Cambridge-based CuspAI has raised $450 million at a $2.6 billion valuation to use AI for materials discovery. What stands out is not just the model. It is the system around it: generative design, simulation, synthesis planning, and experimental validation. That is where AI becomes a product.
We see the same pattern in document processing
Extracting text is the easy part. The hard part is preserving meaning.
Take tables. A weak pipeline flattens a table into a block of text and hopes the model can reconstruct it later. A reliable pipeline does more. It detects table boundaries and merged cells, preserves rows, columns, and repeated headers, and stores the original table as HTML, Markdown, or structured JSON with cell-level links back to the source page. It creates row-level chunks with the column headers attached, then indexes both the table summary and the individual rows, using hybrid retrieval across embeddings, keywords, and structured filters. When one row matches, it retrieves the full parent table rather than the fragment, and validates totals, dates, and relationships before the result is ever used.
In many cases, the same table should exist in two forms: structured storage for filtering, calculation, and validation, and vector storage for semantic retrieval and question answering. That is the engineering most AI demos leave out.
The same applies to agents
Generating an action is easy. Making sure it has the correct parameters, dependencies, permissions, retry behaviour, and audit trail is the actual product work.
A model can perform well while the system still fails, because a header was detached from its row, stale data remained indexed, the wrong table was retrieved, or an operation ran twice.
Model accuracy is not system reliability. The model may be 20%. The other 80% is building the path from messy input to a trustworthy business outcome.