로그인 데모 요청
Solutions · Innovation

One register, from parallel run to the whole firm

Innovation teams are asked to prove value quickly and then make it stick. ExactCov does both: it reads your own agreements in the first week, runs beside the existing process for a monthly cycle, and goes into production without templates, retraining or a data-science team.

데모 요청 How the parallel run works
IndenturesCredit agreements (LMA, LSTA)Fund documents and ISDA schedulesCompliance certificatesScans and heavily amended documentsExcel trackers you already keep
The problem

Pilots end with a deck. Parallel runs end with a comparison.

Most covenant tooling pilots fail at the same point: the tool needs templates before it produces anything, the security review starts after the pilot ends, and the business case is written from the vendor's numbers rather than yours. ExactCov is designed to be run beside the current process on the same inputs, so the case is made by the disagreements it finds.

No templatesIndentures, LMA and LSTA agreements, scans and amendments are read on arrival. Nothing is configured before the first result.
Your documentsThe register is built from your agreements in week one and checked against the source by your analysts.
Security in parallelSSO, mirrored permissions, residency and zero data retention are in the run from the start, not after it.
MeasuredFirst-pass and second-read accuracy, time to recompute and hours returned per facility, all on your fields.
What the run produces

A register, a comparison and a case

By the end of one monthly cycle you have three things a sponsor can look at.

The register

Every covenant, basket, threshold and deliverable from your agreements, cited to the clause and page.

  • Confirmed at onboarding by your team against the source
  • Definitions resolved to their leaf, cross-references followed
  • Amendments read against the original, changed terms flagged

The comparison

The current tracker and the register, side by side, after one cycle on the same inputs.

  • Deliverables the tracker missed or chased late
  • Tests the two disagreed on, with the citation for each
  • Stale figures the book was carrying as current

The case

Numbers from your run, not from a brochure.

  • Hours per facility against the desk's own baseline
  • Accuracy measured on your fields, before and after the second read
  • What production looks like: users, residency, permissions, rollout by desk
Getting started

Run it in parallel. Don't run a pilot.

Nothing changes for the team. The agreements are loaded once, the notices and certificates that already arrive are copied to a mailbox we watch, and for four weeks both processes run on the same inputs. Then the two outputs are compared on five questions, disagreement by disagreement.

How the parallel run works
Agreementsloaded once NAV noticessame mailbox, copied The team's book spreadsheet, process, people: unchanged nobody learns a new screen weeks 1 to 4 ExactCov register runs on its own, every day, cited stale, chase, tests, limits weeks 1 to 4 same inputs Week 4: compare stale we saw, they missed late, and was it chased tests that disagree: why limits that should move what the book never held every disagreement has a citation on one side, so the comparison takes an afternoon
Nothing changes for the team. The register runs beside them on the same inputs, and after one monthly cycle the two outputs are compared, disagreement by disagreement.
The architecture

Where the model sits, and where it does not

Models propose terms at onboarding, behind a page check and a person, and draft or rank on the way out. Everything in between that produces a number the firm relies on is deterministic code. This is what a security and model-risk review needs to see, and it is the reason the register can be audited.

Where AI belongs in covenant work
ONBOARDING: model proposes, code checks, a person confirms Agreement,certificate, notice Model reads proposes test, basis, quote Code verifies quote is on the page Person confirms flagged items only Register from here on, data EVERY DAY: deterministic, no model in the chain headroomformula on cited inputs due datesfrequency + lag state changesdue, overdue, stale limit proposalspolicy rule on NAV audit trailevery write Model ranks, drafts, summarises morning order, chase email, narrative: advisory reads The model may read, propose, rank and draft. It may not compute a figure the bank relies on, move a date, or change a state.
Models sit at the two ends: proposing terms at onboarding, behind a page check and a person, and drafting or ranking on the way out. Everything in between that produces a number the bank relies on is code.
From our research

How much human review is enough?

The question every rollout asks is how many fields a person must look at. Routing by calibrated confidence, a second read on 15 percent of fields catches 88 percent of the remaining errors and takes accuracy from 97.4 to 99.7 percent. Past 30 percent the reviewer is reading correct answers. The review budget is a setting, and the curve tells you where to put it.

Confidence routing and the second read
Errors caught by a second read, by how many fields you send 1,840 fields, 48 errors on the 97.4% first pass. Green: routed by calibrated confidence. Grey: random sample. 0255075100% 0255075100% fields sent for a second read errors caught random sample: 15% of fields, 15% of errors 5% → 60% 15% → 88%, set at 99.7% 30% → 96% 50% → 98% at 15%: 276 fields, about 45 s each 3.5 reviewer hours per 60 agreements 6 errors reach the register, 5 at depth 3+
Confidence is only useful if it is calibrated, and once it is, the curve bends hard. The knee is around 15 percent of fields, which is where 97.4 becomes 99.7. Past 30 the reviewer is reading correct answers.
Go deeper

Reading for innovation teams

Short pieces from the blog and the measurements behind them from Research.

Blog · 3 min

Run it in parallel. Don't run a pilot.

Same inputs, independent output, nothing changes for the teams processing the book. After four weeks you compare. And why a monitoring tool should be priced per fund, with a floor, never per seat.

Read Blog · 3 min

Where AI belongs in covenant work, and where it must not

Headroom, due dates, exposure and state changes are deterministic code, always. Reading, ranking, flagging and drafting are model-assisted, with a check before anything downstream depends on them. The table, and the reasons.

Read Blog · 3 min

Designing for the auditor, not the dashboard

Every figure traces to a page in a stored document and every write is audited. Second-line functions cannot adopt a tool that produces numbers they cannot defend to internal audit or a regulator, however good the numbers are.

Read Research · 4 min

How much human review is enough? Confidence routing and the second read

Reviewing every field is a second extraction. Reviewing none is a bet. Routing by calibrated confidence, a second read on 15 percent of fields takes our first pass from 97.4 to 99.7 percent. How to set the budget and what to do with the rest.

Read Research · 4 min

Open-weights models against the frontier on covenant extraction

We ran four classes of model over 60 agreements and 1,840 cited fields. The frontier still leads on raw reading, but the gap sits almost entirely in definition chains and citations. With the deterministic pre-pass the production first pass reaches 97.4 percent, and the second read 99.7.

Read Research · 4 min

Deterministic first, generative second: an architecture for cited extraction

A parser builds the section tree, the definitions index and the cross-reference graph before any model sees the document. The model then answers from candidates and must cite a span the validator can check. Why the order matters, the failure it prevents, and how it recomputes every affected test inside nine seconds of data landing.

Read
On the platform

The parts innovation teams use most

Everything on this page runs on the same register as the rest of ExactCov. These are the parts this team lives in.

Platform

Agents

Chase requests what is due on the calendar the documents set, Extract reads what lands, Test recomputes every affected covenant within nine seconds, Brief writes the morning list and Escalate treats silence as an event.

Agents Platform

Vault

Every agreement, amendment, certificate and notice, every version, with the page each number came from. Dependencies between documents tracked, regional residency and zero data retention.

Vault Platform

ExactCov on Claude and ChatGPT

Ask the register from the assistant your team already uses. Permissions mirror ExactCov, every answer carries its citation, and drafts wait for a person to send.

ExactCov on Claude and ChatGPT
Getting started

Run it in parallel. Don't run a pilot.

The documents are loaded once. What already arrives by email is copied to a mailbox we watch. Nothing changes for the team, and after one cycle the two outputs are compared, disagreement by disagreement.

Week 0Agreements in, register confirmed against the source, calendar generated, SSO connected.
Weeks 1 to 4Certificates and notices flow to both. Tests, chases and briefs run on our side only.
Week 4Compare the tracker and the register: missed, late, disagreed, stale. Write the case from the result.
AfterRoll out by desk. Every user who should see the register sees it, with their own permissions.

Run it beside your current process

Bring ten agreements and the certificates you already receive. Four weeks later you have a register, a comparison and a business case written from your own numbers.

데모 요청 Contact us