AI Saved Ten Minutes. Did Your Business Actually Gain Anything?

An employee finishes a customer reply in two minutes instead of ten. The owner sees a productivity miracle. The supervisor sees a promise about delivery that the business cannot keep.

Both are looking at the same piece of work.

That is the awkward centre of the AI productivity debate. Generating an answer faster is useful only if the answer survives the rest of the business. A draft can save time at one desk and create rework at three others.

For Indian MSMEs considering AI, the useful question is where assistance improves a complete task, including checking and correction. Counting generated words is the easy part. Counting work that no longer needs doing takes more care.

In India, the policy conversation is current too. A 30 July 2026 Ministry of MSME reply describes government measures intended to support AI and digital adoption. A programme's stated aim is not evidence that an individual firm has already gained productivity. That gap between adoption and a useful result is the business problem this article examines.

The research is encouraging—and specific

The revised November 2024 version of Generative AI at Work, by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, examines 5,172 customer-support agents. AI assistance increased issues resolved per hour by 15% on average, with effects varying substantially by worker. Less experienced workers benefited more; the most experienced saw small speed gains and small quality declines.

This is evidence from an actual workplace, which makes it valuable. It is still evidence about a particular support operation and tool. It does not establish that every Indian firm can reduce staffing by 15%, or that an open-ended chatbot will produce the same result without suitable knowledge and supervision.

Earlier versions of this research report different sample sizes and estimates. A responsible business case should cite the version it actually uses rather than assemble the largest numbers from different summaries.

The implication we draw is that AI can help distribute useful knowledge. Whether it does so in your business depends on what the system knows, what the employee is trying to accomplish and who catches an incorrect answer.

Start with the work that repeats

A small company often contains an unofficial knowledge base distributed across people's memories. One salesperson remembers which customer needs an unusual invoice. Another knows that a product description contains an outdated specification. The owner knows when to make an exception.

Buying a tool does not automatically convert those memories into reliable assistance. It can simply make uncertain answers arrive faster.

Map the recurring tasks before choosing software. Suitable candidates might include drafting responses from an approved FAQ, summarising a meeting for the participants or turning verified product information into a first draft. These are examples to assess, not universally safe deployment recommendations.

For each task, identify its inputs, expected output, common mistakes and reviewer. If nobody can describe what correct completion looks like, you are not ready to measure improvement.

This exercise may reveal that the bottleneck is elsewhere. A salesperson waiting two days for pricing approval does not primarily have a writing-speed problem. A faster proposal draft will sit in the same queue.

Measure the whole task, including the awkward minutes

Suppose a hypothetical employee currently needs twelve minutes to prepare a routine reply. AI reduces drafting to three minutes, but checking takes five and correction another two. The net saving is two minutes, not nine.

Those are illustrative figures, not a study result. Their purpose is to make hidden work visible.

A fair measurement includes preparation, generation, checking, correction and downstream rework. It also records whether the task was completed correctly, whether the customer needed to return and whether another employee had to intervene.

An experiment that excludes review will naturally favour a tool that shifts work to reviewers. That may be good for the demonstration and bad for the business.

Use AI where the source of truth is clear

A bounded assistant has a better chance of being useful when the business can identify authoritative material. That might be a current price list, approved product specification or documented process.

Give those sources an owner and review date. Remove obsolete duplicates. Specify what the assistant should do when the answer is missing: ask for clarification, escalate or state that the information is unavailable.

The escalation instruction matters because a fluent answer can disguise an information gap. An employee may reasonably assume that a confident system retrieved something they had forgotten. A source link and an explicit uncertainty statement make checking easier.

For customer-facing work, separate drafting from sending during the initial pilot. An employee should examine the actual commitment being made, not merely the grammar. Delivery dates, eligibility, refunds and quoted amounts deserve specific attention.

This is an operating recommendation, not a guarantee that review eliminates all errors. Review quality itself depends on the employee's knowledge, workload and access to the underlying records.

The most experienced employee may be the best teacher

The support study's variation across workers suggests a practical question: what can less experienced employees learn from a well-designed system, and what should experienced staff continue to challenge?

Do not turn your best operator into a permanent correction service without recognising the cost. Ask them to help identify common cases, ambiguous cases and the boundary beyond which the tool should escalate.

A useful pilot can also reveal weaknesses in training material. If reviewers repeatedly disagree about the correct response, the company may need a clearer policy before it needs a better prompt.

Invite staff to report when an output looks plausible but is wrong. Reward those reports as improvements to the process. If every mistake becomes an accusation that someone failed to embrace technology, errors will become harder to see.

The point is to build reliable assistance that employees can use thoughtfully. An adoption target measured only by logins tells you little about whether work improved.

Build a business case without imaginary salary savings

Time saved can create capacity, improve response speed or reduce overtime. It does not automatically become cash in the bank. If staff remain employed for the same hours, the financial benefit depends on what they do with the freed capacity.

Record software cost, setup time, knowledge maintenance and reviewer effort. Add the cost of integration only if you actually need it. A simple pilot should not require rebuilding the entire company.

Then describe the intended value honestly. Perhaps more enquiries can receive a useful answer on the same day. Perhaps a new colleague becomes competent sooner. Perhaps managers spend less time rewriting routine communications.

Those benefits can be worthwhile without pretending that the full wage associated with every saved minute disappears. Use financial advice for accounting decisions and keep the operational evidence separate from an ambitious valuation slide.

Our article on Zoho's marketing-budget decision makes a related point: a changing cost structure is a reason to examine your own constraints, not copy another company's allocation.

A pilot that a small team can actually run

Choose one recurring task and a limited group of employees. Define eligible cases so that the assisted and unassisted work are reasonably comparable. Complex escalations should not be quietly excluded from one side and included in the other.

Collect a baseline before introducing assistance. Record task time, quality criteria, reopening and employee experience. Avoid sensitive customer data in external systems unless the organisation has established appropriate authorisation and handling arrangements.

Run the pilot long enough to encounter ordinary variation. There is no universal number of days or cases that makes the result valid. If volume is low, describe the outcome as preliminary and use interviews alongside the numbers.

Have someone review a sample without being told which method produced it where practical. This reduces the temptation to judge a polished AI draft more generously simply because it is new.

At the review, ask whether the benefit persisted after the novelty wore off, whether checking became a bottleneck and whether the task should continue with assistance. Stopping an unsuitable pilot is a useful result.

Decide what the system is allowed to decide

There is a meaningful difference between suggesting wording and authorising a refund. Write that distinction into the workflow. The employee should know which actions require approval and where responsibility sits.

For example, a hypothetical service company might permit AI to summarise a maintenance enquiry and draft a response from approved material. It might require a human to confirm scope, availability and price. That boundary reflects business risk rather than a general claim about what all AI systems can safely do.

Keep an audit trail proportional to the task. You need enough information to understand an error and correct its source, without retaining unnecessary personal data indefinitely.

Also plan what happens when the tool is unavailable. A business that cannot serve a customer without one vendor's assistant has accepted a dependency. That may be reasonable, but it should be a conscious choice.

What about marketing content?

AI can help organise research notes, suggest questions and prepare a rough structure. Publication still requires factual verification, original judgement and someone accountable for the finished piece.

The dangerous shortcut is to automate volume before establishing a reason for the articles to exist. A company can publish more pages and make its website less useful if those pages repeat the same claims or obscure the few strong pieces.

Ask whether assistance increases useful evidence or merely increases the speed of paraphrase. A better product comparison, a carefully verified explanation or a clearer customer guide can justify the work. A quota of posts does not establish reader value.

The same whole-task measurement applies. Include research, source checking, editing and corrections after publication. An article that takes minutes to generate and hours to repair is not a five-minute article.

A plausible future is uneven improvement

Our expectation is that adoption will produce uneven results across tasks and businesses. This is an editorial inference, not a quantified forecast. Work with clear inputs and repeatable checks offers a different opportunity from work requiring negotiation, physical inspection or unresolved judgement.

As tools change, the best candidates may change too. The durable capability is the ability to define a task, establish a baseline and examine the complete result.

That capability is useful even without AI. It forces a company to clarify where knowledge lives, who owns decisions and what customers actually experience.

Buy back useful attention

The most valuable outcome may be a colleague who spends less time reconstructing routine information and more time understanding an unusual customer problem. That is harder to advertise than an enormous productivity percentage. It is also closer to how a small business earns trust.

Before buying the next subscription, choose the task you want to improve and the evidence that would convince you it improved. If that conversation exposes a messy handover, repair the handover too.

The Umbrella works on the connection between business strategy, communication and digital systems. If your team is generating more output while work still stalls, bring us the workflow. The first useful question may be where the work waits.

Sources & Further Reading

• Brynjolfsson, Li and Raymond: Generative AI at Work, revised November 2024. Workplace study; the 15% estimate refers to issues resolved per hour in the studied support operation.
• NBER Reporter: The Economics of Generative AI. Background discussion; uses earlier research estimates, which should not be mixed with the revised paper.

- PIB: Government measures for MSME AI adoption, 30 July 2026. Policy intent, not measured firm-level productivity.