In June 2026, Fortlane Partners published a whitepaper whose title doubles as a thesis: "Patterns over Averages." On the surface it is about liquidity planning — how to forecast cash more reliably. Read it as a process person, though, and something else jumps out. Underneath the treasury language, it is a process mining playbook. The liquidity problem it solves is really a process problem, and the two moves it prescribes generalize to almost any core process a company runs. That is what makes it worth reading well beyond the finance team.
Fortlane's starting observation is simple and uncomfortable. Companies plan cash using averages — a customer is on thirty-day terms, so the money is booked to land on day thirty. The whitepaper's central claim is that this systematically underestimates the uncertainty in cash flow, because actual payment behavior deviates significantly from the contractual assumption, and that deviation is a primary driver of forecasting error. The thirty-day average is clean, defensible, and almost never matched by any single payment.
Here is why that is a process question and not a finance one. The average is a summary of a process — order-to-cash — that produces a wide distribution of outcomes: partial payments, cash-discount behavior, dunning, payment-run weekdays, seasonality. Steering on the average throws away the variance, and the variance is what determines whether you hold too much buffer or run short. Fortlane's answer is a two-step method, and both steps are pure business process management: first use process mining to see the real distribution of behavior (looking back), then use AI-based forecasting to express it as a range (looking ahead). Those two moves are the playbook, and they are the two levers of this edition.
The uncomfortable part of Fortlane's argument is that, in most processes, the variance is not noise around the process. The variance is the process. The contractual payment term defines a single date; what actually happens is a spread — partial payments, cash-discount behavior that depends on invoice size and timing, dunning levels, payment-run weekdays, seasonal acceleration before a customer's year-end, a slowdown in Q1. None of this shows up in "thirty-day terms." All of it shows up in the bank account.
The whitepaper puts a number on it that makes the point hard to dismiss. A customer with ten million euros in gross monthly revenue on thirty-day terms, showing a dispersion of plus or minus twelve days in actual payment dates, produces a weekly forecast bandwidth of around four million euros. That is not a rounding error. It is the difference between holding an expensive liquidity buffer and being caught short — and it is entirely invisible to anyone reading the average.
Purchase-to-pay carries the mirror image of the same blindness, and it is often worse, because companies frequently know their customers' payment behavior better than their own. Invoice approvals stall at predictable points in the month around internal reporting deadlines. Payment runs only execute on certain weekdays. Cash-discount usage is more inconsistent than anyone assumes. These are not random events. They are systematic patterns — a process — that the organization has simply never looked at as a distribution.
Fortlane frames the cost as structural, not behavioral, and that distinction matters. No one is doing their job badly. The planner who uses the average is following the only method available. The cost shows up downstream as buffers and emergency financing: to absorb the uncertainty the average hides, organizations either hold blanket safety reserves or make manual, person-dependent corrections. Blanket buffers tie up capital that earns nothing. Manual corrections do not scale, are rarely reviewed for accuracy, and walk out the door when the person who made them leaves.
There is also a quieter, more expensive error baked into the average mindset: the assumption that individual deviations cancel out at the portfolio level. They only cancel if they are genuinely uncorrelated. But seasonal effects, industry-wide payment cycles, and macroeconomic pressure push many customers in the same direction at the same time. The deviations stack rather than smooth. The portfolio average looks stable right up to the quarter where it isn't, and the organization discovers its real exposure exactly when it can least afford to.
This is what makes it a BPM problem in the precise sense, and why Fortlane's process-mining framing is the right one. The information needed to manage the process well already exists — it is sitting in the ERP system as transaction data the company has been capturing for years. What is missing is the systematic evaluation of that data. The fix is not more discipline from planners or a better spreadsheet. It is the two structural moves of Fortlane's playbook, which turn the existing data into a picture of the real process.
Fortlane's playbook is, stripped to its essence, two moves. The first looks backward and makes the true distribution of process behavior visible. The second looks forward and turns that distribution into a forecast expressed as a range. Neither replaces the existing planning infrastructure; both extend what it can see.
The first move is process mining: reconstructing the actual flow of a process from the event logs already sitting in the ERP system. For order-to-cash, the relevant case is the individual invoice, and the events run from invoice creation through to cash receipt — with upstream steps like order type and partial delivery feeding in as context. Instead of one weighted average, you get the full distribution of how the process actually behaves.
The contrast with the conventional metric is the whole point Fortlane makes. A standard analysis gives you the Weighted Average Terms (WAT) from contracts, and at best the Weighted Average Days to Collect (WADTC) — a single number that compresses the entire spread of payment behavior into one expected value. Process mining goes further: it links every payment to its process path and exposes the variance, the clusters, and the outliers the average erases. On the payables side, the same logic produces the Weighted Average Days to Pay (WADTP) and reveals where your own approval process quietly creates delay.
What emerges is not a tidier average but a behavioral segmentation. A typical order-to-cash process reveals hundreds of variants — standard flows, partial deliveries, credit notes, dunning, complaints — each with its own lead-time profile. Customers stop being grouped by master data like industry or revenue class and start being grouped by how they actually pay: on-time payers, systematic-but-predictable late payers, strongly seasonal payers, and genuinely volatile payers. That last distinction matters enormously, because a predictable late payer carries almost no planning risk while a volatile payer carries a great deal — even when both share the same average.
This is also where process mining stops being a quarterly study and becomes part of the operating model — the same shift toward continuous, agent-supported monitoring we explored in Quick Tips: 5 Mistakes With Process Frontier Agents. The whitepaper is explicit that the approach is not tied to one vendor: it names established process mining platforms — Noreja, Celonis, SAP Signavio, and UiPath — that offer pre-built connectors for SAP and other ERP systems, so the event logs needed to see the real distribution are usually closer at hand than teams expect.
Seeing the real distribution is necessary but not sufficient, which is why Fortlane's second move turns it into a forecast that preserves the uncertainty instead of collapsing it back into a single number. This is where AI-based modeling earns its place. A machine-learning model — in practice usually gradient-boosted decision trees or random forests — is trained on the patterns process mining surfaced, drawing on the customer's payment history, invoice attributes, context factors like month and dunning level, and the process variant itself. The result, for each open item, is not a date but a probability distribution.
The operationally valuable output is the shift from a point value to a calibrated range. Instead of "in week 24 we expect 14.3 million in receipts," the model says "with ninety percent probability, receipts in week 24 fall between 13.1 and 15.8 million." Fortlane points to several established methods that get you there — quantile regression trained directly on percentiles, or conformal prediction, which uses the model's own historical errors to produce calibrated intervals without changing the model. The weekly corridor is not built by adding up expected dates; it comes from a Monte Carlo aggregation of the individual distributions that treats seasonality and payment-run patterns as correlated, so the portfolio effect is realistic rather than assumed away.
A range is only useful if people trust it and can act on it, and this is where the playbook is most pragmatic. Trust comes from explainability: SHAP values decompose each forecast into its drivers, so a planner can tell the CFO exactly why a payment is expected later — invoice above five hundred thousand, customer paid fifteen days slower over the last quarter, plus a Q1 effect — rather than pointing at a black box. Action comes from making the confidence level an explicit steering variable. With a narrow band, the organization can safely reduce its buffer or invest idle cash short-term. With a wide band, it raises the credit line or reschedules a disbursement. The forecast stays honest over time through a learning loop: every actual outcome is reconciled against the forecast and feeds the next model update, so the process model adapts as behavior shifts.
The checking principle for both moves together is simple, and it is the test Fortlane's playbook ultimately comes down to: if you cannot state the dispersion of a process — not its average, but the range and shape of its real behavior — you are not steering the process. You are steering a summary statistic and hoping the variance is kind. Process mining gives you the dispersion. AI forecasting turns it into a corridor you can manage against.
For your most important core process, do you know its real distribution — the spread, the clusters, the outliers — or only its average?
If you applied Fortlane's first move and quantified how far actual behavior in your order-to-cash or purchase-to-pay process deviates from contractual assumptions, how large would the dispersion turn out to be?
How much capital is currently tied up in blanket buffers that exist only to absorb uncertainty you have never actually measured?
Where in your organization are individual deviations assumed to cancel out at the portfolio level — and have you checked whether they are genuinely uncorrelated, or whether they stack in the same direction?
If your planners received a calibrated range with named drivers instead of a single number, which decisions would they make differently?
What would change if the confidence level of your forecast became an explicit steering variable rather than an unspoken gut feeling?
Fortlane Partners wrote "Patterns over Averages" for treasury teams, but its real subject is process. A process has no average — or rather, the average is the least informative thing about it. The behavior that determines outcomes lives in the distribution: the variance, the clusters, the seasonal patterns, the correlations the mean quietly deletes. Steering on the average is not a discipline problem to be solved with more rigor. It is a structural choice to manage on a summary statistic when the full picture is already sitting in your transaction data.
Pick one core process this quarter — order-to-cash is the usual place to start — and run Fortlane's two moves on it. Use process mining to replace its average with its real distribution. Then let an AI model express that distribution as a forecast range with named drivers, and steer along a confidence level you choose deliberately. The data is already there; the ERP system has been capturing it for years. The Fortlane Partners whitepaper "Patterns over Averages", by Meythaler, Mutter and Krawczyk, lays out the liquidity case in full and is well worth reading end to end. Anyone who has once quantified the dispersion in a core process stops accepting the average as a basis for steering.
It is a whitepaper from Fortlane Partners on liquidity planning, arguing that forecasting cash on averages systematically underestimates uncertainty because real payment behavior deviates from contractual terms. Its method is a two-step playbook — process mining to reveal the real distribution, then AI-based forecasting to express it as a calibrated range — which is why it reads, underneath, as a process mining playbook.
An average compresses thousands of cases into one number, which feels like knowledge but discards the variance that actually drives outcomes. A process is a distribution of behaviors, not a single value. When you plan on the mean alone, you systematically underestimate uncertainty — and absorb the gap later through buffers, manual corrections, or missed targets.
Process mining reconstructs the real flow of order-to-cash and purchase-to-pay from the event logs already in your ERP system. Instead of a single weighted average like WADTC, it exposes the full distribution of payment behavior — process variants, clusters, and outliers — and segments customers by how they actually pay rather than by master data. That real distribution is the foundation for a forecast that reflects reality.
Instead of a single figure ("week 24: 14.3 million"), a range forecast gives a calibrated interval ("with 90% probability, between 13.1 and 15.8 million"). Methods like quantile regression and conformal prediction produce t