PyratechAI Methodology — an AI investment framework for real-world decisions
The PyratechAI AI investment framework is a deterministic, versioned methodology for evaluating AI investments over the long term. Every number the platform shows is a formula you can see and reproduce. This page documents each one in plain language — the same math the free calculator, the Analyzer and the paid toolkit files use. Use it to defend an AI investment framework internally, benchmark AI investment processes against other asset classes, or align on definitions across teams. Artificial intelligence (AI) is not treated as a magic input; it is one of several possible solution shapes evaluated on the same economic and evidence bar. The investment framework is deliberately conservative — realized value is discounted for adoption and value realization; evidence is graded, weighted by decision sensitivity, and penalised for staleness where material.
ROI (Return on Investment)
ROI % = (3-year total benefit − 3-year total cost) ÷ 3-year total cost × 100. Costs include the initial investment and three years of annual operating cost. Benefit is the total realized annual benefit multiplied by three. Example: benefit $1,243,200, cost $436,900, ROI = (1,243,200 − 436,900) ÷ 436,900 × 100 = 184.5%.
Payback period
Payback (months) = Initial investment ÷ (Total annual realized benefit ÷ 12). Answers 'how many months of benefit does it take to earn back the up-front spend?'. Example: $379,000 initial ÷ ($414,400 / 12) = 11.0 months.
Benefit-cost ratio (BCR)
BCR = 3-year total benefit ÷ 3-year total cost. Any BCR above 1 means the initiative more than pays for itself over three years; BCR of 2.85 means every dollar spent returns $2.85 in realized benefit.
Net value (3-year)
3-year net value = 3-year total benefit − 3-year total cost. The absolute dollar figure your organization ends up with after three years, in constant terms. We do not add a hidden discount rate — you can layer NPV on top externally if your finance team requires it.
Adoption × Realization discount
Realized productivity value = Theoretical productivity value × Adoption rate × Realization rate. Theoretical productivity value = Affected employees × Fully-loaded annual cost per employee × (Hours saved per week ÷ 40). Direct cost savings and revenue contribution are also discounted by Adoption × Realization; risk avoidance is discounted by Realization only, because avoided losses don't depend on how many users adopted the tool. This is what makes a business case survive scrutiny — most models fail because they book gross theoretical savings without discounting for the fact that not every user adopts and not every reclaimed hour converts to money.
How to estimate adoption and realization
We deliberately don't publish generic benchmark percentages for adoption or realization rates — your own historical data, if you have it, is a better input than any industry-wide number we could give you. What actually drives Adoption UP: a specific pain the target user already feels, a workflow that integrates into their existing tools, executive sponsorship visible to line managers, active champions inside the target team, and time invested in enablement. What drives Adoption DOWN: a solution the user was not asking for, workflow friction (an extra click, tab or login), lack of ongoing measurement, and adoption reported as 'has access' rather than 'used this week'. What drives Realization gaps: reclaimed time that goes to slack rather than higher-value work, quality-review overhead on AI output, partial adoption inside a workflow (assist without acting), and business processes that block converting time savings into headcount or revenue outcomes. Practical rule: if you don't have historical adoption/realization data from a comparable rollout, estimate conservatively and use the sensitivity view to see how much the decision moves. If the decision only holds at high adoption × high realization, it is a Pilot, not a Fund — regardless of the projected ROI.
Revenue contribution (margin-adjusted)
Revenue contribution = Incremental annual revenue × Gross margin %. Only the margin portion is treated as new value — a $1M revenue lift at 60% gross margin contributes $600K to the business case, not $1M.
Weighted scorecard classification
Weighted total = Σ(score × weight) across dimensions, with complexity and risk dimensions inverted as (6 − score) so that a low complexity/risk score contributes positively. Classifications: ≥ 4.0 = PRIORITIZE, ≥ 3.0 = PILOT, ≥ 2.0 = INVESTIGATE, < 2.0 = DEFER. The AI Readiness scorecard uses AI-READY / EMERGING / DEVELOPING / NASCENT at the same thresholds; the Agent scorecard uses PRODUCTION-READY / PILOT-READY / PROTOTYPE / NOT READY.
Two scores in the Analyzer, kept separate
Investment Attractiveness × Evidence Confidence → Investment Decision. Attractiveness aggregates six dimensions (Business Value, Financial Viability, Technical Feasibility, Execution Readiness, Risk Manageability, Time-to-Value). Evidence Confidence weights each material assumption by how much it moves the outcome — so a well-supported small assumption doesn't rescue a shaky big one. PyratechAI never averages them into a single score.
Sensitivity analysis
Each numeric assumption is stressed ±20% to see how much ROI, payback and net value move. Assumptions that swing the result the most while resting on weak evidence become the biggest business-case risks.
Hard blockers override the score
Data that cannot legally be accessed, no accountable owner, or a specific regulatory barrier will override the numerical scores. When a hard blocker is present, PyratechAI will not recommend FUND regardless of how attractive the initiative looks.
Decision gates
Recommendations map to a stage: Discover → Validate → Pilot → Fund → Scale. The system tells you both the current gate and the next one, plus what would move you between them.
Deterministic, not generative
The calculations and recommendations are 100 % deterministic — the same inputs always produce the same result. We use LLMs only to help you generate first-draft inputs when a field is blank; the LLM never invents financial projections or the decision itself.
What we keep private
The exact weight vectors, precise thresholds and the full internal rule list are intentionally not published — this is our methodology. What you always see: your inputs, the major formulas above, your evidence grades, and the reasons behind your recommendation.
Two-score model, kept separate (v1.1)
Every assessment produces two independent scores: Investment Attractiveness (how strong is the case?) and Evidence Confidence (how well is the case supported?). PyratechAI never averages them into a single number. A high-ROI initiative with weak evidence is a PILOT, not a FUND — and the results view is designed to make that call visible in the first ten seconds. Attractiveness is built from six dimensions: Business Value, Financial Viability, Technical Feasibility, Execution Readiness, Risk Manageability and Time-to-Value. Evidence Confidence weights the strength of evidence behind material assumptions.
Value Realization observation (v1.1)
The projected benefit is classified by how it becomes actual business value: Cost Removed (headcount, vendor spend, opex avoided) · Capacity Redeployed (saved time explicitly reassigned to measurable work) · Throughput Increased (more cases, orders, transactions) · Revenue Generated (margin-adjusted) · Loss / Risk Avoided (measured vs baseline) · Unallocated Productivity (time theoretically saved with no defined mechanism to become cash). Unallocated productivity is NOT treated as equivalent to hard cash savings. A business case that leans heavily on unallocated productivity carries a realization risk which the Results page and Decision Receipt now surface.
Sourcing consideration (v1.1)
PyratechAI surfaces a secondary sourcing signal — BUILD / CUSTOMIZE, BUY, CONFIGURE EXISTING PLATFORM, CONVENTIONAL AUTOMATION or DISCOVERY REQUIRED — derived from feasibility and readiness. It is a hint, not the primary decision. FUND / PILOT / REMEDIATE / DEFER / STOP remain the primary output; sourcing tells buyers where the practical path is likely to be. Not every AI-shaped problem should be solved with AI, and not every AI-suitable problem needs a custom build.
AI Suitability (concept)
A companion signal (deepened further in v1.2) asks whether AI is actually the appropriate solution for the task. Deterministic rules, existing platform capability, or classic automation may be preferable when input variability is low or verifiability is weak. The Analyzer flags these cases rather than always assuming an AI build.
Consequence · Autonomy · Reversibility · Exposure (concept)
The same model performance is not equivalent across use cases. A copilot summary error is not the same as an autonomous financial-action error. Risk Manageability considers the consequence of an incorrect AI output, the level of autonomy granted, whether an action is reversible, and how many customers, employees or transactions are exposed. High autonomy on a high-consequence, low-reversibility decision drives stronger governance requirements — and can contribute to a governance blocker.
Evidence maturity
Every material assumption is graded on an evidence ladder: E0 unsupported assumption · E1 expert/management estimate · E2 external benchmark · E3 internal historical data · E4 internal pilot · E5 production measurement. Higher rungs move Evidence Confidence up. The internal ladder values used are proprietary and not published.
Evidence applicability
Strong evidence from an unrelated context is weaker than modest evidence from the same context. Applicability is captured qualitatively for the productivity assumption as HIGH (same process and target users), MEDIUM (similar but materially different) or LOW (different industry or use case). Applicability is factored into Evidence Confidence without arbitrary universal penalties.
Evidence freshness
Freshness matters contextually — model/API pricing becomes stale in weeks, while stable process volumes remain useful for years. Freshness is captured as CURRENT / AGING / STALE and applied only where the age of evidence could materially affect confidence — currently the external/historical productivity evidence and the implementation-cost estimate. Aging evidence is discounted proportionally; recent, pilot and production evidence is not penalised.
Assumption Ledger
Every completed assessment produces a visible Assumption Ledger — the material assumptions driving the decision, with value, evidence code, freshness where relevant, decision sensitivity, and validation status. Trivial inputs are omitted; only assumptions capable of moving ROI, payback, Evidence Confidence, blocker status or the recommendation are surfaced. FTE labor cost is included and clearly labelled ESTIMATE / PROXY when the user did not provide firm data.
Human review, rework and oversight economics
For AI initiatives that generate outputs a human reviews, corrects or handles exceptions on — copilots, agents, generative drafts, decision-support — PyratechAI captures the percentage of outputs reviewed, average review effort, and rework/exception rate. Gross productivity benefit is netted of review cost, rework cost and ongoing oversight before it enters the ROI calculation. Cases with no material human review are not asked these questions.
Sensitivity analysis
Material assumptions are stressed to see how much they move ROI, payback and the recommendation. Each is classified LOW / MEDIUM / HIGH / VERY HIGH sensitivity. Assumptions that combine high sensitivity with weak evidence become Evidence Risks and are prominently surfaced. Monte Carlo simulation is intentionally not used at this stage — we prefer clarity over false precision.
Hard blockers
Certain conditions override the numerical scores: data cannot be legally accessed, no accountable owner exists, a specific regulatory barrier applies, a measurement baseline is missing. When a hard blocker is present, PyratechAI will not recommend FUND regardless of how attractive the initiative looks. High risk alone is not a blocker; it is an input to the recommendation.
Decision gates
Recommendations map to a stage: DISCOVER → VALIDATE → PILOT → FUND → SCALE. The Analyzer shows the current gate, the next likely gate, and the conditions required to progress between them. The investment decision and the gate are related but not identical: a PILOT decision at the PILOT gate can progress to FUND once specific evidence is validated.
Decision Path — 'What would change the decision?'
Every recommendation is followed by the smallest realistic set of corrective actions most likely to change the outcome. Actions are prioritised: (1) hard blockers, (2) high-sensitivity weak-evidence assumptions, (3) conditions preventing the next gate, (4) material value-realization gaps, (5) risk/governance gaps, (6) execution-readiness gaps, (7) technical/AI-suitability gaps, (8) lower-impact optimisations. This is not 'fix the lowest score' — it is 'what is the minimum change that flips the decision'.
Deterministic, versioned, historical-safe
The scoring and recommendation engine is fully deterministic — same inputs produce the same outputs every time. Every completed assessment is stamped with the methodology version that produced it. Methodology upgrades never silently rewrite past decisions: a June DEFER at v1.0 stays a v1.0 DEFER, and a September reassessment at v1.1 sits alongside it as a new record. Generative AI is not part of the decision engine.
What we intentionally do not publish
The exact scoring weights, precise thresholds, blocker override algorithms, sensitivity formulas, sourcing-heuristic thresholds and the full recommendation-engine rule list are proprietary and not published. What you always see: your inputs, the major formulas above, your evidence grades, the value-realization mix, the sourcing consideration, and the specific reasons behind your recommendation.
What PyratechAI is not
PyratechAI is not a regulatory-compliance certification platform, not an AI inventory / enterprise-architecture repository, not a vendor marketplace, not an automated investment adviser, not an enterprise AI maturity benchmark. It is a decision-support tool for a specific question: does THIS AI initiative have a sufficiently attractive and sufficiently evidenced case to justify the next investment decision?
Not professional advice
PyratechAI provides structured decision-support analysis based on information, assumptions and evidence you supply. Scores, scenarios and recommendations are estimates designed to support, not replace, independent business judgment. PyratechAI does not provide financial, investment, legal, regulatory, accounting or other professional advice and does not guarantee outcomes.