RATIO AI Audit Practice · Research Study · Proof of Concept · Hybrid Cognition

Beyond Data Cleanup: Can AI Detect Customer Departure Before It Happens? A Five-Architecture Empirical Study.

Milestone Planning and Research, Inc. 175 synthetic B2B customers · 156 weeks · seed=42 Five architectures · locked economic formula · CSV-verified June 2026 Proof of Concept — research-grade environment Bayesian belief network — machine forecast + human judgment
Study Overview

The business problem. Acquiring a new B2B customer costs five to seven times more than retaining an existing one. Customer churn costs U.S. businesses an estimated $136 billion annually. To make that concrete: a mid-sized B2B operation carrying a $55 million annual-revenue customer portfolio — the scale modeled in this study — can lose several million dollars a year to departures that were detectable months in advance, simply because no one was watching for the right signals. Yet most organizations either deploy no AI for departure detection, or spend years and millions of dollars cleaning and labeling data before any model runs. This study demonstrates a governance-first alternative: a supervised AI departure detector built without a single historical departure label — trained entirely on synthetic behavioral episodes generated from domain expertise. Behavioral surveillance and anomaly detection are the ideal entry point for organizations new to AI: low data requirements, interpretable outputs, measurable economic value, and no multi-year data preparation program required.

What was built. Five model architectures were evaluated against 175 synthetic B2B customers observed over 156 weeks. Every model was scored against the same locked ground truth using an economic framework — Economic Detection Potential, False Positive Penalty, and Net Economic Value — rather than statistical accuracy alone. The result: a corrected Sequential Fusion architecture that combines a Markov Switching Model (Stata), an LSTM trajectory challenger (MATLAB), and a Fourier amplitude detector (Python) to achieve +$3,323k net economic value with only $278k in false positive costs.

How it was built. This study was completed in less than one week of elapsed effort. Claude (Anthropic) wrote all data pipelines, feature engineering scripts, the scoring harness, the GAP branch detector, and the corrected fusion architecture. ChatGPT (OpenAI) implemented the MATLAB Hidden Markov Model and LSTM challenger. The human author served as quality manager throughout — defining architecture, demanding proof for every claim, cross-checking every number against source CSV files, and running OCC 5C governance reviews at each session transition. When AI systems made errors — fabricated results, logic inversions, double-counting — the author identified them through governance challenge and corrected them. The OCC framework is what made the speed possible without sacrificing accuracy. When AI writes the code, schedule risk nearly disappears — quality risk becomes the binding constraint. AI-era project management monitors evidence quality, not milestone dates. This is the paradigm shift.

Traditional machine learning demands clean predictor variables — imputed, normalized, de-duplicated, and validated across systems before a single model runs. Organizations routinely spend millions and years on data engineering before AI development begins. This study demonstrates a different path: the only input required was a clean weekly order history. No CRM integration, no labeled churn outcomes, no external enrichment. Domain expertise converted into synthetic behavioral episodes replaced years of historical label collection. The entire project was completed in less than one week of elapsed effort using AI-assisted development governed by the OCC quality framework.

A full working paper documenting the methodology, governance findings, and twenty lessons learned is available for download below.

Why this is a hybrid cognition study. The five-architecture comparison is the visible half of this exercise. The more general finding sits at the end of it: no single architecture, however well it scored, was sufficient on its own — and the gap the best model left open was not closable by a better model. It was closable by a person who knew something the data had not recorded yet. So the study closes by combining the two in a Bayesian belief network: the machine's forecast and the account team's field judgment enter as separate evidence nodes over a shared prior, each expressed in decibans, and sum into a single auditable posterior. Neither layer overrides the other and both are logged. That structure — machine plus human judgment, combined under one arithmetic rather than one deferring to the other — is what this practice means by hybrid cognition, and it is the thread running through the rest of this site's work. The full treatment begins under Hybrid cognition in practice below.

Summary metrics
Best net economic value
+$3,323k
Sequential Fusion v2
Two-stage fusion net value
+$3,323k
MSM surveillance · LSTM filter
Best discrimination (AUC)
0.858
LSTM + Fourier hybrid
Lowest false positive penalty
$278k
Sequential Fusion v2
Lifecycle stage
POC Complete
Environment setup → system test → UAT → deployment
Belief-network evidence nodes
6
2 empirically calibrated · 4 judgment-only, each logged with rationale
Validation coverage
175
All customers verified · zero discrepancy vs dashboard
Five-architecture comparison — all 175 customers · locked formula · CSV-verified
Architecture AUCTPTNFPFN EDPFPPNetPrecRecall
Markov Switching ModelStage123 · max WoE v2 0.526015441133 $3,405k$2,748k+$656k0.0980.833
Hidden Markov ModelStage123 · state estimation 0.572215251323 $3,571k$3,690k−$119k0.1020.833
Isolation Forest best net Stage123 · anomaly detection 0.439115221353 $3,866k$3,680k+$186k0.1000.833
LSTM + Fourier Hybrid challenger Trajectory learning · routed architecture 0.858116138182 $1,567k$280k+$1,288k0.4710.889
Sequential Fusion v2 best precision MSM watchlist · LSTM confirmation filter 16140162 $3,601k$278k+$3,323k0.5000.889
Visual analysis
Net economic value ($k) by architecture
Net values across five architectures.
Confusion matrix components by architecture
Confusion matrix components.
Precision vs recall — operating points
Precision recall scatter.
Economic Detection Potential vs False Positive Penalty ($k)
EDP vs FPP grouped bar.
Sequential Fusion v2 — three-component architecture
Component 1 · Surveillance
MSM Stage123
Stata 18 · max post-burnin WoE · 2 dB threshold
Scores all 94 CONTINUOUS customers. Broad detection net — all 94 enter the watchlist at 2 dB. Sets detection timing and action intensity for confirmed customers.
Component 2 · Confirmation
LSTM+Fourier Hybrid
MATLAB R2025b + Python · trained on synthetic episodes
Filters the 94-customer watchlist. Cumulative WoE ≥ 400 dB → confirm LEAVING (8 customers). Below threshold → remove, reclassify STABLE (86 customers).
Component 3 · Independent
GAP Fourier Branch
Python 3.12 · amplitude dropout detection
Handles all 81 GAP customers independently. 80 PERIODIC: Fourier amplitude dropout scoring. 1 IDIOSYNCRATIC (C065): ABSTAIN · $0 cost · IOR monitoring only.
All three components run in sequence. CONTINUOUS customers pass through Components 1 and 2. GAP customers go directly to Component 3. Ground truth was withheld from all components during construction. Results verified against locked CSV decision files using the OCC CSV Auditor.
Routing and confirmation flow
Stage 1 — Broad surveillance net
All 175 customersMSM (Stata) max post-burnin WoE scoring · 2 dB threshold
WoE ≥ threshold94 CONTINUOUS customers enter watchlist (all flagged at 2 dB)
Not flaggedSTABLE · no further processing
Stage 2 — Precision confirmation filter
CONTINUOUS watchlistedLSTM cum WoE ≥ 400 dB → confirm LEAVING
CONTINUOUS watchlistedLSTM cum WoE < 400 dB → remove · STABLE
GAP PERIODICFourier amplitude branch · independent
GAP IDIOSYNCRATICIOR interval · ABSTAIN · $0 opportunity cost
Key findings
Architecture choice is an economic decision
Under the locked formula, net economic values range from −$119k to +$3,323k across five architectures. AUC rankings and economic rankings diverge sharply. The LSTM+Fourier Hybrid has the highest AUC (0.858) but not the highest net value (+$1,288k). Sequential Fusion v2 achieves the highest net value (+$3,323k) with a precision of 0.500 — catching the right customers and not wasting intervention budget on false alarms.
Filter, don't escalate — the fusion logic that matters
Fusion v1 used the LSTM to escalate Stage123 alarms — inheriting all 138 false positives. Fusion v2 uses it to filter: Stage123 alarms not confirmed by LSTM are removed from the watchlist. FPP drops from $1,953k to $278k. Net improves by $1.3M. The direction of the LSTM's role is the entire difference between a model that compounds errors and one that corrects them.
Governed acceleration — one week, five models
Claude wrote all pipelines and the scoring harness. ChatGPT implemented the MATLAB HMM and LSTM. The human architect provided direction and governance challenge. When AI systems made errors — fabricated results, double-counted costs, inverted logic — the author identified them through OCC 5C structured review. The OCC framework is what made the speed possible without sacrificing accuracy.
What makes fusion v2 work — exactly
MSM (Stata) flags all 94 CONTINUOUS customers at 2 dB — a broad net by design. LSTM evaluates each independently: 8 reach cumulative WoE ≥ 400 dB and are confirmed LEAVING; 86 fall below and are reclassified STABLE. The 400 dB threshold sits cleanly between the stable 95th percentile (119.7 dB) and the leaving minimum — a 280 dB separation gap. Early detection timing from the MSM is preserved, so EDP credit for catching customers early accrues correctly.
Project management paradigm shift
When AI writes the code, schedule risk nearly disappears. Quality risk does not — it becomes the binding constraint. Five architectures were built in under one week, but the majority of human effort went to quality management: demanding proof for every claim, verifying that numbers foot, documenting governance events. Evidence quality replaces schedule adherence as the primary project management challenge. The OCC 5C framework is the instrument of that quality management.
Validation coverage — 175 customers, zero discrepancy
The Customer Validation Workbook (178 tabs) provides step-by-step verification for every customer: registry facts, model decisions across all architectures, step-by-step economic calculation, and weekly score trace. The CROSS-VALIDATION tab confirms zero difference between bottom-up tab summations and locked dashboard values — EDP, FPP, Net, and all confusion matrix cells — across MSM, LSTM+Fourier Hybrid, and Fusion v2. No faith required: every number traces to a source CSV.
Project lifecycle: POC → environment setup → system test → UAT → deployment
POC designation. This study is designated a Proof of Concept. The models work, the governance works, the economics are verified, and every number traces to a source CSV. The POC qualifier applies to the environment — MATLAB desktop, Stata do-file, Python scripts running on a research workstation — not to the evidence quality. The validation workbook confirmed zero discrepancy across all 175 customers in the research environment. That is POC-level system testing: it certifies the arithmetic is correct in the environment in which the code was written. It does not certify the system is ready for operational use.

Environment setup is next. Before any formal system testing or UAT can occur, the software and data must move to appropriate environments. Platform, licensing, and infrastructure decisions for each model component will be evaluated in this phase in consultation with IT and the relevant platform vendors. Representative data — either a sanitized subset of real customer records or a recalibrated synthetic dataset — must be staged in the target environment. IT approval of the platform stack is a prerequisite, not a parallel track.

System testing in the target environment follows. Once the stack is deployed on approved infrastructure with representative data, formal system testing re-verifies that models produce correct outputs in the new environment — confirming the migration introduced no errors, the scoring harness operates correctly, and audit file generation is intact.

UAT follows successful system testing. User Acceptance Testing is operationally focused. It asks: Do account managers find the departure risk scores actionable? Does the action ladder align with how the organization actually responds? Are the scoring cadence and reporting format compatible with existing workflows? UAT requires real users in a representative operational context and cannot begin until system testing has passed.

Controlled deployment follows UAT. A limited production rollout — one business unit or customer segment — with defined success metrics, a rollback protocol, and ongoing OCC governance reviews.

Downstream deployment decisions by component. Platform, licensing, and environment alternatives for each model component — Stata MSM, MATLAB LSTM, and Python scoring harness — will be evaluated in consultation with IT and the relevant platform vendors during the environment setup phase. The fusion decision CSV is the natural integration point for any downstream system. OCC governance travels through all environments — it is not a development-only tool.
Hybrid cognition in practice — a Bayesian belief network as the human-in-the-loop layer
Why this matters beyond the horse race

Discovery 20 in this study's governance log is the broadest finding of the whole exercise: effective decision support in complex environments rarely comes from one model, however accurate — it comes from structuring multiple evidence streams, including human judgment, so they can be combined and audited together. The mechanism that makes that combination rigorous, rather than a gut-feel override sitting on top of a model's number, is Bayesian Weight of Evidence (WoE), implemented here as a Bayesian belief network: a fixed population prior, six evidence nodes, and one posterior — with the machine's forecast entering as a node alongside the human's, not above it.

Every signal in this study — whether produced by a model or stated by a person — is expressed in the same unit: decibans (dB), a log-odds scale measured relative to a fixed population baseline (10% prior departure rate). Because WoE is additive in log-odds space, a machine's forecast and a salesperson's field judgment can be summed directly into one auditable posterior, instead of one silently overriding the other. That additivity is the entire trick — it's what lets "hybrid cognition" mean something more precise than a person ignoring the model when they feel like it.

None of this matters if it doesn't change what someone does before the customer is actually gone. The entire economic argument in this study — Economic Detection Potential (EDP) weighed against False Positive Penalty (FPP) — rests on one assumption: the value an organization can still recover from a departing customer declines the longer detection takes, because the intervention window narrows as the relationship deteriorates. A forecast that arrives after the customer has effectively decided to leave has academic interest and close to zero economic value. The entire point of looking ahead — statistically, and now with the human layer added — is to surface the signal early enough that there is still a real decision to make and a real action that can change the outcome. Hybrid cognition exists to shorten that detection lag without manufacturing false confidence: the human layer fills exactly the gap where the model has a real but not-yet-confirmable signal, instead of forcing a choice between acting on nothing and waiting for statistical certainty that may arrive too late to act on.

Figure 1 · The belief network — one prior, eight evidence inputs, one auditable posterior
Every input is a Bayes factor expressed in decibans. Because log-odds are additive, the machine's forecast and the account team's field knowledge enter the same summation as peers — neither overrides the other, and every term is recoverable after the fact. Worked example below uses the study's own action ladder (FLAG ≥ 2 dB · OUTREACH ≥ 6 dB · ESCALATE ≥ 10 dB · INTERVENE ≥ 20 dB).
STARTING POINT MACHINE EVIDENCE · MODEL OUTPUT CALIBRATED EVIDENCE · EMPIRICAL, FROM THE REGISTRY JUDGMENT-ONLY EVIDENCE · SALESPERSON, RATIONALE REQUIRED Population prior — 10% departure base rate −9.5 dB Fusion v2 forecast — elevated, not confirmable +4.3 dB Account tier (A) +3.2 dB Revenue concentration (top decile) +1.0 dB Champion turnover — confirmed on call +6.0 dB Competitor mention +2.0 dB Call sentiment — not entered 0.0 dB Budget-cycle signal — not entered 0.0 dB Σ LOG-ODDS SUM no weighting, no override POSTERIOR Σ evidence = +16.5 dB 83.2% P(departure) this week RESULTING ACTION ESCALATE intensity 0.90 · $17.5k if wrong Machine node alone: +4.3 dB → 23.0% → FLAG (intensity 0.40)
Reading the figure. The dashed box is the fixed population base rate, not evidence — it is where every customer starts before anything is observed. The six evidence nodes below the machine node are the human layer: two are empirical Bayes factors computed from the customer registry and apply automatically, four are salesperson judgment and default to 0 dB until entered. Nodes left blank contribute nothing — a node with no evidence behind it does not quietly push the posterior in either direction. Independence assumption. Straight summation treats the nodes as conditionally independent given the customer's true state. That is an assumption, and it is the one most likely to be wrong in production: champion turnover and negative call sentiment plausibly travel together, and adding both at full strength would double-count. The logged-override dataset described below is what would eventually let those correlations be measured and the node values adjusted — which is the reason for logging, not an incidental benefit of it.

The same network as a running ledger — how 10% becomes 83.2%

Each row adds one Bayes factor to the running log-odds total. The right-hand column is that total expressed back as a probability, so the effect of each individual piece of evidence is visible on its own line rather than buried in a single output number.

Evidence node Source Bayes factor Running total P(departure)
Population prior Locked 10% base rate −9.5 dB 10.0%
Machine forecast Fusion v2 weekly WoE +4.3 dB −5.2 dB 23.0%
Account tier (A) Registry — empirical +3.2 dB −2.0 dB 38.5%
Revenue concentration Registry — empirical +1.0 dB −1.0 dB 44.0%
Champion turnover Salesperson — logged +6.0 dB +5.0 dB 75.8%
Competitor mention Salesperson — logged +2.0 dB +7.0 dB 83.2%
Call sentiment Not entered 0.0 dB +7.0 dB 83.2%
Budget-cycle signal Not entered 0.0 dB +7.0 dB 83.2%
Total evidence 1 machine · 2 calibrated · 4 judgment +16.5 dB +7.0 dB 83.2%
Machine only — what the model can defend on its own
23.0%
posterior · +4.3 dB of evidence
Action triggeredFLAG
Intervention intensity0.40
Cost if wrong$1k
A defensible reading of the behavioural data and nothing more. The account gets an internal notification, and the organisation waits for the trajectory layer to accumulate enough history to confirm or dismiss. Every week of that wait is a week of decay against the recovery curve.
Machine + logged human evidence
83.2%
posterior · +16.5 dB of evidence
Action triggeredESCALATE
Intervention intensity0.90
Cost if wrong$17.5k
The same model output, plus two registry facts and two logged field observations. The intensity multiplier rises from 0.40 to 0.90 — a 2.25× increase in modelled recovery on the same customer, in the same week, with every contributing term on the record and reviewable against what actually happened.
What Bayes actually contributes to hybrid cognition
1 — Odds form turns combination into addition
Bayes' theorem written in odds form says posterior odds equal prior odds multiplied by the likelihood ratio of each independent piece of evidence. Take logarithms and the multiplication becomes a sum. That single algebraic move is what makes the whole structure work: combining a model's output with a person's observation stops being a negotiation about whose number wins and becomes an arithmetic operation with a defined answer. There is no blending parameter to argue over, no weighting scheme to tune, and no hidden precedence rule — the posterior is what it is because of what was entered, and removing any single term reproduces exactly the posterior you would have had without it.
2 — Decibans give both kinds of knowing one unit
A deciban is ten times the base-10 logarithm of a likelihood ratio — I.J. Good's unit from the wartime Bletchley Park work, and the same unit used throughout the study's Weight of Evidence scoring. Its value here is that it is indifferent to provenance. A LSTM's confidence score and an account manager's read on a customer call are ontologically different things, but once both are stated as "how much does this multiply the odds of departure," they become the same kind of object and can sit in the same column. The action ladder is denominated in the same unit, so evidence converts directly into a decision without a translation step where judgment can be smuggled in.
3 — A network, not an override switch
Structuring this as a belief network rather than a human veto has a specific consequence: each node is a separate, named, independently maintainable input to one shared posterior. A node can be added, recalibrated from new outcome data, or retired without touching the model or the other nodes. Nothing overwrites anything. The practical difference is auditability — an override tells you only that someone disagreed, whereas this structure records which piece of evidence moved the number, by how much, and on whose authority. That is the difference between knowing a decision was overruled and being able to review why.
Why the gap is worth closing — customer C011 from this study
The argument above is only worth making if detection lag actually costs money. In this dataset it does, and one customer shows it cleanly.

C011 is a Tier A account carrying $1,097.0k of annual revenue, departing by market-share erosion. It is a true departure and every architecture eventually called it correctly — this is not a story about a model getting the answer wrong. The MSM surveillance layer flagged it early and was credited $713.1k of Economic Detection Potential. The confirmation path reached the same conclusion about the same customer and was credited $54.0k. Both are true positives. Both are verified in the Customer Validation Workbook. The $659.1k difference between them is produced entirely by when the call was made, not by whether it was right.

The fusion weekly-score files show why. The trajectory layer's signal on C011 first appears in week 73, and under the escalation policy it is logged as a review-only signal — real, recorded, and unable to trigger an action on its own — for 75 weeks before the surveillance layer independently crosses its threshold and an action fires in week 135. That 75-week window is not a modelling defect to be engineered away; it is the honest state of the evidence, and it is precisely the interval in which a person who knows the champion has left holds information the behavioural data has not yet expressed.

Applying the locked recovery formula to that window makes the stake explicit: an ESCALATE action on C011 in week 73 models $389.2k of recoverable value against the $54.0k the late confirmation was actually credited. The belief network is the mechanism for getting that evidence into the decision in week 73 in a form that survives review — rather than as a phone call nobody wrote down.
How this works in practice — the weekly sequence
1 — The model flags, with its limits attached
The surveillance layer returns a moderate forecast probability on an account. The trajectory-confirmation layer, which needs a longer observation window, has not accumulated enough history to confirm or dismiss it. The honest machine output is "elevated, not yet confirmable" — which is neither silence nor false confidence. It enters the network as one node at its actual strength, +4.3 dB, and is not inflated to justify acting or suppressed to avoid a false alarm.
2 — The account team adds what the data has not recorded
Tier and revenue concentration apply automatically from the registry; they are calibrated and need no human input. What the model cannot see is that the customer's champion left this week, confirmed on a call rather than rumoured. That is entered as a judgment node at a strength the salesperson selects, with a typed rationale required before it can be saved. This is the evidence that closes the gap the observation window cannot close yet.
3 — The combined call, while there is still time to act
The terms sum to +16.5 dB and a posterior of 83.2% — enough to justify a retention call this week rather than waiting for the trajectory layer to confirm on its own. By the time a model-only signal clears its confirmation bar, much of the recoverable value is gone. Every input that produced the number is timestamped, so the decision can be reviewed later against what actually happened to that account, whichever way it went.
Why log the judgment instead of just acting on it
Four of the six human evidence nodes — call sentiment, competitor mention, champion turnover, and budget-cycle signal — are deliberately uncalibrated today. No dataset yet connects "champion turnover, logged at this strength" to actual won and lost outcomes, so each is salesperson judgment entered as a dB value with a required rationale. Stating that plainly is the point: the node values in the worked example above are a considered starting position, not measurements, and the page should not pretend otherwise.

The other two nodes have already been through the process that fixes this. Account tier (+3.2 dB) and revenue concentration (+1.0 dB) are empirical Bayes factors computed from the customer registry — how much more often Tier A and revenue-concentrated accounts actually appear among departures than among stable accounts. They were estimates once too. What converted them was outcome data.

Every logged override is a row in the dataset that would do the same for the remaining four. Once enough overrides accumulate against known outcomes, each judgment node can be re-derived as an empirical Bayes factor, and the correlations between them — the conditional independence assumption noted under Figure 1 — can be measured rather than assumed. The governance discipline applied to the models in this study was that every number traces to a source and nothing is asserted without evidence. Applying that same discipline to the human contribution is what distinguishes this from an override button, and it is the reason the logging requirement is structural rather than administrative.

Full Working Paper — Simulation, Governance & Evidence Fusion

Methodology · twenty governance discoveries · mathematical appendix · case studies · Appendix D results

↓ Download PDF