State-of-the-art techniques applied to every commercial system we build.

AI earns its place in a commercial system where the structure of the problem fits — forecasting, segmentation, scenario modelling, anomaly detection across high-dimensional data. The system uses AI methods where they add value to the underlying decision; it routes around them where they don’t.

The discipline matters because most consumer businesses are now being pitched AI-led commercial advisory by firms whose distinctive capability is the model, not the operating instinct behind it. The practice runs the opposite balance: AI methods earn their place only inside a commercial system that can defend every decision it makes.

Most commercial forecasting is single-point. A pricing model produces a number; a partnership-value calculation produces a number; a market-entry size produces a number. The number is wrong — that’s the easy part. More importantly, the number doesn’t carry its own confidence. The system has no honest read on how seriously to take it.

Bayesian decision systems do two things differently. They produce a distribution rather than a point. And they make the prior — what the system already believed before the data — visible and explicit, so it can be argued over rather than smuggled in. The output is a 70%-confident answer with named bands, which is more useful than a 100%-confident answer with hidden ones.

Inside Design-stage tooling — pricing models, forecasting tools, partnership-value frameworks — Bayesian methods give the system a defensible confidence interval to act on. Inside Embed-stage governance, kill criteria are anchored on those confidence intervals rather than thresholds drawn out of the air.

Most data-driven decisions in a commercial function are built on correlations. Customers who bought X also bought Y. Regions that ran promotion A also saw uplift B. Statistical and machine-learning methods most easily produce correlations. Commercial trading decisions need something stronger.

A pricing decision needs to know: if we raise the price of X, what causes what to happen to demand? An incrementality decision needs to know: did the campaign cause the uplift, or did it coincide with the channel’s natural cycle? A retention decision needs to know: did the loyalty programme cause the repeat purchase, or did the customers who joined the programme already have higher repeat propensity?

The practice uses causal-inference methods — Judea Pearl’s framework of directed acyclic graphs, do-calculus and counterfactual reasoning — to build commercial systems that distinguish cause from coincidence. Inside engagements this shows up as cohort design with proper controls, incrementality testing built into promotional architecture, and pricing models that account for confounding variables the trading function would otherwise miss.

The practical effect: the system stops making decisions that look right against the data but produce outcomes that don’t materialise.

Many commercial decisions look like optimisation problems but are actually combinatorial — there is no smooth gradient to follow; there are constraints, dependencies, and a small number of configurations that satisfy them all. Range planning under shelf-space and assortment constraints. Promo-calendar planning under brand-fairness and stock constraints. Market-entry sequencing under capital and partnership-availability constraints. Supplier-mix design under category-margin and exclusivity constraints.

For problems with this structure, the practice uses logical and scenario-modelling methods — answer set programming, constraint satisfaction, mixed-integer optimisation — where the right tool finds every configuration that satisfies all the constraints simultaneously, then ranks them by commercial objective.

This is not where AI shows up at scale. It is where rigorous logical reasoning, expressed declaratively, lets a commercial system explore a problem space too large for spreadsheets and too structured for machine learning.

The practical effect: the system can answer questions like “what range can we stock under this margin, shelf and brand-mix constraint?” or “in what order should we sequence these four market entries given partnership availability and capital limits?” — and produce all feasible answers rather than one heuristic guess.

The library is the visible core of the WM IP estate. It is a Streamlit-based forecasting tool whose core is a polynomial-surrogate prediction recipe — delay embeddings, per-window standardisation, whitened PCA, an ND tensor-product Chebyshev polynomial fit by a minimum-derivative-norm KKT solve. The pipeline is ported from a founding advisor’s day-trading research and is now part of the practice’s IP estate.

Inside engagements the library underpins LTV dashboards, partnership-value scenario models, channel-economics forecasters, and any Design-stage tooling that needs a probabilistic forecast with honest uncertainty bands.

Hold-out benchmarks on public data

The library has been benchmarked against random-walk and seasonal-naive baselines on standard public time series. Lower ratios are better; ratio is the polynomial-surrogate RMSE divided by the random-walk RMSE on the held-out test set.

On a synthetic test (four sinusoidal series, length 400), the surrogate beats random walk by roughly 7× on RMSE and the 95% confidence band achieves nominal coverage. Macroeconomic data with structurally unrelated columns is a known weak point — for that case the library auto-detects and switches to per-column fit, restoring competitive accuracy.

The analytical methods above are necessary but not sufficient. A commercial system also needs the right routing of its own outputs: RACI for the commercial functions, decision rights mapped to the consequential decisions, named review meetings where the system gets re-tested, kill criteria the system can defend without having to argue them every time.

Decision-rights design is a systemic problem, not a people problem. An analytically-correct decision degrades when the routing is wrong — when it lands at the wrong cadence, sits inside a committee that can’t act, or runs against incentive structures that quietly veto it. The practice treats decision-rights as a part of the commercial system itself; the Embed stage carries this work explicitly, because no model survives a misaligned decision-rights structure.

The quantitative half of the work is what makes governance defensible: kill criteria anchored on named confidence intervals, review cadences keyed to forecast accuracy, and a documented basis for any consequential decision the system produces — defensible against the question every board eventually asks: how would we have known if this was wrong?

Off-the-shelf competitor-pricing tools either cover one source poorly or aren’t tuned to the questions a commercial system actually needs answers to: are they cutting price on these SKUs because of stock pressure, or because of a planned promotion? Is their range pull deliberate or a stock-out? What’s their take-rate doing on the seller side? The practice builds custom scrapers for the specific competitors and metrics that matter inside a given commercial system.

Scraping is treated as commercial infrastructure: the goal isn’t more data, it’s tighter feedback loops on the four or five competitor moves that actually require a response.