AI ROI Calculator - Is My Company Ready for Enterprise AI?

A local-first, open-source AI ROI calculator for exploring whether a company is ready for enterprise AI adoption, when AI may improve economics, when it may increase losses, and when financial pressure may be incorrectly attributed to workforce capacity.

Live calculator · Traditional Chinese documentation · Model specification · Agent guide · Contributing · Security and privacy

Created by Protico.io and released under the MIT License.

Why this project exists

Enterprise AI discussions often jump from tool benchmarks to company-wide productivity, revenue, or workforce conclusions. That jump hides several different questions:

• Which tasks are actually exposed to AI?
• How much gross speed remains after review, rework, and AI operations?
• Does released capacity meet real customer demand?
• Does AI revenue retain enough gross margin after inference cost?
• Is a target-margin gap caused by AI, or did it already exist?
• Do first-year cash costs differ from after-rollout annual economics?
• How do labor costs change the payback of globally priced models?

This project makes those assumptions explicit and editable. It is intended for research, internal scenario design, and public review of the equations. It is not a prediction engine and does not provide financial, legal, HR, or workforce advice.

Privacy-first by design

The calculator is a static browser application:

• All calculations run locally in the browser.
• The interface supports English, Traditional Chinese, Japanese, French, and Spanish, with the preference stored only in the local browser.
• The calculator has no application backend or account system, and the model does not intentionally upload company scenario inputs.
• Local state is stored only in browser localStorage.
• PDF reports are generated through the browser print workflow and include the analysis panel plus the tested paths toward positive economics.
• Public filing links are opened only when the user selects them.
• The hosted site at all4.ai is delivered through Cloudflare Pages and loads a Protico frame service for product guidance and user support.
• The hosted site also uses Google Analytics to understand whether the site meets visitor needs. The application does not send calculator input values, scenario outputs, or custom field-level events to Google Analytics.
• These services may receive normal browser request metadata and anything a visitor intentionally shares through that service.

Organizations can clone the repository, remove or block the hosted Protico and Google Analytics scripts, disconnect from the internet, and run the calculator on an internal machine for stricter privacy requirements.

Run locally

Requirements:

• A modern browser
• Python 3 for the static server
• Node.js 18 or later for tests

git clone <repository-url>
cd ai-efficiency-calculator
python3 -m http.server 8000

Open http://127.0.0.1:8000.

With npm scripts:

npm start
npm run build
npm test
npm run check

There are no runtime package dependencies.

Use with a coding agent

An agent can clone and operate the project without a build step:

git clone <repository-url>
cd ai-efficiency-calculator
npm test
npm start

The authoritative model is in calculator.js; interface rendering is in app.js; locale registration and translations are in locale-packs.js and i18n.js. Read docs/AGENT_GUIDE.md before changing an equation, preset, translation, or privacy boundary.

Deploy to GitHub Pages

The included workflow deploys the static application when changes reach main. Before the first deployment, set Settings > Pages > Build and deployment > Source to GitHub Actions in the repository.

Deploy to Cloudflare Pages

The production calculator is hosted at www.all4.ai with the Cloudflare Pages fallback at ai-efficiency-calculator.pages.dev. Build output contains only the browser assets required by the application:

npm run build
npx wrangler pages deploy dist \
  --project-name ai-efficiency-calculator \
  --branch main

Wrangler authentication and permission to deploy to the target Cloudflare account are required. Company scenario inputs remain in the visitor's browser; the deployed site has no application backend. The production deployment also uses Cloudflare delivery services, Google Analytics, and the Protico frame described in the privacy section.

What the calculator models

The current version includes:

• Current and AI-adjusted Revenue per Employee
• Net productivity after review, rework, and AI operations overhead
• Negative-productivity AI scenarios
• Demand-capped AI-attributed revenue
• Internal AI cost, fixed platform cost, transformation cost, and product inference cost
• No-AI versus AI target-margin baselines
• AI-attributed change in modeled workforce pressure
• Evidence-adjusted automatable capacity
• First-year incremental cash flow
• Multi-year NPV and payback period
• Conservative, base, and upside sensitivity cases
• Independently scrollable input and analysis panels on desktop
• App-style mobile input and analysis workspaces with persistent tabs and single-open research sections
• Dynamic adjustment paths that test how a non-positive scenario could improve
• Local PDF reports containing the analysis and tested improvement paths
• AI-native future hiring avoided
• Labor-value payback across editable country benchmarks
• Public-company examples across SaaS, platforms, marketplaces, media, semiconductors, and hardware

Core equations

Let:

• s = gross speed uplift on exposed tasks
• x = task exposure
• a = effective task adoption
• q = output value retention
• h_review, h_rework, h_ops = task-time overhead
• E = employees
• R = current revenue
• RPE = R / E

Gross task time saved:

gross_time_saved = s / (1 + s)

Net task time saved:

net_task_time_saved =
  gross_time_saved
  - review_overhead
  - rework_overhead
  - ai_operations_overhead

Organization time-release rate:

time_release_rate = exposure × adoption × net_task_time_saved

Positive productivity is discounted by output value retention. Negative productivity is retained as a loss rather than hidden by the quality factor.

AI-attributed revenue:

capacity_revenue =
  growth_equivalent_FTE
  × current_RPE
  × growth_allocation
  × revenue_conversion

AI_revenue = min(capacity_revenue, validated_demand_ceiling)

Annual AI cost:

annual_AI_cost =
  employee_AI_cost
  + fixed_platform_cost
  + transformation_cost_amortization
  + product_inference_cost

Evidence-adjusted workforce capacity:

capacity_limit =
  released_FTE
  × structurally_automatable_share
  × non_growth_share
  × (1 - risk_buffer)

The target-margin solver is run twice:

AI_attributed_workforce_pressure =
  required_adjustment_with_AI
  - required_adjustment_without_AI

This separation is critical. A low-margin company may have substantial target-margin pressure before any AI program exists.

The complete definitions, cost treatment, timeline model, and sensitivity transformations are documented in docs/MODEL.md.

Input groups

Company baseline

Revenue, operating profit, gross margin, baseline growth, employees, primary labor market, and fully loaded labor cost.

AI operating assumptions

Task exposure, gross speed uplift, effective adoption, value retention, human review, rework, AI operations, internal AI cost, fixed platform cost, transformation cost, amortization, and product inference cost.

Value realization strategy

Growth allocation, revenue conversion, validated demand ceiling, structurally automatable share, risk buffer, year-one realization, target operating margin, workforce revenue-loss elasticity, transition cost, analysis horizon, and discount rate.

Every input in the application includes a definition, an explanation of how it changes the model, and an example.

The company-scenario bar can be collapsed and automatically minimizes after the user begins scrolling either desktop panel. The research introduction and model equations are available from the question-mark button beside the project title.

Scenario library

The repository includes editable examples for Meta, Alphabet, Amazon, Apple, Cloudflare, TSMC, NVIDIA, Airbnb, Uber, Netflix, Spotify, Salesforce, Tesla, and a synthetic AI-native startup. It also includes synthetic SMB research scenarios for a 75-person retail e-commerce company, a 14-person independent restaurant, and a 750-person Taiwan hardware manufacturer.

Public revenue, operating profit, margin, and employee values are used only where disclosed. Internal AI productivity, cost, adoption, demand, and organization assumptions are model inputs, not claims about those companies.

The SMB scenarios are illustrative operating profiles, not industry averages. Their financial and AI assumptions should be replaced with internal data before interpreting model outputs.

Research basis

The model is informed by evidence showing that AI effects vary materially by task, worker experience, workflow, and measurement method:

• Generative AI at Work, NBER Working Paper 31161
• Navigating the Jagged Technological Frontier, Harvard Business School
• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR
• The Economics of Generative AI, NBER
• Firm Data on AI, Stanford SIEPR

These studies provide context, not universal default effect sizes.

Contributing

The model should improve through transparent disagreement and reproducible evidence. Contributions are welcome for:

• Equation corrections
• Better cost and accounting treatment
• Task- or industry-specific model extensions
• Additional public-company data
• Localization
• Accessibility and visualization
• Tests for edge cases and unintended conclusions

Read CONTRIBUTING.md before opening a pull request. Model changes must include the proposed equation, economic rationale, assumptions, and tests.

Current limitations

• Company-wide averages hide role and task heterogeneity.
• RPE is a comparison metric, not a causal estimate of individual output.
• Public filings cannot identify AI's causal effect.
• Demand, adoption, quality, and review burden usually require internal measurement.
• Conservative/base/upside outputs are deterministic scenarios, not confidence intervals.
• Workforce pressure is a financial stress-test output, not an executable organization recommendation.
• The model does not include tax, financing, working capital, country-specific employment law, or minimum role coverage.

License

MIT License. See LICENSE.

The application footer identifies the project as open source, links to the GitHub repository, credits Protico.io, and discloses that the hosted site uses Cloudflare, Protico, and Google Analytics services. That attribution does not change the MIT permissions granted in the license.