The agents return €5.89M a year after a loss of €2.24M from bad changes.
Of the scenarios tried, Triple the volume shows the largest net value: +€11.78M a year compared with Today. The table is not a recommendation. The conditions for each scenario are listed under the table.
37 of 37 prices are illustrative examples, and none has been researched. Do not quote these figures until the prices are replaced. The measurements come from two synthetic Rootplant files with 400 applications and 250 shipments over about 12 hours. The figures are an estimate from a model and are not accounting figures.
The scenarios, per year
Today
A: Catch it sooner
B: People check denials
C: Triple the volume
Work done in place of people
€8.53M
€8.53M
€8.53M
€25.59M
Extra business
€0
€0
€0
€0
Cost to run the agents
−€52k
−€52k
−€52k
−€155k
People
−€343k
−€343k
−€1.11M
−€1.03M
Loss from bad changes
−€2.24M
−€1.12M
−€673k
−€6.73M
Programme
€0
€0
€0
€0
Net
€5.89M
€7.02M
€6.69M
€17.68M
Against Today
–
+€1.12M
+€801k
+€11.78M
1 year in 10 is worse than this
€4.25M
€6.20M
€6.20M
€12.76M
1 year in 10 is better than this
€7.28M
€7.71M
€7.11M
€21.83M
Loss, 1 year in 20
−€4.46M
−€2.23M
−€1.35M
−€13.37M
Years below zero
0.1%
<0.1%
<0.1%
0.1%
A: Catch it sooner. Bad changes are caught after 3.3 h.
B: People check denials. People check 100% of denials.
C: Triple the volume. Run volume is ×3 today's rate.
Who helps and who harms: Catch it sooner
Hatched areas show work done, solid areas show run cost, and red areas show the share of loss. The 10 agents share the loss because their own output changed together when a feed changed, measured on runs side by side. The lowest net values today are Thermal agent (€251k), Corridor agent (€251k) and Dispatch orchestrator (€254k).
Catching bad changes sooner has the largest single effect on Catch it sooner: +€1.12M.
What the figures are based onpage 2
Measured from the run files
Loan decisions: runs
400 in 11.95 h
Health status
UNHEALTHY
Change record_feed_version 3.3
70 runs
Extra denials compared with the old version
84% (72%–89%)
Still active at file end
6.6 h
Doubling alarm after
3.3 h
Cold-chain dispatch: runs
250 in 11.97 h
Health status
UNHEALTHY
Change record_feed_version 4.1
43 runs
Extra denials compared with the old version
81% (64%–88%)
Still active at file end
6.6 h
Doubling alarm after
3.2 h
Assumed
Bad changes a year per system
4
estimate
Share affecting all systems
25%
estimate
A person catches a wrong outcome
70%
estimate
Working days a year
250
example
A wrong outcome becomes a complaint
0.1
example
Cost of a complaint
150
example
Loan decisions: minutes a person would spend
30
example
Loan decisions: cost of a person's minute
1
example
Loan decisions: minutes to review a run
12
example
Loan decisions: wrong outcome, share of amount
0.03
example
Cold-chain dispatch: minutes a person would spend
20
example
Cold-chain dispatch: cost of a person's minute
1
example
Cold-chain dispatch: minutes to review a run
12
example
Cold-chain dispatch: cost of a wrong outcome
2000
example
How each figure is made
Work done is runs a year × minutes a person would spend × the cost of that minute. The value is shared equally between agents, which is a fair share of the total when every step is needed for a decision. The sharing rule is a convention and never changes the total.
Run cost is the tokens, tool calls and seconds of step time measured per agent, at the prices listed and scaled to one year.
Loss is wrong outcomes per hour × hours until caught × bad changes a year × the cost of one wrong outcome and its complaint. The hourly number is measured against runs that stayed on the old version at the same time. For one incident at the observed exposure, this equals the loss on the Value sheet.
The range comes from 3,000 simulated years with fixed random numbers. Incidents happen at random, and detection time and the cost of a wrong outcome vary. The shared part of the incidents affects every system at once.
Not in this number
Fines: a fine depends on the regulator and the situation, so the sum excludes fines.
Wrong outcomes when nothing changed: the run files cannot show them, so the normal error rate counts as zero.
Wrong approvals: the loss here is only the rise in the share of failed runs after a change.
Customers who leave after a complaint: the sum does not count them.
Proof of cause for the fault path: the agents changed together while other runs stayed on the old version. Only the causal analysis can show a cause chain.
A fault rate that changes over time: the model continues the observed rate unchanged after the 6.6 hours that were actually seen.
The What-If page made this report from Agent Health reports. Every figure can be calculated again from the page and these assumptions.
The What-If page writes this report for the scenarios shown, with illustrative prices. Your browser's print function gives the same two A4 pages.
How to read the figures
Range of years
In the table, 1 year in 10 is worse than the low figure and 1 year in 10 is better than the high figure.
The low figure is P10 and the high figure is P90 of net value over 3,000 simulated years with a fixed seed.
Shared work and shared loss
Each agent gets a fair share of the work done, and the loss is shared by the agents that changed together.
Equal sharing of work done is the Shapley value when every step is required for a decision. Loss is measured under a concurrent control. The fault path is an association, and a cause chain needs the causal analysis.
Spreadsheet
In the prototype
The What-If prototype writes agents-what-if.xlsx, which opens in Excel and other spreadsheet tools. The file holds values, so a controller can continue working with them.
The sheets
Scenarios
Every line of the report's table, per scenario, and what each one changes
PerAgent
Work done, run cost, loss share and net for each agent in each scenario
Assumptions
Every assumption with its value and its source
Measured
What the run files say, per system
NotInThisNumber
What the figures exclude
The workbook from Agent Health's Value sheet does more: its figures are live formulas based on an Inputs sheet.
Presentation
Illustration
Three slides would be written from the report, for the group that prepares the meeting to add to its own presentation. A .pptx would open in PowerPoint and other presentation tools.
1 · The messageThe agents return €5.89M a year after a loss of €2.24M from bad changes.Illustrative prices. The figures are an estimate from a model and are not accounting figures.
2 · The choicesNet per year, by scenarioToday€5.89MA: Catch it sooner€7.02MB: People check denials€6.69MC: Triple the volume€17.68M
3 · Who helps and who harms
The presentation file is not built. Nothing writes this file yet, and the slides above only show what it would hold.
Document
Illustration
A draft decision paper would be written from the report, for the board secretary or CFO to finish. A .docx would open in Word and other word processors.
Decision paper · draft
AI agents: what they return, and what to change
Background
The agents return €5.89M a year after a loss of €2.24M from bad changes.
The options
Option
What changes
Net a year
Against Today
Loss, 1 year in 20
A: Catch it sooner
Bad changes are caught after 3.3 h.
€7.02M
+€1.12M
−€2.23M
B: People check denials
People check 100% of denials.
€6.69M
+€801k
−€1.35M
C: Triple the volume
Run volume is ×3 today's rate.
€17.68M
+€11.78M
−€13.37M
Proposed decision
The group writes the proposed decision. The report compares the options and does not recommend one.
Basis
The figures are an estimate from a model and are not accounting figures. The prices are illustrative and must be replaced before these figures are quoted. Fines, wrong approvals and customers who leave after a complaint are not in the number.
The document file is not built. Nothing writes this file yet. The options would follow the scenarios, and a person would still write the proposed decision.
Power BI
Illustration
The figures would be a live data source, so the report in Power BI would follow the scenarios. Nobody would need to paste them in before each meeting.
How it would connect
Open the connection file, or paste the query into Power BI's Advanced Editor.
Sign in with a viewer token, kept as a Power Query parameter and never in a shared file.
Refresh the data before the meeting. Every figure shows its scenario and its assumptions.
let
BaseUrl = "https://acme.augalabs.com",
Doc = Json.Document(Web.Contents(BaseUrl,
[RelativePath = "api/whatif/scenarios", Query = [dataset = "credit-nov"]]))
in
Table.FromRecords(Doc[rows])
What the group would see
Today€5.89M
A: Catch it sooner€7.02M
B: People check denials€6.69M
C: Triple the volume€17.68M
What exists today
Agent Health has this on the branch feat/agent-health. GET /api/health/value/powerbi.pq gives a Power Query with every table, and powerbi.pbids gives a connection file. /api/whatif/scenarios above is a proposal. What-If runs in the browser and keeps its scenarios there, so they must first be saved on the tenant.
Board portal
No integration
The board report does not need a connection to reach the board. The report is a two-page A4 PDF, and it goes into the meeting's documents like the group's other papers.
How it would work
Replace the illustrative prices, so that the report's warning no longer applies.
Open the report as PDF and save it.
Upload it to the meeting in your portal, with the decision paper beside it.
Portals this is for
The portals include Diligent Boards, Admincontrol, OnBoard, BoardEffect and Boardeaser.
The report has not been tried in any of these portals. A direct integration waits until a customer organisation asks for one.