Affine wants AI to finish the job

In 2013, when the fashionable phrase was “big data,” Affine had a small but revealing complaint about its industry. Too many analytics companies, its founders believed, responded to difficult problems by adding people. The young Bengaluru firm wanted to keep more of the solution. Solve something useful, turn the repeatable parts into software, and let the next customer start further ahead.

That objection is a useful way into Affine today. Its vocabulary now includes agentic AI, conversational analytics and generative imagery. Underneath sits an older commercial question: why should every new engagement pay to rediscover what the last engagement taught you?

The objection that came before the agents

Manas Agrawal, Vineet Kumar and Abhishek Anand founded Affine in 2011. By July 2013, contemporary reporting described 40 employees and eight customers, five of them Fortune 500 companies. An angel round had arrived in 2012. The ambition was already to package domain knowledge into analytical platforms, with enough flexibility to fit different customers.

The founders’ reasoning was more interesting than a promise of faster mathematics. Data needed business meaning. A churn model mattered because a telecom operator could use it to retain customers. A report that stopped at a number left the difficult decision with the buyer.

“Data tells its own story, not always the one you want to listen to.”

Manas Agrawal · 2013

In 2021, a new brand identity emphasized three connected capabilities: AI, data engineering and cloud. Academic collaborations with IIT Bhubaneswar, IIT Kharagpur and Berkeley Haas’s Fisher Center for Business Analytics were part of that account. The firm was expanding the machinery around its analysis.

Two kinds of questions, two different tools

Consider a merchandising manager asking why sales slipped. Some answers live in tables: transactions, prices, inventory. Others live in documents: product notes, policies, customer feedback. Treating these as one undifferentiated pile of information misses how businesses store what they know.

Affine’s Quin addresses the structured side. A user asks a question in ordinary language; the system interprets intent, generates SQL, queries enterprise sources and produces insights or visualizations. Its stated controls include enterprise authentication and role-based access. The attraction is straightforward: fewer routine questions waiting for someone who knows the database syntax.

Eryl addresses enterprise knowledge retrieval. Its sequence is retrieve, reason, evaluate, refine. One agent gathers relevant material, another constructs an answer, and a critic checks qualities including relevance and faithfulness. That design makes evaluation part of answering, rather than leaving the reader to discover a confident mistake.

The catalog travels beyond question answering. IntelliTag and IntelliSearch target product discovery. CreAItive generates marketing visuals and evaluates them, with human feedback in the refinement process. Telescope segments satellite imagery to help analyze land and infrastructure. These are specific workflows, each with its own inputs, users and consequences.

The pipeline is part of the product

A clever interface still needs reliable information underneath. Affine’s data-modernization practice covers migration, pipelines and governance. In one anonymized consumer-goods example, the company describes replacing nightly batch processing with an event-driven lakehouse spanning more than 500 SKUs. It reports eliminating a four-hour daily processing window. The practical change is fresher information when somebody needs to act.

Manufacturing offers an equally concrete test. Affine describes a deep-learning inspection system for a consumer-goods production line, reporting roughly 99% defect-detection accuracy and about 95% less manual inspection effort. Those are company-reported results from a particular deployment. Their significance lies in the workflow: cameras, models and operating processes have to function together.

The same principle applies to business intelligence. Affine builds semantic layers that define metrics consistently, then connects dashboards and natural-language access. If two departments mean different things by “revenue,” an eloquent answer can deepen the confusion. Agreement about the measure is part of the engineering.

The company also describes a way to replenish that technical knowledge. Quarterly internal hackathons give researchers and engineers room to work on problems outside client briefs. Three continuing research groups then focus on frontier methods, production engineering and industry applications. The useful distinction is between an experiment and an asset: an experiment teaches the people who ran it; a documented, evaluated component can teach the next delivery team as well.

Make the next project inherit the last

Agentic AI Launchpad makes Affine’s original reuse argument visible. A process owner describes a problem; the platform turns it into a workflow blueprint and identifies existing components. Shared connectors, evaluation and governance are intended to prevent each department from building a separate version of the same agent.

Affine markets roughly halved build time and around 40% lower build costs on repeat use cases. These are its platform claims, not a quoted engagement price. The economic logic depends on genuine overlap between projects. Reuse is less valuable when requirements, permissions or integrations differ enough to force a rebuild.

Its business model keeps services close to the software. Affine says it prefers managed engagements, while also offering embedded teams, outcome-based arrangements and focused sprints. It competes for work an enterprise might otherwise give to its internal AI team, a specialist analytics firm or a large technology consultancy.

For a buyer, this changes the demonstration worth requesting. Bring a real question, a difficult document and the permissions the eventual user will have. Ask who checks the answer, how an exception reaches a human, and what happens when a data source changes. A working path through those conditions says more about fit than a polished exchange with a chatbot. It also exposes the work that belongs in the engagement.

The lesson a reader can borrow is modest and demanding: choose the decision first, agree on the data, test the output, and preserve useful components. Assign responsibility for operation and handover. Affine’s approach makes most sense where an organization has recurring, consequential workflows and people prepared to own them. The next project then begins with something better than another blank page.