When we built KlaritiQ, we started from a hypothesis: most Indian companies believe they are further along on AI than they actually are. The published research backs it up, with one important nuance. (For how we turn that into a single comparable number, see how the AI Readiness Index is calculated.)

India's MSME Digital Maturity Index sits at about 58 out of 100 (Vi Business, 2025), and only 12% of MSMEs reach full digital maturity. For mid-size manufacturers specifically, CII-KPMG puts maturity at 2.9 out of 5, early-stage but real. Not laggards, but not leaders either.

What "early-stage but real" looks like

It usually means a company has done several things right: leadership has named AI a priority, a pilot or two is running, someone owns a roadmap on paper. The gap is rarely ambition, it's execution.

Companies are building on a foundation they haven't finished constructing, making AI announcements while data quality and governance go unaddressed beneath the surface.

Where the real gaps are

Talent is the binding constraint. ML, data-science and architect roles run a 60–73% demand-supply gap in India, and a data scientist costs ₹12–28 LPA fully loaded. For most mid-market firms, hiring a team for a first project is costly and high-risk, buying or renting capability beats building it.

Data is the under-counted bottleneck. Companies report having "a lot of data," but volume isn't readiness. Data preparation alone is 30–60% of an AI project's cost, the single most under-budgeted line item.

Build-vs-buy is where budgets are won or lost. Purchased or partnered AI tools succeed roughly 67% of the time; first-time internal builds succeed about a third as often (~22%, per NANDA, research with varying success definitions). Moving a custom build from proof-of-concept to production raises cost 250–400%.

Governance is the most urgent gap. With the DPDP Rules gazetted in November 2025 and core obligations enforceable from May 2027, 2026 is a build year, not a grace period. Using existing personal data for a new AI purpose requires fresh consent, and liability stays with you even when the AI is an outsourced tool.

What separates the movers from the stuck

Three moves consistently mark the firms that pull ahead:

1. Name an owner. A single accountable person for AI, policy, data, and a first use case, beats a committee. Enthusiasm without a mandate stalls.
2. Fix data before models. Audit the data behind your top use case first. Most stalled initiatives are blocked by data problems, not model problems, and data prep is where the cost hides.
3. Buy or rent your first win. At this scale, off-the-shelf SaaS or a managed API gets you a working result faster and cheaper than building, and lets you prove value before committing.

One variable predicts AI success in Indian mid-market firms more reliably than budget, headcount, or the number of AI tools deployed: the presence, or absence, of a dedicated internal AI champion.

It's a pattern that repeats. The firms that move have a named, empowered champion driving adoption; the ones that stall have enthusiasm spread across a committee and owned by no one. Against a talent market where ML and data-science roles run a 60–73% demand-supply gap, an internal owner who can direct AI, rather than an expensive new specialist hire, is often the difference between a pilot that ships and one that quietly dies.

What an AI champion actually does

This isn't a data scientist. It's not the CTO doubling as "the AI person." It's a dedicated operator, someone whose explicit job is to drive AI adoption across the organization, not ship models.

1. They speak both languages, engineering and business, without code-switching between teams.
2. They run toward the messy organizational problems rather than deferring to later.
3. They can identify a 90-day quick win, execute it, and tell the story in a board meeting.
4. They have a personal conviction about AI that isn't easily eroded by setbacks or skeptics.

The job description you actually need

What they're responsible for: Identifying high-value AI use cases across departments. Running the internal AI literacy program. Coordinating between data, engineering, legal, and business units. Tracking the company's ARI score over time.

What they are not responsible for: Building models. Managing the data platform. Generating R&D output. Those are engineering functions. Mixing them with the champion role is how organizations burn out good people.

If you can't hire yet, name someone

Early-stage companies often can't make a full-time hire immediately. The fallback isn't to leave the chair empty, it's to formally name an existing team member as the AI champion, with 20–30% of their time and explicit mandate. Our data shows that even a part-time champion, with real authority, produces measurable ARI improvement within one quarter.

The most common thing founders and CTOs say when we start a KlaritiQ assessment: "Our data situation is pretty good, we've been collecting it for years." And then the Data Foundations dimension scores come back, and the conversation changes.

Across Indian mid-market firms, the bottleneck is data readiness, not data volume. Companies have enormous datasets; what they lack is data that's owned, clean, and accessible to a model. Volume and readiness are completely different things, and the gap is expensive: data preparation alone runs 30–60% of an AI project's cost, the single most under-budgeted line item.

The four failure modes we see in every assessment

1. No data ownership

Data that "belongs to everyone" belongs to no one. Without a named owner for each key data domain, there are no quality standards, no update cadences, and no one to call when an ML model starts behaving strangely because its training data drifted.

2. No data lineage

Can your team answer: "Where did this column come from, and what transformations has it been through?" If not, you can't trust what you'd use to train a model. Lineage is the chain of custody for data.

3. Siloed systems that can't talk

The data is in three CRMs, two data warehouses, an old MySQL database from 2019, and three different BI tools. Each team's definition of "a customer" is slightly different. This isn't a ML problem, it's a data architecture problem that blocks ML entirely.

4. No quality gates in the data pipeline

Data enters the system and nothing validates it. Duplicate records, null values in critical fields, schema drift from upstream API changes, these compound silently until they surface as a model that starts making confidently wrong predictions.

The 30-day fix that changes the trajectory

1. Pick one use case. Identify its three core data inputs. Don't generalize. What data, specifically, would this model consume?
2. Run a data quality audit on those three inputs. Completeness, accuracy, consistency, timeliness. Measure it. Quantify the gaps.
3. Assign an owner to each data domain. One person. Accountable. In writing.
4. Add one quality gate to the pipeline. A dbt test, a Great Expectations check, an anomaly alert in your warehouse. One gate, running in production, before you train anything.

When governance comes up in an AI readiness conversation, the default reaction from founders and product teams is a kind of respectful dismissal. "Yes, we'll get to that." In 2026, that posture is a business risk.

The numbers

Three concrete risks that land on your P&L

1. DPDP Act compliance

India's Digital Personal Data Protection Act is not a future consideration. AI systems that process personal data require data principal consent frameworks, breach notification protocols, and data fiduciary obligations. Companies without governance infrastructure aren't just ethically exposed; they're non-compliant.

2. Enterprise customer requirements

If your company sells to mid-market or enterprise customers, AI governance questionnaires are now standard in vendor security reviews. Customers are asking: Do you have an AI usage policy? Do you conduct bias audits? Without answers, deals stall or don't close.

3. Model failure liability

When an AI system makes a consequential error, the question immediately becomes: what oversight was in place? Companies with governance frameworks have an answer. Companies without one have exposure.

What to build first (the minimum viable governance stack)

The Responsible AI Policy (2–4 pages): What AI systems are you building or using? What decisions do they inform? What are the prohibited use cases? Who reviews AI systems before deployment?

The Model Risk Assessment Template (1 page per model): What data does this model use? What's the failure mode? Who monitors it in production? When was it last audited?

The AI Inventory (spreadsheet is fine): A list of every AI tool or model in use across the company, including the third-party SaaS tools with embedded AI features that most companies forget about entirely.

The 90-day governance sprint

A focused 90-day governance sprint, write the policy, complete the inventory, run one model risk assessment, is enough to move a firm from improvising compliance on every project to a defensible baseline. And for an ordinary mid-market manufacturer, the heaviest obligations (like algorithmic due diligence, which bites only on notified significant data fiduciaries) don't even apply yet. That's exactly why the build year is the cheap time to act.