Savings register: from an insight to a decision to a checked result
An insight that says "you could save 400 dollars a week" is worth nothing until someone acts and someone checks. The savings register is that loop, measured in money. It is on the Cost page, under "Savings register", and builds on the follow-ups on findings.
In the product
Product capture, 15–16 November 2026. Open the image for full size. Counts and dates are examples from the captured workspace.
The Savings register sits below Cost per outcome. It separates an estimated opportunity, a decision and a checked result. This captured workspace has no populated savings entries; the screenshot does not demonstrate realised savings.
An entry
Any avoidable-cost item can become an entry for one agent: an item from the cost of failure and repetition, from declared tools (such as calls to a tool not on the list), or from where the tokens go. An entry records:
- What was found, with the estimated saving a week when it was entered.
- Who owns it.
- What was decided:
- Act: something was done. It needs a note on what, and a window of 7 to 60 days (14 by default). Provantix reads the result after the window.
- Accept: "we need this" (for example, "we need the retries"). It needs a reason. An accepted insight is not raised again for that agent until its weekly cost has grown past the margin you set (50% by default).
- Dismiss: it needs a reason. A dismissed insight is also dismissed in the section it came from.
Only an admin can create or cancel an entry. Everyone can read and export the register. Entries are never deleted: a cancelled entry stays, with who cancelled it. Each change is audited.
The result
For a decision made on day D with a window of W days, "before" is the W days up to D (not D itself, which the decision split) and "after" is the W days from D+1. Provantix reads the result once D+W is over, for the same agent, the same item and periods of the same length. It reads the item's cost again by the same method that raised it. Each result is one of:
| Result | Meaning |
|---|---|
| Fell | The item's cost fell in total, and per run and per 1,000 model calls by at least 20%. Only this counts as a checked saving. |
| Did not fall | The cost did not fall by 20% per unit. If the total fell only because the agent did less, the result says "lower traffic only" and claims nothing. |
| Not enough data | Fewer than 5 runs of the agent on a side, or the item cost almost nothing before, so there is nothing to see fall. |
| Not comparable | The agent's volume (more than double or less than half), its models (a model with a tenth or more of its calls appeared or went), or the configuration it reported changed, so before and after are different populations. No saving is claimed. This is the same rule as for follow-ups on quality findings. |
A checked saving is stated per week: (cost per run before − cost per run after) × the runs after, scaled to a week. It is per unit so that a fall from lower traffic is not counted. Both unit figures and the totals are shown.
The causation caveat applies to every result. The cost fell after the decision; that does not show the decision caused it.
The view
- Open, by estimated saving: current suggestions from where-the-tokens-go with an estimated saving and no entry.
- Decided, awaiting a result.
- Results, with the reason for each.
- Accepted or dismissed.
- Two totals, kept apart: the saving checked (entries whose result is "fell") and the saving estimated (open suggestions and entries awaiting a result). They are never added together: one is read from the evidence afterwards and the other is a projection.
Export and the cost statement
GET /api/savings.csv exports every entry, one line each, with estimated and checked figures in different columns. GET /api/savings/statement?from=&to= returns the entries decided or assessed in a period, with the caveat, as the section the period's cost statement will carry. That statement is not built yet, so the section is available but not yet placed in one.
API
GET /api/savingsis the view.GET /api/savings/entries?agent=lists entries.POST /api/savings/entrieswithsource(cost_line,tool_itemorinsight),item,agent_id,subject(for an item about one model),decision,note,owner,window_daysand, for accept,accept_margin(admin).POST /api/savings/entries/<id>/cancel(admin).POST /api/savings/assessreads any entry whose window is over now (admin); it also runs hourly.
Limits
- Estimates follow your price table.
- A change you made on purpose, such as moving to a cheaper model, makes before and after different populations by the rule above, so it is not comparable and no saving is claimed. Compare the costs by hand in that case.
- Only the agent's own volume is controlled for: a change in what the agent is asked to do is not visible to Provantix.
