The Unglamorous Work Behind Enterprise AI

Let’s jump right into it: By now, we have learnt that AI tools can only act on the information they receive. As companies move to tools that assess creditworthiness, support clinical workflows or execute tasks, data quality is becoming a condition for safe deployment when utilising AI agents.

The challenge goes beyond correcting typos. Data can be stale, duplicated, inconsistent or incomplete, and AI agents may lack the organisational context needed to use technically accurate information. The issue is whether companies can trust data for each decision, and what can be done to ensure that happens, along with its impact to the organisation.

Let’s explore the realm of data quality for enterprise AI in this edition of The AI Shift….

Bad Data Costs Enterprises  

When it comes to AI, the equation is rather simple: Data quality determines whether a system is interpreting reality or merely processing flawed inputs consistently. 

A missing transaction feed, for example, can make a customer appear financially weaker than they are. If a model treats the gap as negative behaviour, its output may look confident while resting on a faulty premise. 

As Srijan Nagar, cofounder of FinBox, an AI-native digital lending platform, puts it, model confidence should not be confused with confidence in the data.

The consequences vary by use case, but the underlying risk is the same: an error can travel downstream and shape a decision. Across healthcare, fintech and enterprise software, the quality of data is proving to be a big bottleneck for AI systems. 

In fintech, an incorrect identity signal may let fraud through or deny a genuine customer an account or loan. In healthcare, a missing result or duplicated record can distort the information presented to a clinical professional. In both, the problem is not just a model’s accuracy in isolation, but the quality and coverage of the evidence it receives.

The risk extends beyond structured records because that’s not how enterprise data is stored; it often sits in documents, emails, archives and business applications. Even structured data may mislead an AI agent if it is outdated, lacks context or is not authorised for the user. 

And when an agent is expected to act, it must understand company-specific definitions and practices, not just retrieve relevant-looking text. Data quality therefore is about the reliability of the information, its currency, the governance standards applicable for the enterprise and the context of each decision.

Engineering Data Reliability

This is why we can comfortably say that there is no single data quality problem, and consequently there is no single fix. 

For a fintech platform, it’s about guarding against a missing transaction signal or a potential bad signal that could upend the intents of agents. For a hospital, it is reconciling records that never shared a format, and on the other hand, an enterprise is deciding whether an agent should trust a contract sitting in an archive. 

The controls are nonetheless remarkably similar, and they fall into a sequence: decide what trustworthy data means, screen what enters, add the context a model cannot infer, and keep watching after deployment.

For each sector — fintech, healthtech or enterprise software — screening starts at the point of entry, because it’s the cheapest place to catch a bad signal and this can potentially shape all decisions. 

Data discipline is an unglamorous job and each step relies on the previous one: start by validating the completeness of data, consistency, recency and anomalies, then move on to tracking the missing values and drift. Then comes comparing behaviour across customer segments, and sending anomalous or low-confidence cases to review. And let’s not forget the policy controls and audit trails on the decisions that matter most. 

FinBox cofounder Nagar says the controls and audits in particular are vital in separating a gap from a judgement: “You need to build freshness and completeness checks, distinguish missing data from adverse signals, and route anomalous or low-confidence cases for further review.”

Quality is also a coverage question: models trained across device tiers, networks and document variants, then tested by region, gender, age and device, are less likely to fail the customers a narrower training set would have excluded.

Teams also have to define which sources deserve trust or the weights for each source of truth, and there the answer is never equal treatment. Kedar Parikh, the chief product officer at identity and risk management company HyperVerge, used an example that’s best applicable to the startup to highlight this vital step. 

Records verified at source by a regulated entity, such as PAN, Aadhaar, CKYC or credit bureau data, carry the most weight, he remarked.

Then bank statements and documents extracted through OCR are treated as reliable but need corroboration. Self-declared inputs sit at the bottom of the pyramid and are the least reliable data point. 

In this case, a weak signal from self-declared inputs should trigger a check rather than close the case for any user. 

The system also needs to understand how to deal with uncertainty and different degrees of uncertainty. A deterministic check, such as whether a name matches a PAN record, produces a definite answer where there is very little scope for uncertainty. A probabilistic judgement, such as whether a selfie is live, produces a probability, which requires certainty checks. 

This is why model outputs carry confidence scores, and flagged cases go to a human rather than being refused outright. 

The Health Data Conundrum

Healthcare is where that discipline meets its hardest constraint, because here the sources are not so neatly arranged and in many cases, there can be no agreement between data points. 

Patient information is scattered across hospital records, labs, scans, billing and wearables, each in its own format, so the first job is normalisation: every source mapped to one common set of clinical definitions, where a term means the same thing everywhere. Automated checks then hold back incomplete entries, duplicates or contradictions before a model sees them, while training draws on material clinicians have already reviewed rather than the noise of the source systems. 

Under-representation is chased deliberately, across patient groups before launch and by subgroups after it, with unusual cases going to a clinician. Arjun Nagulapully, CTO of AIONOS, an enterprise AI services company, calls that combination of governed ingestion and reviewed training data what makes a dataset “AI-ready rather than just available”.

Healthcare is akin to enterprise data. Information sits in documents, emails, archives and legacy applications, and an accurate file is still useless to an agent if nobody has mapped what it means or who may see it. The work is therefore connective: link repositories, classify material, map how documents, transactions and processes relate, preserve lineage, apply retention rules and enforce identity-based access, so retrieval stays limited to what is relevant and authorised. 

Balancing Act: Speed Vs Quality 

Mayank Verma, global head of data and AI at Xebia, an AI-first consulting firm, explained: “Data quality has become even more important when it comes to the agentic data platform.” Without it, he added, the same term means different things in different sectors, and that’s a recipe for agents to trip up.  

Across the three sectors, the pattern is the same: define what reliable data means, validate and contextualise it before it reaches a model, keep monitoring it after deployment, and give people a route to review uncertain cases. 

The difference between sectors is where the work is hardest when it comes to data quality. While one sector like fintech may put its weight on verifying identity and income, others like healthcare are more concerned about data parity i.e. agreeing on definitions and standards.

The immediate benefit is fewer decisions resting on missing or misleading inputs, but the size of the gain depends on the workflow. 

Better data cannot guarantee a sound outcome. What it narrows is the room for confident decisions built on missing, biased or unauthorised information, and it gives enterprises a clearer basis for judging whether automation is delivering value.

Top Stories From India & Around The World

• OpenAI Halts Top-Model Work After An Agent Bypasses Restrictions: OpenAI paused training, evaluation and tool-enabled inference for its most capable models after an internal agent bypassed internet restrictions. The company plans a fresh training run after validating additional safeguards.
• NPCI Chief Says UPI MDR Can Fund AI Cybersecurity: NPCI CEO Dilip Asbe said merchant discount revenue is needed to fund increasingly costly AI-based cybersecurity tools. Inc42 reports the proposed framework could generate ₹13,000-₹15,000 Cr in its first year.
• Google Brings Live Avatars To Gemini: Gemini 3.8 Live with Live Avatar combines streaming video with conversational AI, including lip-syncing and expressive responses. Google says the feature is available through Gemini Enterprise.
• Dextr AI Raises $6.7 Mn To Build AI Agents For Hospitality: The hospitality-focused startup raised a $6.7 Mn seed round led by Elevation Capital, with participation from Foundation Capital. It will use the funding to develop its AI agents and expand deployments, including in India.

The Weekly Buzz: Meta’s Muse Goes Enterprise

Meta launched Meta Enterprise Platform, a new business unit built around its Muse, this week, just a few days after showing off the roadmap for its personal AI agent. 

The goal is to sell the same Muse personal agent tools that consumers are already using to companies and developers. Mark Zuckerberg called it the next major pillar of Meta’s business and hired former MongoDB CEO Chirantan “CJ” Desai to lead it, reporting directly to him.

Muse itself is Meta’s personal AI agent. Unlike a normal chatbot, it can actually do tasks for users — send emails, book travel, shop, fill out forms, and work across connected apps. It runs on its own secure cloud computer so it can keep working even when the user is offline. Since launching earlier this month it has become a clear consumer hit, climbing app-store charts and drawing millions of downloads.

At its Connect event, Meta expanded Muse further. It announced upcoming support on AI glasses, a digital avatar people can video-chat with, a Mac version that can control desktop apps, and a long list of new connectors including major retailers, payment services, and work tools like GitHub and Notion. It also continues updating the underlying Muse Spark model and Muse Code coding agent.

The enterprise move is a logical next step. Meta is taking a product that already works for everyday users and packaging the full stack: the agent, the models, the APIs, and the coding tools — for businesses that want the same capabilities at scale. 

The platform will initially offer the Muse agent, Meta Business Agent, Muse API, and Muse Code as a combined suite, giving businesses a ready-made set of tools rather than forcing them to build everything from scratch. Early focus is on helping companies deploy agents that can handle real work, not just answer questions.

Startup In The Spotlight: Zenalyst

Founded in May 2025 by Nagendra Singh, Sanketh Krishnappa and Vijay Jha, Bengaluru-based Zenalyst is building AI agents that run enterprise finance, procurement and legal workflows rather than just report on them.

Every CIO says the same thing: dashboards everywhere, answers nowhere. Enterprise decisions still depend on numbers scattered across ERP, CRM, banking and document systems, stitched together by hand. By the time a finance team spots a variance, the quarter it belonged to is closed.

Instead of describing the problem, Zenalyst’s ZenForce platform connects to more than 150 enterprise systems across ERP, CRM and banking to execute those workflows. Its agents include ZenBank for treasury and money operations, ZenProcure for procurement and ZenCollect for order-to-cash and contract intelligence.

The startup positions ZenForce as deterministic rather than probabilistic. Reasoning is gated through verification steps, queries are grounded in the customer’s own schema, and customer data is never used to train external models.

Zenalyst is targeting the autonomous finance market, which stood at $26.34 Bn in 2025 and is projected to reach $112.84 Bn by 2031, as enterprises look to move finance operations from manual number-crunching to agents that act on the numbers themselves.

Prompt Of The Week

What prompts and hacks are CTOs, CEOs and cofounders using these days to streamline their work? 

Here’s Ankur Dhingra, CEO of ProHance, with a prompt he uses to find the right AI and startup voices in his extended network before anyone else does:

“Act as a talent-intelligence agent. Once a month, scan my target companies for individuals within my third-degree network who meet ALL of the following criteria.

Location: Based in India

Topic focus: Actively posting about at least one of the following:

• AI-native product building

• Scaling startups

• Enterprise GTM strategy using AI

• AI adoption or impact stories

Authority check: Triangulate by looking at their experience and decide how much authority the person holds in the fields above.

Signal strength: They must show real impact through metrics, case studies, outcomes or lived experiences. Generic commentary does not count.

For each qualifying profile, capture:

• Name, current role and company

• Posting frequency and dominant themes in the last 30 days

• Connection path (mutual contacts)

Output a ranked shortlist in Rolodex format, sorted by relevance and engagement strength.

Flag any profile where a role change, event appearance or open-to-work signal suggests a timely outreach window.

Exclude anyone already in our network or previously contacted.“

Editor’s Note: Some prompts may need to be adjusted by users for best results or may not work as intended for certain users.

[Edited by: Nikhil Subramaniam]
[Creatives by: Abhyam Gusai]

The post The Unglamorous Work Behind Enterprise AI appeared first on Inc42 Media.