AI equaliser for mid-sized firms

The company has already experimented with AI and found several promising uses. Its CEO now faces a different question: should it keep buying AI as a service, or build a private AI capability of its own?

That is becoming increasingly practical.

Open-weight models such as Qwen, a family of large language models released by Chinese technology company Alibaba, can be downloaded and operated by the company itself. Unlike a proprietary model accessed as an external service, this one can run on infrastructure the enterprise controls. The company does not own the intellectual property in Qwen, but subject to its licence, it can operate a copy without sending every request to the model’s developer.

Think of it as installing an AI engine inside the company.

Employees would access a simple internal application, call it Company AI, using their normal corporate login. Behind it sits an open model, connected with appropriate permissions to the company’s existing information and systems.

Suppose a salesperson is preparing a proposal for furnishing a 2,000-room student hostel. The system can retrieve specifications from previous projects, approved product catalogues, current prices, and relevant design standards, then help draft the proposal. An engineer might upload an architect’s specification and ask it to identify matching products. Procurement could compare a bill of materials against supplier data. Separately, a project manager could ask which components for a hospital installation are delayed. Customer service could retrieve maintenance instructions for a chair installed three years ago.

The model supplies language and reasoning. The company’s own systems supply the facts.

The competitive advantage does not lie primarily in owning the model. Models will improve and can be replaced. The harder asset to reproduce is the company’s accumulated knowledge: completed projects, pricing histories, drawings, supplier performance, and the judgement embedded in past decisions.

There is, however, an unglamorous prerequisite: the facts need to be usable. Drawings live in one system, quotations in spreadsheets and email, specifications in inconsistently named PDFs, and valuable knowledge in the heads of experienced employees.

Before connecting AI to everything, the CEO must invest in data plumbing. File naming needs standardisation. Product master files need a reliable home. Old drawings and project records need to be searchable. Versions must be controlled. Legacy paper archives containing useful technical knowledge may need digitisation.

Retrieval-augmented generation, or RAG, allows the model to retrieve relevant information from company records before producing an answer. But retrieval cannot repair bad source data. If three versions of a drawing exist and nobody knows which is current, AI merely makes the confusion easier to access. In manufacturing, an obsolete specification can turn into wrong procurement, wasted material, or a costly installation error.

Next comes infrastructure. Running a capable open-weight model privately does not necessarily require a massive server farm. The company could operate dedicated GPUs on premise, use a private cloud environment, or ask a systems integrator to manage dedicated infrastructure. The choice depends on workload, cost, and data sensitivity.

The important point is control. The company should know where its data goes, who can access it, what is logged, and which systems the model can touch.

Security therefore cannot be bolted on later. Company AI should inherit the permissions employees already have. A salesperson should not suddenly gain access to supplier negotiations because an AI assistant can search them. Procurement should not see executive compensation. A project team working for one client may need to be prevented from retrieving another client’s confidential drawings. The principle is simple: employees see only what they are authorised to see.

The company should also resist the temptation to automate everything immediately. AI can draft a quotation, flag inconsistencies, or prepare a comparison. It should not silently approve a commercial bid, change an engineering specification, or issue a purchase order merely because it can. In consequential workflows, AI prepares; people remain accountable.

Change management may prove harder than model deployment. Factory supervisors, procurement officers, and experienced salespeople will judge Company AI on whether it helps with Tuesday morning’s problem, not on benchmark scores. The sensible approach is to begin with a few high-frequency tasks where results can be measured: tender drafting, product matching, invoice reconciliation, service-document retrieval, or checking bills of materials.

Measure time saved, errors avoided, and whether employees actually use it. Then expand.

This also changes the economics of enterprise AI. The CEO need not make a 10-year bet on Qwen or any other model. The architecture should allow the underlying model to be replaced as better ones appear. What the company is really building is the layer around the model: its data connectors, permissions, workflows, evaluation methods, and user interface.

That layer is much harder to buy off the shelf.

Building a private AI capability is therefore no longer the exclusive domain of technology giants. For a mid-sized Indian company, the task is less about training a spectacular new model and more about disciplined data housekeeping, sensible infrastructure, security, workflow redesign, and choosing problems worth solving.

The model may be open and replaceable. The company’s knowledge is neither. That is where the real value lies.

The author is a technology consultant and venture capitalist.

Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.