402 billion tokens. That is the volume of data processed by large language models daily in the current enterprise sector, and Sqwish Labs just moved to own a slice of that pie with a new training methodology that effectively allows companies to “fine-tune” general-purpose decision engines for their specific, messy, proprietary workflows. The move isn’t just a product update; it is a signal that the era of off-the-shelf AI is ending, replaced by a niche market of hyper-customized, data-heavy decision models that rival the complexity of human analysts. Sqwish’s new tool, released this week, lets firms ingest their own historical datasets—supply chain logs, customer service transcripts, or financial ledgers—to teach an AI not just what to do, but how to weigh conflicting variables in real-time, a capability that has been the holy grail of operational intelligence for a decade.
This features directly challenges the current dominance of general-purpose large language models (LLMs), which are great at generating text but terrible at making consistent, data-driven operational decisions without extensive, costly prompt engineering and external retrieval-augmented generation (RAG) pipelines. For tech leaders, the question shifts from “Can we build AI?” to “Can we afford to make it unique?”
The Death of the Off-the-Shelf Algorithm
Sqwish Labs did not invent the concept of custom training. But it made it accessible. The new platform, called “Decision Core,” allows non-engineers to upload CSV files and JSON logs, then use a no-code interface to define “decision weights”—essentially telling the AI that on-time delivery matters more than cost savings in a logistics scenario, or that customer retention is worth $200 in credit per churned user in a subscription model. It’s a clever tweak on a familiar problem.
Previously, companies had to hire squads of data scientists to fine-tune BERT-based models or build complex decision trees in Python, a process that took months and required massive compute resources that only FAANG-level companies could afford reliably. Now, Sqwash is commoditizing the “last mile” of AI customization. That is a huge shift in power dynamics for mid-sized enterprises.
The technical leap is subtle but significant. Most current AI tools are “stateless”; they treat every input as a blank slate, which leads to inconsistent decisions when context matters. Decision Core introduces a lightweight memory layer that doesn’t store the full conversation history (a privacy nightmare) but instead stores the “state” of the decision logic as it evolves. This means the AI can remember that a specific client has a high churn risk and adjust its pricing recommendation accordingly, without needing to re-process the entire contract every time.
It’s a move from reactive AI to proactive, context-aware decisioning. And it works remarkably well in controlled trials where Sqwash claims a 34% reduction in human override rates for automated approvals. But the real test is in the wild, where data is dirty, contradictory, and often incomplete.
Why “Decision Models” Are Outperforming Chatbots
For years, the AI industry focused on natural language generation. We spent billions training models to write email, code, and poetry. But the real value in enterprise AI isn’t in chat; it’s in operations. A chatbot that answers HR queries is a nice-to-have.
An AI that correctly approves a $50,000 loan while flagging a $5,000 loan as high-risk is a business-critical asset. This is the “decision model” space, and it is growing faster than consumer AI. Sqwash estimates that decision-model spending will grow 28% year-over-year, outpacing the 18% growth rate of general enterprise software. The drivers are clear: efficiency, consistency, and the sheer volume of small decisions businesses make every day that human staff cannot handle at scale.
Every invoice, every shipment, every refund, and every risk assessment is a data point waiting to be optimized. The problem with previous approaches to this was the “black box” issue. Traditional machine learning models would make a decision, but you couldn’t easily explain why. Regulators in finance and healthcare hate that. Sqwash’s approach attempts to solve this by building the decision logic as a series of interpretable rules that the AI learns to weight, rather than a opaque neural network.
It’s a hybrid model. The AI learns the pattern, but the logic remains transparent enough for audit trails. This is a crucial differentiator for industries where explainability is the law, not just a best practice. And it suggests a future where AI doesn’t just guess, but reasons through a structured, auditable process that feels less like magic and more like a very fast, very precise clerk.
The Data Bottleneck: Why Your Data Is the Product
Here is the uncomfortable truth: the model is only as good as the data you feed it. Sqwash is betting that companies have more high-quality, proprietary data than they ever have, and that this data is currently sitting underutilized in data lakes and legacy systems. But the reality on the ground is messier. Most companies aren’t ready.
They have siloed data, inconsistent formats, and poor metadata. To use Decision Core effectively, you need clean, labeled data that defines what a “good” decision looks like. If your historical data is full of human bias, errors, or inconsistent criteria, the AI won’t fix it—it will learn and amplify it. This is the “garbage in, garbage out” problem, now scaled up to the enterprise level.
It’s a hard sell to data teams who have spent a decade complaining about data quality. There’s a deeper implication here: data is becoming the primary competitive moat. If your competitors are using generic OpenAI or Anthropic models, their insights are limited by the public data they were trained on. But if you train a decision model on 10 years of your own proprietary supply chain data, you have an edge they can’t buy. You have a “taste” for your specific market that no generic model can replicate.
This shifts the value proposition from software-as-a-service (SaaS) to data-as-an-asset. Companies that invest in data hygiene and labeling now aren’t just preparing for AI; they’re building a proprietary advantage. And that changes how COOs think about their IT budgets. It’s no longer just about buying software; it’s about curating a dataset that will become smarter over time.
What This Means for India’s Tech Ecosystem
India stands to benefit disproportionately from this shift. The country is already a global hub for data services, with firms like TCS, Infosys, and Wipro processing massive volumes of data for European and American clients. But most of that work is commoditized; it’s about moving data, not making sense of it for decision-making. Sqwash’s model creates a new tier of demand: the need for “decision engineering.” Indian tech firms are uniquely positioned to capitalize on this.
They have the cost advantage to handle the data cleaning and labeling work that Western companies struggle to make profitable. More importantly, India’s own digital public infrastructure—UPI, Aadhaar, and the National Digital Health Mission—creates a massive, high-quality dataset for decision models in finance and healthcare that no other country can match. Indian startups can build vertical-specific decision models using this data, offering solutions that are cheaper and more relevant than Western alternatives.
For Indian readers, this is a signal to upskill. The demand for “AI engineers” who just know Python is dropping. The demand for “data engineers” and “decision scientists” who understand the business context behind the data is skyrocketing. If you work in a bank, a hospital, or a logistics company, you are not just a user of AI; you are the source of its intelligence.
The ability to define what a “good” decision looks like in your industry is now a high-value skill. And for investors, this is a green light for Indian AI startups that focus on vertical-specific decision models rather than general-purpose chatbots. The market is moving away from generic tools and toward specialized, data-heavy intelligence, and India has the talent and the data to lead that charge.
The Security Risks of Custom Training
Every time you train a model on your own data, you are creating a new attack surface. If an attacker can inject malicious data into the training set, they can poison the model’s decision-making. This is known as “data poisoning,” and it’s becoming a major concern for regulators. Sqwash claims to have built in “anomaly detection” during the training process to flag suspicious data points, but security experts warn that no system is truly immune.
The risk is amplified because decision models are often given higher privileges in a company’s system than chatbots. If a poisoned model approves a fraudulent transaction or releases a sensitive document, the damage is immediate and financial. That’s a harder sell to CISOs than “we’ll make your customer service faster.”
There’s also the issue of intellectual property. When you train a model on your proprietary data, does the resulting model become your property? Or do parts of your data leak into the model’s parameters in a way that could be reverse-engineered by a competitor? The legal framework for this is still non-existent.
We’re flying blind. Sqwash, like most AI companies, requires a long-form Enterprise Data Agreement that shifts liability to the customer. That’s a standard clause, but it raises the stakes. You’re not just buying a tool; you’re taking on the legal and security risk of training a system that makes money-critical decisions.
And that’s a barrier to entry that only larger companies can easily overcome, potentially leaving smaller firms behind. The timing coincides with a surge in enterprise AI adoption, where companies are moving from experimentation to production. General-purpose LLMs are hitting a ceiling in terms of operational reliability, creating a market gap for specialized, customizable decision models. Sqwash is moving to fill that gap before larger competitors like Microsoft or AWS can lock down the niche with their own native solutions.
Frequently Asked Questions
Why did Sqwash Labs release this new tool now?
How does this affect Indian data service providers?
It creates a new, higher-value segment for data services. Instead of just cleaning and moving data, Indian firms can now offer “decision engineering” services, helping global clients structure and label their data for AI training. This shifts the value chain upstream, allowing Indian companies to capture more margin and build deeper relationships with their clients.
Is it safe to train AI models on private company data?
It is safe if implemented with strict governance, but it carries inherent risks. Data poisoning, IP leakage, and regulatory non-compliance are real threats. Companies must adopt robust data hygiene practices, use isolated environments for training, and consult legal experts to define the ownership and liability of the resulting model.
Who are the main competitors to Sqwash Labs in this space?
Direct competitors are few, but indirect threats come from major cloud providers like AWS, Azure, and GCP, which are building native fine-tuning tools. Additionally, specialized AI labs like Fine-Tuning.io and various open-source projects offer similar capabilities but lack the enterprise-grade support and security features that Sqwash provides.
What is the next step for companies considering this technology?
Start with a pilot project. Choose one specific, high-volume decision process—like invoice approval or customer support triage—and test the model on historical data. Measure the accuracy and cost savings before scaling. This approach minimizes risk and allows teams to understand the data quality requirements before committing to a full deployment.
Source: TechCrunch
Source: TechCrunch
