Software WatchToday
None of the five publishes what happens when the extraction is unsure
TC Bulletin put the questions from its July story to the published material of ListedKit, Rebillion.ai, DocJacket, Trackxi and Nekst. Every one of them describes the speed. Not one describes the citation, the confidence level, or what becomes of your clients' documents.
What this story establishes
- Zero of the five state whether an extracted value cites the page or clause it came from.
- Zero of the five mention confidence levels or the flagging of values the model was unsure about.
- Zero of the five publish a policy on whether customer documents are used to train AI models.
- Four of five describe a human review step, in language ranging from 'you approve' to 'a quick review'.
- Two publish performance claims, 3X accuracy and 4X faster, with no methodology attached to either.
In July this publication set out five questions worth asking any platform selling contract extraction to coordinators. This month we put them to the published material of the five platforms named in that story. The exercise took an afternoon, which is roughly how long it should take a buyer.
The result is not that the products are bad. It is that on the two questions which decide whether an extraction can be trusted, the category is silent.
What each one publishes
Where they do say something
Human review is the one question the category does answer, and mostly in the same words. DocJacket puts it as AI preps, you approve. Rebillion.ai describes going from worker to reviewer, with the user approving and the model handling the rest. ListedKit says Ava does the work and your team reviews. Nekst describes a quick review before you are ready to go. Trackxi's material describes automatic extraction of names, prices, dates and contingencies without describing a review step at all.
Four out of five is a reasonable showing. It is also the easiest of the five questions to answer, because saying a human approves costs nothing and commits to nothing about how the approval is meant to work.
The claims that arrive without arithmetic
Two of the five publish performance figures. Rebillion.ai states it is up to three times more accurate than manual entry. Trackxi states contracts are processed four times faster with AI-powered data extraction. Nekst puts a number on the clock instead, describing extraction in about ninety seconds.
No methodology accompanies the first two. Three times more accurate than which manual baseline, measured on what document set, scored by whom. Four times faster than what. These may well be defensible internally. As published, they are not checkable, which is the same problem this publication identified in the unsourced market sizing circulating around coordinator insurance.
Speed is the easiest thing to measure and the least useful thing to promise. Nobody was ever sued for reading a contract slowly.
What this survey is not
This is a survey of what these companies publish, not of what their products do. A platform may cite sources beautifully inside the application and simply not mention it on a marketing page. Several probably do.
That distinction matters and it does not rescue the position. A coordinator evaluating five products cannot see inside any of them before buying. The published material is the entire basis for a shortlist, and on the questions that determine whether an extraction is auditable, that material currently says nothing at all. The first vendor in this category to publish a citation and confidence model in plain terms will have a genuine advantage, and it is available to any of them this week.
The five questions, unchanged
- Does every extracted value cite a page or clause in the source document?
- Are low-confidence values visibly flagged rather than silently included?
- Does anything reach the calendar before a human approves it?
- Is AI usage priced into the plan, or metered separately in a way that punishes a busy month?
- Are client documents excluded from model training, in writing?