Watch what actually happens when an analyst hands a memo to an investment committee. Nobody reads the headline IRR and moves to approve. Someone asks where the exit cap came from. Someone else asks whether the rent roll reflects the March amendment. The analyst walks the room back through the model, the lease file, and the comp set until the committee is satisfied that the number has a pedigree. The meeting was never about the number. It was about the trail.
That behavior is not culture, it is fiduciary law expressing itself as a meeting. Institutional capital is regulated, audited, and answerable to LPs, which means every material figure must be defensible after the fact, sometimes years after, to someone hostile. The industry's work products, the IC memo, the valuation pack, the covenant certificate, the investor letter, are not collections of numbers. They are collections of claims, each carrying an implicit warranty: trace me and I will hold.
What an institutional number actually is
Strip any figure out of an institutional artifact and it decomposes into three parts: the value, the computation that produced it, and the sources the computation consumed. A DSCR of 1.24x is really "NOI of $7.76M from these GL lines and this rent roll, divided by debt service of $6.25M from clause 6.2 of this loan agreement." The value is the smallest part. The other two parts are what make it usable in front of a committee, an auditor, or a lender.
This is why "the AI got the right answer" is a weaker claim than it sounds. A right answer with no trail still has to be re-derived by a human before it can be used, which means the AI saved the typing but not the work. The expensive part of analyst time was never arithmetic. It was establishing, and being able to demonstrate, where everything came from.
Why provenance cannot be bolted on
The obvious fix, and the one most AI products ship, is to generate the answer and then ask the model to explain it. Attach the explanation as a footnote and call it a citation. The problem is that a language model's explanation is generated the same way its answer was: it is the most plausible-sounding justification, produced after the fact, with no structural connection to the computation. Two runs can produce the same answer with different explanations, or worse, different answers with equally confident explanations.
Auditors have a phrase for evidence produced after the conclusion: they do not accept it. A citation that a model wrote to accompany its own output is a second opinion from the same witness. It reads like provenance, but nothing enforces that the cited clause was actually the input that determined the figure. When the memo is challenged, the trail evaporates exactly when it is needed.
Take any figure in the output and ask the system to show its inputs, then change one of those inputs and ask again. If the figure moves the way the math says it should, and the trail updates with it, the provenance is structural. If the system re-explains rather than re-computes, the provenance is decorative.
Citation by construction
The alternative is to build outputs out of their sources rather than annotating them afterward. That requires two commitments that go deeper than a feature list. First, a knowledge graph: every document, lease clause, GL line, and assumption lands as a node with identity, so that "the rent for suite 1204" is a specific thing that can be pointed at, versioned, and superseded by an amendment. Second, a deterministic engine: every financial figure is computed by an auditable calculation over those nodes, never generated as prose, so the same inputs produce the same output every time.
On that foundation, citation stops being an act of diligence and becomes a property of the system. The IC memo's exit value cites the cap rate assumption and the stabilized NOI because it was literally computed from them. The variance narrative cites the two move-outs because the bridge was built from the transactions. Nothing needs to remember to cite. The trail is the structure the number was built on, exposed.
2,148 cells, same answer every run.Why the trail is the moat
Here is the commercial consequence, and the reason this is a thesis rather than a compliance note. The bottleneck in institutional workflows is not production, it is review. Drafts were never scarce; senior time to verify them was. When the artifact carries its own audit trail, review changes shape: instead of re-deriving the work to trust it, the reviewer spot-checks the trail and applies judgment to the decisions. That is where the hours actually come back, in the most expensive seats in the firm.
- Adoption follows defensibility. The first AI-drafted memo that survives a hostile IC question does more for adoption than any training session.
- The trail compounds. Every cited artifact enriches the graph that the next artifact is built on. Decorative citations compound nothing.
- Regulation is a tailwind. Auditors and LPs are converging on the same question: show me how the machine got this. Firms with structural provenance answer in a click.
Built AI was built on this thesis: a knowledge graph so every source has identity, a deterministic engine so every figure is computed rather than generated, and citation by construction so every artifact can be walked back to its inputs, by an IC, an auditor, or an LP, years later. The audit trail is not a feature we added. It is the product. To see a cited artifact built from your own documents, explore how the platform works or book a walkthrough.