AI can read a credit agreement in seconds. It can compare thousands of positions across systems and draft the entry that corrects a difference. An investment firm still needs a person who is accountable for the result.
On ReconoSense, anything the AI produces is a proposal. A person with the authority to approve it decides whether it goes into the record.
Where agents work
Agents on the platform do three kinds of work.
Extraction. Agents read agreements, amendments and notices and note the section each term comes from. A reviewer approves the terms, and anything extracted with low confidence goes to the reviewer before it is used anywhere.
Matching. Agents propose matches between sources. An analyst approves them.
Resolution. Agents draft a fix for each difference, naming the source that governs and the section that decides it. An accountant approves the fix.
In each case the decision belongs to a person.
Where a second person is required
Some actions require two people. A payment is prepared from the approved record, and a second person releases it. The person who initiates an action cannot approve it. This is segregation of duties, and the platform enforces it on every approval.
What is recorded
The log keeps the agent's proposal, the evidence it cited, who approved it and when, and nobody can edit it afterwards. An auditor or examiner can trace a number back to its source, its clause and its approver.
The record also shows what the AI did and did not do. A firm can see how many proposals its reviewers accepted and how many they corrected.
How the models improve
Reviewer corrections train the firm's own models. When a reviewer corrects an extracted term, the correction becomes an example for that firm's models on the document formats that firm receives.
A firm's data trains that firm's models only. There is no cross-tenant training or benchmarking.
Where the models run
The models run inside the firm's own environment, and firm data never goes to an outside AI company. The platform runs in the firm's own cloud, on premises or in a dedicated single-tenant environment hosted by ReconoSense.
Running without AI
Each AI feature has its own switch. The reconciliation rules don't depend on the AI, so the platform runs with it switched off. People then do the extraction and drafting themselves, under the same approvals and in the same record.
That means a firm can turn the AI on one feature at a time, see how each performs and switch any of them off without disrupting its work.
Why this design
An investment firm answers to investors, lenders, auditors and regulators for every number it reports. Accountability rests with people. A system that lets a model write directly to the record removes the point at which a person takes responsibility.
Keeping approval with a named person preserves that point. The record shows the agent's proposal and the person's approval.