Three Days to an Hour: Where the Hour Goes

Selector Funds Management cut its DDQ turnaround from three days to about one hour. We name Selector as a lighthouse client on our research page. We use that result as one of the two client-derived inputs to our savings calculator, which shows it as "DDQ 3 days to 1 hour".

A turnaround figure makes a poor argument by itself. Every vendor in this market has a three-to-one story. A fund COO should look at the split underneath the figure: some of those three days went on work that no person should have done. Part of the remaining hour is work a sensible firm would refuse to automate even if it could.

Where the three days went

Most of a due diligence questionnaire from an asset consultant repeats questions the fund has answered before. The earlier answers went to other consultants on other templates, some under an older edition of a policy. Writing the answers takes little of the three days. An analyst spends most of that time finding each previous response and checking that it still holds. The analyst then re-cuts each answer to fit a form that asks the same thing in a new shape.

The analyst searches by hand, between other jobs. The search covers old DDQ packs, board papers, investment committee papers and policy documents with several live editions. It also reaches the mailbox of whoever handled the last questionnaire. One search like that takes an afternoon and produces nothing a client would pay for.

Our retrieval pipeline cuts this search from an afternoon to seconds without costing the firm anything of value.

A matched question returns an answer the firm already stands behind

We documented the March 2026 pipeline in our benchmark paper. That pipeline extracted and verified question and answer pairs at ingestion time. When an incoming question matched a pair, the system returned the verified answer instead of generating a new one. The system wrote nothing new. The firm reused an answer it already stood behind, the same reuse the analyst had been attempting by hand.

Two other design choices in that pipeline cut work on repeat questions. Recency ranking picked the current edition when several editions of a policy existed, a check an analyst used to make by reading the version history.

Retrieval was also hybrid. Text embeddings from nomic-embed-text, at 768 dimensions, ran alongside ColPali visual embeddings. The ColPali embeddings read layout, tables and document structure, which a flattened text stream loses. Much DDQ source material sits in fee tables and governance matrices. A retriever that cannot read a table will miss an answer the fund already has.

The saving sits in the repeat questions, where all three choices pay off. For a question the fund has never answered, pair matching has no verified pair to return and recency ranking often has no current document to rank. Hybrid retrieval still searches. It can surface material that looks relevant without answering the question.

The pipeline hands some questions back to a person

We built the March 2026 pipeline to keep part of every questionnaire with a person. The first model call extracted an answer from the located region of a document. A second call asked whether that answer addressed the question and assigned a confidence score between 0 and 1. When the score fell below a threshold, 0.5 by default, the system returned a structured refusal instead of a low-confidence answer.

In our benchmark paper, we credit that second call with the 98.3% proper refusal rate. That call caught cases where retrieval found something that looked right and the extracted answer missed the question. The system sent each refused question back to a person with a statement that it had found no supported answer. A confident paragraph would have hidden that gap.

The refused questions are the interesting ones. A consultant might ask about a structure the fund adopted last quarter. Another might ask how the firm would handle an event that has not happened yet. No document in the set holds a verified answer to either question. The right output is a refusal. The person who knows the answer should then write it. A system that writes a fluent answer to either question has made an unsupported representation to an investor.

Sign-off stays with a person

Selector's Portfolio Manager gave this testimonial: "The system paid for itself in the first quarter. We're now responding to DDQs faster than our competitors with larger teams, and our compliance officer actually trusts the outputs because every claim includes source documents."

The compliance officer in that quote still checks the outputs. The source documents make each check cheap. Because the March 2026 pipeline kept document provenance through retrieval and extraction, every answer traced to a document, a page and the extraction method. The officer checks each claim against its source and skips the hunt for where a sentence came from.

The compliance officer reviews faster and still owns the signature. A completed DDQ is a representation to an existing or prospective investor. No retrieval architecture can carry that responsibility. A firm that hands the signature to software has mispriced what it bought.

We have not measured the remaining hour

Our evidence for "three days to one hour" is an outcome figure from Selector. No instrumented trace of the hour sits behind it. We have no per-question timing to show how the hour divides among review, refused questions and formatting for the consultant's template. We inferred the split from the design of the pipeline our March 2026 benchmark ran on. We have not timed Selector's workflow.

A sceptical reader could argue that template wrangling fills most of the hour and that new questions make up a handful of each questionnaire. If both points hold, the refusal path matters less than we have implied. We cannot settle that with the evidence we hold. Per-question timing across a full questionnaire would settle it, split by matched pairs, generated extractions and refusals. We have not run that measurement.