In 2008 a Mountain View company put a sentence on its website that we would sign today:

80-90 percent of the ESI collected for a case is non-relevant. Yet, enterprises continue to send the entire dataset to outside services for processing at an average cost of $1,800 per gigabyte.

Clearwell Systems, 2008

The $1,800 is a 2008 number and processing is far cheaper now. The idea underneath it has not aged: the legal team should control processing, because most of what it pays to process is junk, and nobody should pay by the gigabyte to find that out.

What it proved

Clearwell's version was a 2U appliance a legal department racked on its own network. By 2010 it advertised one-click NIST-list filtering, deduplication across the whole data set, integrated OCR, 400 file types, a terabyte a day, and export to any review platform in EDRM XML, DAT or CSV. Microsoft, Boeing, GE and a run of Am Law 200 firms bought it. By 2011 it had 400 customers and $56 million in revenue, Gartner put it in the Leaders quadrant of its first eDiscovery Magic Quadrant, and Symantec paid $392 million for the company. Taking control of processing was not a theory. It was a market.

Where it fell short

It was a culling tool, not a review platform. Analysts at the time called Clearwell “more of a specialist in ECA,” and the market agreed: the common workflow was to cull in Clearwell and review in Relativity.

It was built for a different time. The record of a dispute in 2010 was email and Office files; Slack did not exist until 2013 and Teams until 2017. Search meant keywords. Review meant people. AI meant a statistical model that had to be trained on a sample for every matter and could not say why it made a call. None of that was a flaw then. All of it is a limit now.

It priced by data volume, and its owners still do: $65,000 per 100 gigabytes in 2009, a Gartner caution in 2011, and a terabyte of processing licensed at $835,773.91 over four years in 2026. The customer's cost was tied to the one input that has since become nearly free.

And it stopped being a product. Under Symantec, Clearwell revenue came in below expectations within a year, the CEO left for Sequoia, and Gartner moved it from Leader to Challenger by 2015 before the whole division was sold. A company became a feature of an archive.

What we kept

The idea, and the outputs. Marbury's engine takes control of processing the way Clearwell did: DeNIST, deduplicate, OCR, index, and hand a load file to any review platform. The OCR is a neural model now and reads images as well as scans, but the job is the same. That is the parallel, and it is the whole of it.

What we built for the modern workflow

The data. Slack, Teams and text exports are converted at ingestion into readable, producible documents, in a format that did not exist when Clearwell shipped.

The index. Every document is chunked and embedded as it is processed, so the collection can be searched by meaning and not only by keyword, and read by an agent. Neither the technique nor the models existed in 2010.

The AI. First-pass review returns a call and a written reason for each document, banded rather than scored, with no training set to build and every call open to be overturned. Term reports and privilege logs are things you ask for.

The operator. There is no console to learn and no professional-services line. You say what you need, the agent plans it, and nothing runs until you approve.

And review is not someone else's product. It lives in the same place as the processed data, so a matter can be culled and reviewed here, or culled here and reviewed in the platform your firm already runs, with the coding and the OCR text attached. Clearwell's customers had to choose. Ours do not.

Sources