Your data
For a pilot, files exported from the system you already use. No connection to your accounting system is needed to start.
How Lumiveo learns from your team
The agent proposes, your reviewer decides, and each correction becomes a rule you can read. Nothing that changes an answer is used until a person approves it.
The steps
The agent proposes.
For each exception, the agent gives the likely cause and the fix, and shows the records it used.
Your reviewer decides.
The reviewer accepts the proposal or corrects it, with a one-line reason. This is review work your team already does.
The correction becomes a rule.
A rule is a short written instruction for one client fund, in plain language your team can read and edit.
The rule is tested on past cases.
Lumiveo checks the rule against that fund’s past exceptions to confirm it agrees with what reviewers decided.
A supervisor approves or rejects it.
The supervisor sees the rule, the test result and the cases it would have changed.
The rule is applied and recorded.
The agent follows the rule on later cases. The record shows each decision, the rule used, who approved it and when.
It also learns from its own work
For every case, Lumiveo saves the steps the agent took: what it looked at, what it tried, what failed and what worked. After the case is closed, it keeps what is useful. Your team does nothing extra.
- From a case that went right: it keeps the shortest set of steps that found the answer, and tries them first on the next case of the same type.
- From a case that went wrong: it finds the first wrong step and saves an instruction to avoid it.
Example with made-up data
| What it saw | 30 cash differences at one broker for Fund A. Checking open dividends found the cause in 22 of them. |
| What it changed | For cash differences at this broker, it now checks open dividends first, then currency settlements. |
| Tested | Run again on the same 30 past cases. Same answers, fewer steps. |
How it knows a case went right
- The records matched after the fix.
- The reviewer accepted the proposal with no edit.
- The case was not reopened later.
The agent learns this way only where the result can be checked. Where the answer is a matter of judgment, it learns from your reviewer, and a person approves the rule.
What is recorded for every decision
- The exception and the data the agent read
- What the agent proposed, and why
- What the reviewer decided, and the one-line reason
- The rule that applied, its version, and who approved it and when
- Whether the case was reopened later
What it works with
AI models
Any of the main models. If you already run your own agent software, Lumiveo can add the rules, testing, approval and record around it.
Your model stays unchanged.
The agent improves by following your approved rules and the steps saved from past cases. If you want, your saved corrections can later be used to train a model that belongs to you.
What you see each month
Before a pilot starts, we agree with you on the measures, the starting point and the target. Then we report them for each client fund every month.
| Measure | What it means | Which way it should move |
|---|---|---|
| Accepted with no edit | Share of the agent’s proposals your reviewers accept as they are | Up |
| Repeat mistakes | How often the agent makes a mistake a reviewer has already corrected | Down, toward zero |
| Closed without a person | Share of items the agent closes alone, only for item types you have allowed | Up, within the limits you set |
| Review time per item | Minutes a reviewer spends on each item, from a time sample | Down |
| Time to learn a new client | How long until the agent handles a new fund’s exceptions correctly | Down |
| Rules | Rules proposed, approved, rejected and switched off | Shown, no target |
How the starting point is set
We take three to six months of your past exceptions and how they were resolved.
We run the agent on them without rules to see where it starts.
Your team times a sample of reviews in the first week, so time saved is measured, not estimated.
Example report, one client fund
Example with made-up data
| Month | Items | Accepted with no edit | Repeat mistakes | Review time per item | Rules approved (total) |
|---|---|---|---|---|---|
| Month 1 | 1,180 | 58% | 41 | 6.5 min | 12 |
| Month 2 | 1,240 | 71% | 9 | 4.8 min | 27 |
| Month 3 | 1,205 | 79% | 2 | 3.9 min | 34 |
The report also lists each rule used that month, how often it applied, and any case where a reviewer reversed it.