AI-first Series – Part 1
Scroll through any B2B enterprise website and count the number of times the keyword “AI”has been infused within the content: you would be forgiven for thinking AI has already been embedded into every organization.
Have enterprises truly become AI-first? Not quite, though we see a dramatic surge in investments in AI-initiatives. According to Mckinsey, two thirds of surveyed enterprises reported incorporating AI into at least one of their business functions in 2025, a significant jump from 2017, when it was 20%.
As organizations race towards plugging AI into their core functions, they seek data engineering consulting expertise to help build bridges between their existing enterprise systems and an AI-first future. But becoming AI-first requires a clear vision of what AI-first truly means in practice. This is where artificial intelligence integration comes in.
What’s AI-first and How to Become One?
AI-first could mean different things to different types of businesses.
- For a SaaS startup, it could mean AI integrated deep into their product functionality.
- For a large enterprise operating in a regulated ecosystem such as healthcare, AI-first represents responsible AI operating in an assistive capacity, their outputs always compliant and explainable.
- For a mid-sized manufacturing business, AI-first could mean agents performing tasks autonomously with minimal human interaction.
Well, the best way to understand AI-first is to clearly define what it is not.
What is Not AI-first?
Trigent’s Data Engineering Consulting Team, through its on-the-ground experience, outlines what is not AI-first:
- When AI is added as an afterthought or add-on to the existing product or service. For example, AI being added after a product has been designed.
- When AI operates in isolation, i.e., it is disconnected from core systems and not integrated with real-time data flows.
- When there is little synergy between human workforce and agentic workforce, i.e., Agents are deployed within organizations, but thanks to widespread distrust, the human workforce tends to ignore or not use them.
Most Enterprises are Currently AI-last and That Needs to Change
Unlike AI-first that can take many forms, AI-last can be understood in a relatively simple measure. If an organization has blob storage of data, scattered across multiple enterprise systems, unstructured, unlabelled, and inaccessible to AI models— we can safely presume that the organization is operating in AI-last mode.
However, going by that definition, the number of enterprises that are AI-last may likely be very large. Consider this IDC statistic: 80% of enterprise data is unstructured, and Forbes reports that 90% of unstructured data still remains untapped in organizations.
How to Shun the AI-last Status?
Trigent’s data engineering solutions and artificial intelligence consulting engagements have helped unearth the following insights:
The first step in avoiding the AI-last status is for humans to take stock of the massive unnamed datasets, understand what each data item pertains to, and make explicit connections between data items. It is what we call data labelling. Much of the effort is likely to be manual.
Enterprises will either hire internal or external teams to manually review, tag, and categorize data assets such as documents, emails, videos, and images. The objective is to look at these raw images, say a scanned form or a product image, and accordingly add meaningful labels. The labels provide a basic structure, meaning, and context for the ML models to make sense of the data items.
The amount of grunt work will differ for an enterprise with more analog data than digitized data. In such cases, the labelers often have to start from scratch. They need to digitize paper records using OCR, and label structured fields (Patient ID, Policy number, Diagnosis) as well as unstructured data (e.g., doctor’s notes, claims narratives).
In contrast, digital enterprises that have done away with analog ways of working may already have digitized data that are structured yet scattered across systems (emails, logs, CRM entries). In such cases, the data is brought together through warehouses and data lakes. After this, natural language processing (NLP) is employed to extract entities (product name, return reason, sentiment). Labelers will then focus on establishing semantic meanings for unstructured data, spending significant effort on creating intent labels (returns, escalation, delay) and semantic labels (frustrated, confused).
Is Manual Data Labelling Necessary for Digital-first Companies?
Not necessarily, because these SaaS-type companies operate in well-defined environments where much of the data is either structured or semi-structured, while their unstructured content has clear semantic anchors. Their products inherently capture enough information about the users (who they are, what they are scrolling, and what issues they are facing). The product environment would have already captured user role, user actions, error events, feature usage, and so on.
In such cases, since there are enough meaningful signals around a datapoint, enterprises can bypass the manual labelling work and start employing pre-trained AIs (Open AI or AWS comprehend) to begin tagging the content at scale. Remember, generative AI can build on the semantic anchors and further elaborate the context. Obviously, you would need humans in the loop to review, edit, and approve the AI tags.
So, When is an Enterprise Ready for AI-assisted Labelling?
As mentioned earlier, it is relatively easier for a digital-first enterprise to get a head start in AI-assisted labelling. Needless to say, other enterprises will need to transition to a digital-first approach. By digital-first, we mean their data should meet the following criteria:
1 Digitized (Data is stored in electronic form and not locked in PDFs or paper forms)
2 Centralized (Data is stored in accessible and unified systems like a warehouse or lake)
3 Standardized (Data stored in consistent schemas and structures across systems)
4 Consistent (Data flows in a reliable and repeatable way right from ingestion to transformation, to storage – a maturity milestone we often help clients reach through Trigent’s data engineering consulting services
.What’s Next After AI-assisted Labelling?
Can your ML models look at your tagged data and start deriving patterns and make predictions? And how far are you from creating your first assistants and agents – proving you are indeed AI-first?
Stay tuned for more insights from our data engineering consulting experts in Part 2.
