Payback
8.47 months
Illustrative cost-benefit
Change containment, volume, wages, or build cost. Payback, year-1 ROI, the six scenarios, and the sensitivity tornado recompute from the same formulas as the Python model.
Committed inputs
Pays back inside year 1.
Payback
8.47 months
Year-1 ROI
0.4175
41.75% of build cost.
Year-1 net cash
1,043,640.00
KES (illustrative)
Headline bot scenario. No triage routing and no assumed quality value.
Months to recover build cost from net annual benefit. The mark is 12 months. Select a row to read that case in the scoreboard.
Payback 12 months
Each bar is year-1 net cash for the bot programme at base containment when that one assumption moves by the relative swing (0.20). The other assumptions stay put. Lookup USSD offer moves against the voice-agent share of that class. Base year-1 net cash is 1,043,640.00 KES (illustrative).
Voice-agent handle time
Low input 6.4 → 188,262.40High input 9.6 → 1,899,017.60
Annual contacts
Low input 96000 → 214,912.00High input 144000 → 1,872,368.00
Monthly productive hours
Low input 112 → 1,997,320.00High input 168 → 407,853.33
Voice-agent monthly wage
Low input 64000 → 280,696.00High input 96000 → 1,806,584.00
Lookup USSD containment
Low input 0.56 → 446,648.00High input 0.84 → 1,640,632.00
Bot build cost
Low input 2000000 → 1,543,640.00High input 3000000 → 543,640.00
Lookup USSD offer share
Low input 0.64 → 455,864.00High input 0.85 (clamped) → 1,227,320.00
Wage burden rate
Low input 0.24 → 867,576.00High input 0.36 → 1,219,704.00
Bot annual licence
Low input 480000 → 1,163,640.00High input 720000 → 923,640.00
Transaction USSD containment
Low input 0.32 → 958,888.46High input 0.48 → 1,128,391.54
USSD cost per session
Low input 0.8 → 1,055,145.60High input 1.2 → 1,032,134.40
Year-1 net cash from the low input to the high input Base case
The demo levers cover containment, volume, the voice-agent wage, and bot build cost. Routing shares, channel charges, triage, demand weights, and the remaining inputs are here. Each one writes the field the model reads.
Each class sums to 1. Moving USSD, chat bot, or IVR sets the voice-agent share to the residual when that residual stays inside 0–1. Voice agent means the contact is served with no automation attempt.
Sum 1
Sum 1
Sum 1
Precision and recall are assumptions. They have to be able to sit in one confusion matrix. The quality value is reported on its own and is removed in the last scenario.
A weight divided by the sum of the weights is the demand share. The Bitext training count stays at 1,000 for every intent and never enters the arithmetic.