India Is Using AI. So Why Is Everyone Still Chasing Updates?

A joint study by the Confederation of Indian Industry (CII) and the Indian School of Business (ISB), reported in October 2026, found that 65.8% of surveyed micro, small and medium enterprises were using AI tools. Only 21% reported using enterprise resource planning (ERP) or customised business software.

The study covered 2,919 confirmed businesses. Participation was voluntary, so the findings cannot be generalised to every Indian MSME. The figures also measure different kinds of technology use: occasionally using an AI assistant is not equivalent to implementing software across business operations.

Even with those qualifications, the contrast raises an interesting question.

Why might an employee find it easier to start using advanced AI than a business finds it to connect its sales, production, accounts and dispatch information?

The answer lies in what each change demands.

An employee can begin using ChatGPT without asking anyone's permission to reorganise the company. Within minutes, they can write better emails, prepare presentations or summarise documents.

The benefits are immediate. Existing habits can remain largely unchanged.

Connecting business operations is different.

Someone must decide how orders are recorded, when production milestones are updated, which department confirms dispatch and who is responsible when information is missing.

These decisions affect several people. They require agreement, consistency and sometimes uncomfortable changes in responsibility.

AI gives individuals new capabilities. Better processes require organisations to develop shared capabilities.

One is relatively easy to adopt. The other is harder to negotiate.

That difference matters more than the sophistication of the software involved.

Consider what happens when a customer pays an outstanding invoice by bank transfer.

The money reaches the company's account. But accounts hasn't matched the payment against the invoice. Two days later, a sales executive sends the customer a reminder.

The customer replies, annoyed, with proof of payment.

Accounts checks the bank statement. Sales apologises. Someone updates the records.

Nobody has necessarily done their job badly.

The bank processed the payment. Accounts followed its reconciliation routine. Sales acted on the information available.

Yet several people have spent time resolving a problem that should never have reached the customer.

This is an important category of work, because businesses rarely measure it separately.

It appears as follow-up calls, clarification messages, repeated data entry, status meetings, reconciliations and corrections.

Every department sees its own tasks. Few see the cumulative cost of information failing to move properly between them.

A production manager thinks he is answering a sales enquiry. A sales executive thinks she is following up on an order. An accountant thinks he is resolving a payment discrepancy.

From the perspective of the business, they may all be doing something else: compensating for gaps in a shared process.

And because these activities are spread across departments, they can become accepted as an unavoidable part of running a company.

They aren't always unavoidable.

Sometimes the work exists only because the organisation doesn't know what it already ought to know.

Now introduce an AI assistant into the payment example.

It helps the sales executive write the reminder in ten seconds instead of two minutes.

Measured at the level of the employee, productivity has improved.

But the reminder was unnecessary. The customer was still inconvenienced. Accounts still had to investigate. Sales still had to apologise.

The organisation has become faster at performing a task that shouldn't have existed.

This distinction is easy to overlook because most productivity discussions begin with the task.

How quickly can an AI tool prepare a quotation? Summarise a document? Answer an enquiry? Generate a follow-up?

Those are reasonable questions, but they leave a more fundamental one unanswered.

Why does this task need to happen at all?

An employee spending an hour collecting order updates might use AI to reduce that work to twenty minutes. That's a meaningful saving.

But if reliable production information were available when needed, much of the collection process could disappear.

The difference is between reducing the cost of an activity and removing the condition that creates it.

Both can improve a business. They are not equally transformative.

There is also a less obvious risk.

When AI makes scattered information easier to search, summarise and reconcile, it can make fragmented operations more tolerable. Employees experience less friction, even though the underlying causes remain.

A company may consequently feel less urgency to improve how information is recorded and shared.

In that situation, AI isn't creating the inefficiency.

It's making the inefficiency easier to live with.

This isn't merely a theoretical distinction.

A 2025 study published in Scientific Reports examined an automotive tooling manufacturer, AB Dies and Moulds, in Chennai. The company introduced a combination of Lean manufacturing practices, standardised operating procedures, employee training and connected ERP modules.

Among the reported improvements, the time required to trace and resolve quality-related information fell from 3.5 days to 0.8 days. Production scheduling accuracy also improved, from 68% to 93%.

The study involved a single company, and the gains cannot be attributed to software alone or assumed to apply across Indian manufacturing.

But the case illustrates something important about how improvement occurs.

The business didn't simply acquire a more intelligent interface. It changed how work was carried out, documented, checked and coordinated.

Technology made those changes easier to sustain.

The implementation was not frictionless, either. Employees required training, existing practices had to change, and manual and digital processes operated in parallel during the transition.

That is a more realistic picture of digital transformation than installing a tool and expecting productivity to appear.

It also explains why this kind of progress is harder than giving employees an AI assistant.

A shared system exposes inconsistencies. It requires people to record decisions they previously communicated informally. It can reveal that different departments have been using different definitions of the same status.

Those problems existed before the software arrived.

The software merely makes them harder to ignore.

None of this means every small business needs an expensive ERP installation. A modest system connecting order records, payment confirmation and dispatch updates may deliver substantial value.

The objective isn't to replace every spreadsheet.

It's to ensure that critical information has an agreed meaning, a responsible owner and a dependable way of reaching the people who need it.

This becomes especially important as AI moves beyond drafting text and begins taking actions.

Imagine an AI agent answering customers' questions about their orders.

It can retrieve spreadsheets, read WhatsApp messages and consult dispatch records.

But suppose those sources disagree.

The production spreadsheet says an order is ready. Dispatch has no record of collecting it. A supervisor mentioned a delay in a message yesterday.

Which version is correct?

The agent might interpret all three sources perfectly. That doesn't tell it which one reflects what actually happened.

It can identify the contradiction and ask for confirmation. What it shouldn't do is confidently tell the customer the goods are on their way.

This is the difference between having access to information and having a reliable basis for a decision.

For an operational system, knowing who last confirmed a status, when it was confirmed and what remains uncertain can be more valuable than producing a fluent explanation of incomplete records.

And not every problem requires artificial intelligence.

If a payment has been matched to an invoice, a conventional rule can stop an unnecessary reminder.

If dispatch has confirmed that goods left the factory, a connected system can make that status available to sales.

There is little reason to involve a language model merely to retrieve a well-defined fact.

AI becomes more useful when the information is difficult to structure or the situation requires interpretation.

It could extract order details from unfamiliar purchase orders, understand messages mixing Hindi and English, recognise discrepancies across documents or help employees investigate exceptions.

It may even help smaller businesses connect older software and messy records without replacing every existing system.

So this is not an argument for choosing traditional software over AI.

It's an argument for giving each technology the right responsibility.

Use dependable systems to establish what is known. Use AI to interpret complexity, handle exceptions and make that knowledge easier to act upon.

Without the first, the second can produce remarkably polished confusion.

Return to the Ludhiana manufacturer.

Suppose the company wants every customer to receive an accurate order update without repeated phone calls.

The technical requirements might be modest.

Each order needs a unique identifier. Production milestones need agreed definitions. Outside-processing delays need to be recorded. Dispatch must confirm when goods actually leave. Sales needs access to the latest confirmed status.

But those apparently simple requirements raise organisational questions.

Who records a delay at an outside vendor? Who updates an expected dispatch date? Can sales promise delivery before production confirms it? What happens when a supervisor makes an exception?

These are decisions about accountability, not algorithms.

An AI assistant can help implement the resulting workflow. It cannot legitimately decide, on its own, how responsibility should be distributed across the business.

Nor will any system eliminate uncertainty entirely.

Suppliers will be late. Employees will forget updates. Customers will change requirements. Physical inventory will sometimes disagree with the records.

The aim isn't to create a fantasy in which every piece of information is always perfect.

It's to build an organisation that knows when information is uncertain, who is responsible for resolving it and what should happen next.

That is what makes a system dependable.

And it is why even sophisticated software can fail when the habits surrounding it remain unchanged.

For India's smaller businesses, the next stage of AI adoption should be judged less by how many employees have access to intelligent tools and more by what changes in everyday operations.

A useful starting point is to examine the work that connects departments.

How many calls are needed to establish an order's status? How long after receiving a payment does the business recognise that an invoice has been settled? How often do employees discover contradictory records? How frequently does a customer have to ask twice?

These questions expose the cost of poor coordination.

They also reveal opportunities that may not require AI at all.

A business might start by connecting one critical workflow, assigning responsibility for its records and measuring what changes.

Once the information is reliable, AI can help employees use it more effectively: answering customer enquiries, spotting unusual delays, translating communications and handling cases that require judgement.

The combined improvement is more valuable than either technology in isolation.

And the results become visible in outcomes that matter: fewer unnecessary calls, fewer incorrect reminders, more dependable delivery commitments and less dependence on individual employees remembering where information is kept.

Technology adoption is easy to count. Operational capability is harder to measure.

But the distinction is worth making.

A manufacturer doesn't particularly care whether a customer update was generated by an advanced reasoning model or retrieved through a simple database query.

He cares that it is correct, that the customer receives it and that his production supervisor can get on with running the factory.

A business that reduces the time needed to chase an update from ten minutes to one has improved productivity.

A business that makes the chase unnecessary has achieved something more fundamental.

The real test of AI adoption isn't how much work a machine can perform. It's how much unnecessary work the business no longer has to do.