AI-Threshold methodology & sources

AI-Threshold is a model, not an oracle. Its structure rests on four literatures; the numeric calibrations and the AI-obsolescence framing are our judgment, labelled as such.

The economics — cost, payback, break-even, NPV

Status-quo cost vs. a full Total Cost of Ownership for the automated option (build, subscription, maintenance, and the residual human work). Break-even frequency tells you how often you must do the task before it ever pays off.

• Brealey, Myers & Allen — Principles of Corporate Finance — NPV is the correct rule; simple payback ignores the time value of money and post-payback cash flows.
• Horngren, Datar & Rajan — Cost Accounting — Cost-Volume-Profit / break-even analysis.
• Ellram (1993), Int. J. Purchasing & Materials Mgmt — Total Cost of Ownership — recurring costs dominate, not just upfront.
• Coase (1937) · Williamson (1981) — Make-vs-buy as a transaction-cost decision (frequency + specificity).
• Kaplan (1986), HBR — Justify automation with better DCF and a realistic baseline — not faith.

Residual human work — the “80% automated, 20% review”

Automation is rarely total. The leftover slice is treated as a first-class input, costed at the full rate, recurring monthly.

• Bainbridge (1983), Automatica — “Ironies of Automation” — The residual is selectively the hardest part of the job, not a smaller copy of it.
• Parasuraman, Sheridan & Wickens (2000), IEEE — Automation is a continuum of levels, not a binary.

Why freed time is discounted

Saved minutes only have value if you can convert them. The model discounts freed time by how it’s used (billable vs. relief) and how it arrives (one block vs. scattered).

• Becker (1965), Economic Journal — The opportunity cost of time; freed time is only worth what you redeploy it to.
• De Langhe & Puntoni (2016), J. Marketing Research — “Time-savings bias” — people systematically overestimate time saved.
• Leroy (2009), OBHDP — “Attention residue” — fragmented time carries a switching tax.

Setup overruns & uncertainty — the Monte Carlo engine

Uncertain inputs (setup time above all) become ranges, sampled thousands of times to report the probability of paying back in time, plus a tornado of what matters most.

• Kahneman & Tversky (1979) · Buehler et al. (1994) — The planning fallacy — estimates from the “inside view” run systematically low.
• Flyvbjerg (2002, 2006) — Overruns are the norm; fix via reference-class forecasting.
• Malcolm et al. (1959), Operations Research — PERT / three-point (min, likely, max) estimation.
• Clemen & Reilly · Eschenbach (1992) — Tornado / sensitivity analysis — what to nail down first.

Task half-life / AI-obsolescence — our synthesis

If an off-the-shelf AI feature will likely do the task end-to-end before a custom build pays back, the verdict says “don’t.” This framing is ours, built on top of task-based automation economics — no single paper makes the make-vs-buy claim.

• Autor, Levy & Murnane (2003) · Acemoglu & Restrepo (2018, 2022) — Routine/codifiable tasks are exactly the automatable ones — and the easiest for a vendor to productize.
• Frey & Osborne (2017) · Brynjolfsson, Mitchell & Rock (2018) — Per-task susceptibility to computerisation / machine learning.
• Eloundou et al. (2023), “GPTs are GPTs” — The gap between bare-LLM and LLM-plus-software exposure is the obsolescence risk.