Enterprise AI Value Model

Objective

Turn the framework in Beyond the Productivity Promise into a tool an investment committee can use: estimate the value of an AI initiative across all five dimensions, make every assumption visible, and show which assumptions the case depends on.

Technical architecture

A self-contained calculation engine in TypeScript, with a React interface. It runs in the browser, so no figures leave the user's device.

• Engine. A monthly cash-flow model with three functions: value at full adoption by dimension, a month-by-month projection, and sensitivity.
• Interface. Grouped inputs, headline metrics, a J-curve chart of cumulative cash flow, value by dimension, a tornado chart of sensitivities and the scorecard.
• Export. Inputs and results can be downloaded as JSON, and the scorecard can be printed.

Methodology

Value at full adoption is calculated for six dimensions:

Dimension	Calculation
Efficiency	Users × hours saved per week × 46 weeks × loaded hourly cost × share of time redeployed
Quality	Work items × change in error rate × cost per error (negative if errors rise)
Revenue	Incremental revenue × contribution margin
Decision quality	Decisions improved × value of a better decision
Risk and resilience	Expected annual loss × share avoided
Organizational capability	Cost of the next use case × saving from reusable assets × use cases per year

Monthly projection. Usage-driven value is scaled by adoption, which rises linearly to its target over the ramp period. Capability value accrues as the reusable assets are built. The counterfactual share, the benefit that would have happened anyway, is removed from every dimension. Build cost falls in month one, change cost is spread over the first six months, and run cost is monthly. Cash flows are discounted at the chosen annual rate.

Competitive position is scored from 1 to 5 rather than given a dollar value. If competitors can buy the same tool, any advantage it creates can be competed away, so putting a number on it would overstate the precision.

Sensitivity. Each driver moves 25% down and up while the others are held constant, and drivers are ranked by the resulting swing in NPV.

Scorecard. The six executive questions from the paper are answered from the inputs. For example, a case gets a warning if less than 40% of its value comes from outcomes other than time saved, or if no counterfactual is set.

Evaluation

The engine was tested in three ways:

1. Accounting identities. NPV equals discounted benefit minus discounted cost. Payback is the first month cumulative cash flow turns positive. Value by year sums to total value.
2. Expected behaviour. Higher adoption, more redeployment or lower cost raise NPV. A lower error rate raises quality value, and a higher error rate turns it negative.
3. Shortcut comparison. The model was run against the hours-saved shortcut: every saved hour counted at full salary, no counterfactual, no quality check, efficiency only.

It has not yet been tested against the actual results of real AI investments.

Results

All figures below come from the illustrative example: an AI assistant for a 400-person operations team, $0.9M to build, $0.3M for change and training, and $0.42M a year to run.

Measure	Result
Three-year NPV	$0.40M
Return on cost (present value)	17%
Payback	Month 26
Deepest point of the J-curve	−$1.19M, in month 6
Share of value from time saved	62%
Same case, valued with the hours-saved shortcut	$1.62M NPV

What the example shows:

• The shortcut overstates value about fourfold. Most of the gap comes from assuming every saved hour becomes value. Assuming half is redeployed is more realistic, and even that requires a plan.
• Adoption decides the case. Adoption is the largest sensitivity, at ±$0.59M. Below about 58% adoption the investment loses money. At 40% adoption the NPV falls to −$0.61M.
• Redeployment matters as much as the technology. With only 25% of saved time redeployed, NPV falls to −$0.44M.
• The horizon matters. The same case is worth $1.47M over five years, which is how a J-curve looks: capability investments that seem expensive in year one pay off later. Slowing the adoption ramp from 9 to 18 months turns the three-year NPV slightly negative.
• Quality cuts both ways. If errors rise by one percentage point instead of falling by half a point, NPV falls to $0.25M.

Limitations

• Garbage in, garbage out. The model structures judgment but does not replace it. Decision, risk and capability inputs are hard to estimate, and the tool cannot check them.
• Independent dimensions. The model treats dimensions as independent, while in practice they interact. Better decisions may reduce risk, for example.
• Linear adoption. Real adoption often follows an S-curve, or stalls.
• Single-driver sensitivity. Sensitivity moves one input at a time and does not capture combined downside scenarios.
• No tax, financing or headcount effects. Redeployed time is valued at cost, not at the value of the work it moves to.
• Illustrative figures. The example inputs are not drawn from a real organization.

Code and resources

The full model runs in your browser on this page. A public repository will be linked here once it is published.