AI for MSMEs: what it can actually do today

AI today is reliable for drafting, summarising, reading documents and messages, answering routine questions from your own data, and moving data between WhatsApp, sheets and accounting software. It is not reliable unattended for advice, negotiation, or any output a customer or regulator will act on without a human check. Studies reported in October 2026 put basic use high and deep use low: a CII and ISB study of 2,919 MSMEs, as reported in the guides reviewed for this article, found 65.8 percent use AI tools but only 21 percent run ERP or customised business software. Start with subscriptions costing a few thousand rupees a month. Treat projects costing lakhs as a second step, after your data is clean.

The short version: two completely different spends

Mixing these two spends is the most common budgeting error. A owner who approves a project in lakhs before the team uses a drafting tool daily has bought the roof before the walls.

For tool-by-tool prices, see Best AI tools for a small business in India and How to use ChatGPT in your business.

What it can do today, by job

Writing and content. Drafting quotations, replies, product descriptions, job posts and first versions of documents is the most widely reported use. In the CII and ISB figures as reported, business communication accounts for 34.9 percent of use and content creation for 22.3 percent, with training at 18.9 percent. Tool anchor: ChatGPT Plus is Rs 1,999 per month including GST in India, as of November 2026, with a lower Go tier at Rs 399 per month. Human-check rule: every price, date and commitment in a draft gets checked by a person before it goes out.

Customer replies on WhatsApp. Routine questions on price, availability, timings and order status can be answered from your own approved information, with assignment to a person when the question leaves the script. Tool anchor: WhatsApp platforms of the AiSensy type start at Rs 1,500 per month plus Meta per-message charges (anchor card: Marketing Rs 0.8631, Utility Rs 0.1150 per message, effective 1 October 2026; a BSP may price above the anchor). Human-check rule: discounts, credit terms and complaints always go to a person. For the full system choice, read CRM for a WhatsApp-first business.

Reading orders and documents into your systems. AI can read a WhatsApp order, a supplier invoice or a scanned document, extract the fields, and prepare an entry for Tally or Zoho, instead of a staff member retyping it. This is project territory, not a subscription feature. Consultancy bands reported for this kind of work are wide: reading WhatsApp orders into accounting vouchers is banded at Rs 1.5 to 5 lakh, and supplier invoice reading and matching at Rs 1 to 4 lakh, in the guide material reviewed. Treat every such band as quote-dependent. Human-check rule: entries are prepared by the system and posted by a person until error rates are measured and low.

Reports and dashboards. A daily sales and stock summary, prepared from your existing sheets or accounting data and sent to you on WhatsApp, is a realistic early win. Reported bands for a dashboard with a summary run at Rs 1 to 3 lakh as a project, while simple reporting inside a CRM you already pay for may cost nothing extra. Zoho CRM, for reference, is Rs 800 (Standard), Rs 1,400 (Professional) and Rs 2,400 (Enterprise) per user per month billed annually before GST, with its Zia AI features listed at Enterprise level on the pricing page checked. Human-check rule: reconcile the automated report against the source system weekly for the first month.

Design and voice. Social creatives, catalogue images and draft voiceovers are cheap to produce now. Canva Pro is reported in guide material at Rs 499 per month or Rs 4,500 per year in India; verify on canva.com before paying, since that figure is guide-reported in this check. Human-check rule: check every specification, size and claim in a generated creative against the real product before publishing. General AI assistants can state incorrect facts confidently, so prices, claims and specifications must be checked before anything is published.

What it cannot do reliably yet

Keep this list on the wall. It is the trust section of this article.

It cannot give reliable legal or tax advice unattended. It can summarise a circular or draft a reply, but filing positions and tax judgement remain your CA's responsibility.

It cannot negotiate with a buyer or a supplier on its own. Tone, context and relationship history decide negotiations, and an automated message can commit you to terms you never approved.

It cannot judge credit or hiring. It can organise applications and histories, but the decision to extend trust to a person or a firm is a human decision with human consequences.

It cannot fix broken data. Most AI projects fail on data and process before AI is involved. If you have three price lists, two customer lists and orders scattered across personal phones, the project will automate confusion at speed.

It cannot be trusted with unchecked customer data in free tools. Staff pasting customer lists into free chat tools is the most common leak identified in the guide material reviewed. The DPDP Act 2023 applies to customer data your firm holds, regardless of which tool your staff used.

Are you ready? A five-point checklist

Score your firm honestly. One point for each statement that is true today.

You have one current price list, in one place, that everyone uses. You have one customer master list, not separate lists per salesperson. At least twelve months of orders sit in one system or one set of sheets. One named person inside the firm will own any AI setup after the vendor leaves. Approval rules are written down: who may approve a discount, a dispatch or a customer commitment, and up to what limit.

If you tick fewer than three boxes, fix data first, at near zero software cost. Consolidating a price list and a customer master takes staff time, not licences, and it raises the value of every tool you buy later. If you tick four or five, a small subscription stack and then one measured workflow project are reasonable.

Pradrix is building technology for MSMEs together with the owners, CAs and experts who will use it, and early access will open to firms that want to be part of that process. If you would like to be considered when it opens, you can register for early access.

Costs in full: subscriptions and projects

Subscriptions first, using the worked example stack. Project bands second, stated honestly. The consultancy guide reviewed for this article gives conflicting bands inside the same page: its opening puts a full first automation at Rs 4 to 13 lakh, while its detailed section puts a first workflow automation connecting WhatsApp or email to Tally or Zoho at Rs 3 to 10 lakh, single-tool setups at Rs 25,000 to Rs 1 lakh, and a discovery sprint at Rs 1 to 3 lakh. The correct conclusion is not to pick one band. Project costs are wide and quote-dependent, driven by how clean your data is and how many systems must connect. Get the scope, the exclusions and the post-project support cost in writing.

A guide suggestion worth using as a guardrail, attributed as such: keep total AI and software spend under 2 to 3 percent of monthly revenue until savings are measurable. The worked example below sits inside that guardrail.

Worked example: an Indore packaged foods firm

Take an Indore packaged foods firm with turnover of Rs 5 crore, an owner plus a three-person sales and support team, starting with subscriptions only and no project in year one.

Monthly stack: ChatGPT Plus for the owner, Rs 1,999 per month including GST, as of November 2026. Canva Pro on the annual plan, Rs 4,500 per year, which is 4,500 divided by 12 = Rs 375 per month effective (guide-reported figure; verify on canva.com). AiSensy Basic, Rs 1,500 per month. Zoho CRM Standard for three users: 800 x 3 = Rs 2,400 per month.

Arithmetic: monthly total before Meta message charges is 1,999 + 375 + 1,500 + 2,400 = Rs 6,274 per month. Annual: 6,274 x 12 = Rs 75,288 per year. As a share of turnover: 75,288 divided by 5,00,00,000 (Rs 5 crore) is about 1.5 percent of turnover, inside the 2 to 3 percent guardrail when measured against revenue.

Add Meta charges illustratively: 1,000 marketing messages a month at the BSP-published Rs 1.09 figure used in the source material is 1,000 x 1.09 = Rs 1,090 per month, or Rs 13,080 per year, taking the illustrative total to Rs 88,368 per year. At the Meta anchor of Rs 0.8631, the same volume would cost less, which is why the BSP rate card question in the CRM article matters here too.

Payback, as a method rather than a promise. The consultancy guide reviewed calculates manual order handling for ten field staff at five hours a week each and Rs 150 per hour as 50 hours x 50 weeks x Rs 150 = Rs 3,75,000 a year. Run the same three-line calculation for your own team, with your own hours and wage rate, before approving any project in lakhs. If the arithmetic does not clear the project cost within a period you accept, do not approve the project yet.

Risks and safeguards

Data leakage through free tools is the first risk. Move customer-facing work to business accounts whose terms exclude training on your data, keep personal customer data out of general tools, and write a one-page staff policy naming what may never be pasted into a free tool.

Vendor lock-in is the second risk. Protect yourself with export rights for your data, documented workflows that another vendor can maintain, and a clear answer to one question before signing: if we leave in month seven, what do we take with us, in what format?

Cost creep is the third risk. Subscriptions renew quietly. Review the stack quarterly, cancel seats nobody uses, and measure time saved against the baseline you recorded at the start. The digitising roadmap in Digitising your business: a practical roadmap places AI steps inside the wider sequence, so AI does not outrun your basic systems.

A 30-day starting sequence

Week one: one drafting tool for the owner. Use it for quotations, replies and summaries. Record time taken against the old method.

Week two: WhatsApp routine replies. Put your ten most common questions into approved answers, with a clear handover to a person for anything sensitive.

Week three: one report automated. Choose the report you currently compile by hand every morning or every Friday, and automate only that one.

Week four: measure. Compare time saved against the baseline, list errors the human check caught, and decide whether month two earns a second tool or a first project conversation. No large project in month one.

How the adoption picture should change your decision

The reported studies point the same way, even though their exact percentages differ by sample and method. A Dun and Bradstreet study, as reported in the same guide material, says three in four MSMEs use AI in some form but only 5 percent place it at the core of operations. Read together with the CII and ISB figures, the pattern is consistent: basic tools have spread widely, while systems that run daily operations remain rare. That gap is not a reason to rush. It is a reason to sequence correctly. Firms that already use an assistant daily, keep one customer list and measure one automated report have the habits a workflow project needs. Firms without those habits should build them first, at subscription cost, and let the project decision follow the evidence of their own month-one measurements. The technology will still be there in a quarter, usually cheaper and better documented. Your cleaned data will be worth more by then as well, and it is the asset no vendor can sell you.