcile whatever comes back, then make the call on instinct anyway.
These companies have spent years and serious money on ERP, CRM, historians, SCADA and BI platforms. The data exists. What nobody has done is close the distance between where that data sits and the person who needs it.
Episode 3 of the Octopus Digital webinar series tackled that gap directly. Leaders from Lucky Core Industries and Servis Group described what changed after they deployed OmniChat, and what they would tell any company about to start. This article covers what they said, along with the architecture that turns “your data never leaves your network” into a technical fact rather than a marketing line.
Watch the full session: Instant answers, zero data leakage.
Instant answers, zero data leakage: How industry leaders deploy GenAI that speaks their language
How Lucky Core Industries and Servis Industries Limited deployed OmniChat GenAI to go from 48-hour manual reporting to instant, on-premise answers across their enterprise data, without a single record leaving their infrastructure.
Key takeaways
- Companies now collect data faster than they convert it into decisions, and that conversion speed decides who wins.
- Public LLMs know nothing about your fiscal calendar, your plant codes, or the way your business defines contribution margin, so they answer generically.
- OmniChat sends the LLM metadata only, meaning table names, column names and descriptions. The generated query executes locally and your raw data stays inside your environment.
- Both companies spent more time cleansing data than deploying AI. LCI put the cleansing effort at three to four weeks.
- Both teams began with one business problem and one user group, which in each case was leadership.
Why reporting still takes days in companies that already digitized
Tanveer Karamat, CEO of Octopus Digital, opened by describing how the problem has shifted. Heavy investment across OT and IT, increasingly converging, means these companies produce more data than at any point in their history. Collecting data used to be the hard part. Karamat argued that the difficulty has moved to converting that data into decisions quickly, accurately and at scale, and he stressed that companies need all three. Pilots that never scale fail to return the money that produced them.
His framing of competitive advantage drew a strong reaction in the chat. Companies will compete on how fast they learn from their data relative to everyone else, rather than on how much of it they have accumulated.
He also raised an effect that gets less airtime. Generative AI raises the intellectual floor of an organization, because people who lack seniority and deep domain knowledge can now reach analysis that used to sit with a handful of experienced hands. Karamat called this genuine democratization of institutional knowledge, and he expects it to lift the whole company to a different operating level.
What scattered information costs
McKinsey estimates that employees spend roughly 1.8 hours a day, or about 9.3 hours a week, searching for and gathering information. IDC puts knowledge worker search time at around 30% of the workday. Do the arithmetic and the picture gets blunt: for every five people you employ, one effectively spends the entire day hunting for information instead of acting on it.
The upside carries equally concrete numbers. McKinsey analyzed 63 generative AI use cases across 16 business functions and estimated $2.6 trillion to $4.4 trillion in annual value, concentrated in customer operations, marketing and sales, software engineering, and R&D. Adoption climbed quickly too, from 65% of organizations using generative AI regularly in at least one function in early 2024 to 71% by 2025.
None of those numbers make AI inevitable. They do show that information friction has always carried a cost, even though no P&L ever recorded it as a line item.
The security gap traditional DLP was never built to catch
LayerX telemetry collected insights from enterprise browsers and found that 77% of employees paste data into GenAI prompts, with 82% of that activity running through unmanaged personal accounts. Around 40% of the files employees upload to GenAI tools contain PII or payment card data. IBM’s 2025 Cost of a Data Breach report attributes 20% of breaches to shadow AI, which added roughly $670,000 to the average cost of a breach.
Security teams should pay attention to why their existing controls miss this. Vendors designed conventional DLP around email, file transfer and sanctioned SaaS applications. Nobody built it to inspect what an employee pastes into a browser prompt. The leak vector moved and the controls stayed where they were.
Even if you set security aside, public models create two business problems. They cannot see your plant, your KPIs or your SAP orders, so they answer in generalities. And if your fiscal year starts in July, they have no way of knowing that unless someone tells them every single time. Traditional BI solves part of the problem, though only for people who already know how to drive it. Ask a commercial director to build an ad hoc cross functional view in Power BI and you are back to raising a ticket.
Khurram, product lead for OmniChat, walked through the architecture. OmniChat gives users a ChatGPT style interface for enterprise data without sharing that data with any LLM. Users ask questions in natural language, the system translates them, executes the result locally, and returns text, tables, charts or complete dashboards.
The six steps:
- A user asks a question in natural language.
- The application packages that question with metadata only, meaning table names, column names and descriptions. It sends no records and no values.
- The metadata and the question travel to the LLM.
- The LLM returns code, either a Python script or a SQL query. It never sees your data.
- The code goes to the OmniChat runtime, which sits on a virtual machine inside your network with read only access to your database, and executes there against your actual data.
- The result returns to the user and displays.
After step four, nothing reaches the LLM. A secure VPN tunnel connects the two networks, and unplugging that connection cuts the model’s visibility immediately.
Khurram explained the principle with an analogy that stuck. Imagine you need to calculate OPEX from a bank statement in Excel and you don’t know how. You could send the file to your accountant friend, which hands over your data. Or you could call and say there are ten columns, here is what they are named, tell me how to do this. You get the same answer and the data never leaves your desk. OmniChat works the second way.
Two transparency details surfaced in the Q&A. Users can inspect exactly which metadata went to the LLM, and they can review the code that came back before the server runs it.
Customers choose from four deployment models, each with role based access control so you decide who sees what:
- On premises, air gapped. Everything stays inside your network, including an open source LLM. This gives you the strongest security at the highest cost, since you buy the GPUs. Banks, insurers and healthcare organizations usually pick it.
- Hybrid. Data stays on premises while the LLM sits outside. Both LCI and Servis run this model, and Octopus Digital recommends it as the default for most use cases.
- Private cloud. Octopus Digital deploys onto your own cloud infrastructure.
- Public cloud. The most economical option, with some security tradeoffs.
That right-hand column is the whole point, so let’s get into how it works.
Lucky Core Industries: when a dashboard shows the data but not the story
Zain A. Khawaja, CTO at Lucky Core Industries, spoke candidly about where good dashboards stop being enough. LCI, formerly ICI Pakistan, operates across life sciences, chemicals, polyester and soda ash, and the company had already built dashboards that earned national recognition. They still fell short.
He gave an example every commercial leader recognizes. A product slips below its usual sales pattern for a month and then drops to zero. The dashboard reports the drop without explaining it. Demand may have collapsed. Demand may have held while stock ran out. A competitor may have arrived at half the price with better credit terms. Somebody may have discontinued the product deliberately because it no longer fit the portfolio. Each explanation points to a different strategy, and as Khawaja put it, those decisions sit 180 degrees apart. A newly appointed chief commercial officer looking at that chart cannot choose between them, because the context lives in people’s heads and few organizations still keep twenty year veterans in every seat.
LCI therefore pulled ERP data, CRM and sales records, visit logs, complaints, stock levels and the Excel files holding budgets and pricing into a single model. They trained it on their own semantics, covering what a plant means to them, how they calculate contribution margin and what counts as a discount, then pointed it at their C-suite.
Three things stand out from their approach.
They chose executives as the first users on purpose. They went after the people who make strategic decisions and have no time to pull reports, rather than the ERP power users who already knew their way around the systems.
Cleansing consumed the most time, three to four weeks of it. Khawaja put the reasoning plainly: you can trust AI insights only as far as you trust the data underneath, and companies tend to keep isolated data badly because nobody ever expected it to tell a story.
The feedback loop built trust. Every response carries a thumbs up and thumbs down control, and a thumbs down routes to a designated trainer, either inside LCI or at Octopus Digital. Most of those cases reveal a missing business rule or a gap in context. Khawaja’s reasoning here is sharp. When people dislike an application they rarely complain, they quietly stop using it, and no one can mandate usage at leadership level. Resolving issues fast kept executives engaged.
LCI also built a governance layer that isolates exactly which datasets the LLM can reach. Even with the architectural assurances in place, they exposed only the data they were willing to experiment with, a posture other companies could usefully copy. They went live in late 2025 and are now scoping OT data as the next step.
Servis Group: five spreadsheets, three teams, one question
Bilal, Chief Commercial Officer at Servis Group, came at the problem from a different angle. His group manufactures footwear, tyres and automotive parts, exports to several markets, and runs a retail network, so each business carries different data richness and moves at a different tempo.
His inbound supply chain example makes the cost concrete. The banking team keeps trade finance data across a digital platform and an Excel sheet. A second team moves material from port to factory. A third handles clearing and document retirement against a bank deadline of noon, while logistics operates between 7pm and 9pm. Ask for a consolidated position at 1pm and none of those timelines line up.
Management used to WhatsApp one or two key people, who would open a few sheets, chase a few colleagues, and produce an answer three hours later. By then the moment had passed. You would still make the decision, just with worse information.
Servis moved their static dashboards into OmniChat and added natural language querying on top. A manager now asks how many letters of credit are pending, how many documents are clean versus discrepant, and what foreign exchange exposure sits on contracts already entered. The answer arrives in about ten seconds, with no middleman.
Two of Bilal’s observations deserve more attention.
Why an off the shelf tool would not work. Frontier models keep converging in capability, so eventually the data each one can reach will separate them, which means industries will differentiate themselves on how well they structure and protect their own data. Put your pricing sheet into a public assistant and you may never query it back by name, though the exposure is real. With OmniChat, the data stays inside his system and the model only understands the file structure.
Probabilistic matching. Write “black” in one row and “blck” in another, and a deterministic Excel or ERP query misses the typo. AI catches it, because it works probabilistically. Ask for everything steel related and it picks up variants, specifications, hyphens, underscores and spacing differences. That capability alone breaks a familiar cycle in which a manager reviews a report, says something looks off, and an analyst hunts down a spelling variant and rebuilds the whole thing.
Bilal made one further point, about the act of asking itself. Under the old model, when a logistics lead wanted data from a commercial manager, the first response was always “why do you want this?” Sometimes people were protecting territory, sometimes they were protecting their own workload. Direct access removes that negotiation entirely.
What the speakers recommend if you are starting now
- Start with a business problem, not an AI initiative. Karamat urged companies to avoid running generative AI as a technology project and to look for use cases that deliver a quick, visible win. Khawaja’s team did exactly that and credited the approach for the project’s success.
- Budget properly for data cleansing. LCI needed three to four weeks for its dataset, and Khawaja warned that timelines climb steeply once you reach back ten, fifteen or twenty years, because definitions shifted over time and different people kept data according to their own preferences. Karamat put it bluntly: why invest in analyzing data when nobody has cleansed the source?
- Name your data owners. Master data governance matters, and defining who owns which dataset is what makes AI viable at scale. Bilal’s version: anything left to its own accord drifts, both in quality and in currency. The X factor is that anyone can check the data at any time, and that visibility changes behavior.
- Train the model on your business semantics. Teach it contribution margin, gross margin, what a plant is and what counts as a discount. Skip that step and you get technically correct answers to questions you never asked.
- Build the feedback loop before you launch. Set up thumbs up and thumbs down controls, name a trainer, and commit to fast turnaround. Tell users at the outset that accuracy will fall short of 100% and that vigilance forms part of the deal. Visible correction builds trust, while claiming perfection destroys it.
- Keep humans accountable. Karamat’s third leadership principle holds that AI augments judgement without replacing it, and accountability stays with the user. What changes is the speed of the analysis feeding the decision.
LCI added a seventh lesson. Once leadership started asking sharper questions, the team discovered that their CRM data lacked the richness to answer them. So they ran training programmes, used OmniChat itself to score how well salespeople completed their reports, and built rewards around the results. The system flagged where the narrative failed to match the numbers and where people were copying and pasting. Data quality turned into a managed performance metric instead of a hope.
From the audience
Which LLMs does OmniChat use? The platform stays LLM agnostic and works with closed source models, open source models, or Azure OpenAI where a customer prefers it. On premises deployments typically run open source models, and where customers express no preference, OmniChat routes each question to an appropriate model.
Can it connect to SAP and Oracle? Yes. More than 500 connectors cover the major business applications, databases, CSV and Excel files, and OT sources including devices and sensors. The platform also supports MCP, which extends it into the wider ecosystem of open source tools.
Did anyone independently verify the security architecture? During the LCI deployment, their security team tested the setup and confirmed compliance.