Linking Platform Investment to Business Outcomes

A fictional online grocery company spent 9.6 million over two years on a data platform program. At the review, the platform team reported that 41 data products now ran on the platform, that a new data product took three weeks to build instead of eleven, and that 85% of eligible teams had adopted it. Over the same two years, company revenue grew 14%, from 850 million to 969 million. Gross margin rose from 31.0% to 32.5%, and monthly customer churn fell from 4.0% to 3.7%. The CFO asked one question: how much of that came from the platform? The platform lead said the platform had made it all possible. The head of merchandising pointed to a pricing program. Nobody had a number, and the board could not tell whether the 3.0 million a year it now cost to run the platform was money well spent. The company and every figure in this article are invented for illustration.

The platform team had reported what it delivered and how widely it was used. Those are outputs. The CFO was asking about outcomes: changes in revenue, margin, and customers, and how much of each change would not have happened without the platform. Answering takes a chain of claims from the capability to the money, a comparison with what would otherwise have happened, a clear line between early signs and results that have arrived, and named people who answer for each number. The Case Said Four Million and the Cost Was Nine showed how to establish what a technology investment costs. This is the other half: what it earned, and what a leader should decide while the answer is still uncertain.

A chain from capability to cash

An output is something a team delivers; an outcome is a change in behavior or results that follows. Between the two sits a chain of claims, and each can be wrong. The grocery company’s largest bet ran like this. An hourly inventory feed (the capability) let the substitution engine suggest replacements based on what was actually on the shelf, not on yesterday’s stock. Fewer order lines went unfilled (an operational result). Customers spent more each week (a business variable), and the extra sales produced gross profit (the money). Writing the chain down with an owner and a piece of evidence for every link turns a vague belief into the kind of value hypothesis that Fourteen Pilots and Nothing in Production asks a strategy to state before funding. A link with no evidence is the first place to look when the numbers disappoint.

The cleanest evidence that a link holds is a randomized split. Vodafone published a case study in which half the traffic to a landing page saw a version with faster loading, through server-side rendering and image optimization, and half saw the baseline. The faster version improved largest contentful paint by 31% and produced 8% more sales. A shared platform rarely gets that kind of test, because its capability reaches everyone at once. The business question then has to be answered from comparisons the company can construct.

Revenue grew 14% while the platform was being built, so the observed growth says little about the platform. The question is how revenue differs from what it would have been without the platform, and that second number is the counterfactual. It is never observed, for the reason the causal effect entry gives: the same region cannot both receive the feed and go without it in the same weeks. The UK Treasury’s Green Book describes the same idea as the business-as-usual baseline and separates deadweight, outcomes that would have happened anyway, from additionality, outcomes that happen because of the intervention. Only the additional part belongs to the platform. The Green Book also warns that an unfair description of business as usual makes an investment look better than it is.

The feed was not switched on everywhere at once. Operations enabled it in six of the twelve regions first, chosen because their warehouses were ready, and the other six followed two quarters later. That sequence is a comparison group. The table shows average weekly revenue per active customer in the quarter before and the quarter after the first six regions switched on, while the other six had not yet received the feed.

Group	Before	After	Change
Regions with the feed (6)	62.00	64.30	+2.30
Regions without the feed yet (6)	61.50	63.00	+1.50
Difference between the changes			+0.80

Revenue per customer rose 2.30 where the feed was on, but it also rose 1.50 where it was not, because of price inflation, seasonal demand, and other initiatives. Difference-in-differences subtracts the comparison group’s change and leaves 0.80 per customer-week, or 1.3% of the starting 62.00. With 300,000 active customers over 52 weeks, that is 0.80 × 52 × 300,000 = 12.48 million of revenue a year at full rollout. Had the platform team reported the raw 2.30, the claim would have been 35.88 million, 2.9 times larger.

The result rests on an assumption: without the feed, the early regions would have changed as the later regions did. Because operations picked regions for readiness, that assumption needs a check. The team plots the quarters before the switch-on for both groups and confirms the lines moved together; the method fails when they do not. The Magenta Book, the UK government’s evaluation guidance, lists this dependence on parallel movement among the weaknesses of the method, along with its need for enough data. In this example the check passes. If every region had received the feed together, no comparison would exist. The Magenta Book then offers contribution analysis: show that each link in the chain occurred, assess the other factors that could explain the result, and conclude that the platform contributed. That gives a reasoned claim with some confidence, not a measured effect, and it should be reported under that name.

Where the money shows up: revenue, margin, retention, and growth

Each financial channel has its own arithmetic, its own delay before evidence exists, and its own natural owner. They also draw on the same customers, so an effect has to be counted in one channel and not again in another.

Channel	Platform mechanism in the case	What to measure	How it misleads
Revenue	Fresher inventory data leaves fewer order lines unfilled	Revenue per active customer against a comparison group	Reporting the raw rise, or reporting revenue as if it were profit
Margin	Retired pipelines and clusters; fewer support contacts	Gross profit in dollars and as a percentage; invoices that actually stopped	Savings that never reach a ledger; platform cost moved between cost lines
Retention	More accurate delivery windows mean fewer late deliveries	Churn and cohort retention against a comparison group	Extending an early churn rate into a customer lifetime
Growth	Shorter data product lead time brings launches forward	Launch dates for projects where data was the binding constraint	Crediting speed to the platform when another constraint set the date

Revenue and margin

The 12.48 million is revenue, not profit. At the company’s gross margin of 32.5%, it carries 12.48 × 0.325 = 4.056 million of gross profit a year. Gross margin is gross profit as a share of revenue, so adding revenue at the average margin leaves the percentage unchanged. If the substituted items sell at a 25% margin, the same revenue yields 3.12 million, and the platform raises gross profit dollars while pulling the company’s margin percentage down slightly. Over the two years, gross profit rose from 263.5 million (31.0% of 850) to 314.9 million (32.5% of 969), an increase of 51.4 million. The revenue channel is equal to about 8% of that increase (4.056 divided by 51.4). Pricing, supplier terms, and delivery routing explain the rest in ways this comparison cannot see.

The cost side of the margin story is less flattering. The platform retired nine separately run pipelines and clusters, and the invoices for them stopped, saving a verified 1.8 million a year. The platform costs 3.0 million a year to run, so the consolidation increased annual run cost by 1.2 million. The argument that teams would otherwise have bought more infrastructure is a cost avoidance, and it appears in no ledger. Consolidation did not pay for itself, which is consistent with the lesson of the budget article that total cost has to be counted before value is claimed. Classification adds a reporting effect. If the 3.0 million run cost is booked in cost of revenue, gross margin is 3.0 / 969 = 0.31 percentage points lower than if it is booked in operating expenses, and no cash has moved. A margin comparison across periods is valid only when the cost lines are classified the same way in both.

Retention

The same six-against-six design applied to monthly churn shows 4.0% falling to 3.5% where the feed was on and 4.0% falling to 3.9% where it was not. The difference is 0.4 percentage points. The value of that depends on how long the lower rate lasts. Retention follows a starting group over time, and at a constant monthly churn of 4.0% the share of the group still present after 12 months is 0.96 to the twelfth power, or 61.3%. At 3.6% it is 0.964 to the twelfth power, or 64.4%. For a starting group of 300,000 customers, that is about 9,400 more customers at month 12. A customer spends 62.00 × 52 = 3,224 a year, which at 32.5% leaves 1,047.80 of gross profit, so 9,404 customers represent about 9.9 million of annual gross profit. That figure is a run-rate after a full year, it assumes churn stays 0.4 points lower, and it is more than twice the revenue channel’s 4.06 million. It is also the claim with the least evidence behind it.

Growth

Cutting data product lead time from eleven weeks to three can bring launches forward, but the value depends on whether data was the constraint. Suppose a same-day delivery tier launched eight weeks earlier than the old lead time would have allowed, and the tier earns 90,000 of gross profit a week once running. Eight weeks of that is 720,000. If the courier contracts took sixteen weeks to sign, the data product was never on the critical path, and the platform’s share of the launch date is zero. The cost of delay is what a later date costs, and the platform earns credit only for the weeks it actually removed. The delay should also be counted once: the weeks of waiting and the sales they postponed are the same loss, not two.

Early signs and results that have arrived

A leading signal moves before the result it is thought to predict. A realized outcome is a change that has already shown up in the numbers finance reports. Both are useful, and confusing them is how a program gets credit for benefits that never arrive, the pattern traced in Live for a Year, Used by a Third. Evidence also arrives on different calendars, so a report should say which kind of evidence each number is.

Evidence	When it exists	What it supports	What it cannot show
Capability delivered; teams adopting it	Weeks after launch	The platform exists and is used	That anything changed for customers
Order lines filled; on-time delivery rate	Weeks to a quarter	The mechanism is working	That customers spent more or stayed longer
Revenue per customer, feed regions against others	A quarter or two	A realized revenue effect, with a comparison	Whether it persists, or what it cost to serve
Monthly churn, feed regions against others	Two quarters or more	An early retention effect	A 12-month or lifetime effect
Twelve-month cohort retention; gross profit in the statements	A year or more	The realized financial result	Which initiative produced it, without the comparison

A leading signal should be tested against the outcome it is meant to predict before it is trusted. Suppose the company’s older data show that customers with a late delivery churned at 7.0% a month, against 3.8% for customers whose deliveries were on time. Suppose the on-time share rose 4 percentage points more in the feed regions than in the others. Fewer late deliveries then explain 0.04 × (7.0 − 3.8) = 0.128 percentage points of churn, or 32% of the 0.4-point difference. The rest may come from better substitutions, from regional differences, or from chance, and the evidence does not say which. A cautious plan therefore uses three retention figures. If only the tested mechanism holds, the churn gap is 0.128 points and the month-12 run-rate is about 2,960 customers and 3.1 million of gross profit. If the whole measured gap holds, it is about 9,400 customers and 9.9 million. If none holds, it is zero.

Take the baseline for each measure at the start, so that project work does not affect it, as the UK Infrastructure and Projects Authority’s benefits guide advises. The guide also advises matching reporting frequency to the availability of data, because monthly benefit reports are pointless when the data arrive quarterly. Label every reported benefit as delivered, as a leading change observed, or as realized in the financials, and never present the second as the third. The adoption and task-success measures in Twenty-Three of Twenty-Five Shipped are the first links in this chain, not the end of it.

Who answers for the number, and how sure the board can be

The platform lead can answer for the capability: availability, adoption, lead time, and the 3.0 million run cost. She cannot answer for revenue per customer or churn, which merchandising, operations, and the customer organization also move. If the platform lead alone owns those numbers, she is accountable for results she cannot control, and the business owners are accountable for nothing. The IPA guide assigns benefits owners the responsibility for realizing the benefits handed to them, and the same split works here. The senior vice president of customer experience owns the churn and revenue-per-customer targets and agrees to them before funding. The platform lead owns delivery of the feed and its lead time. Finance owns the baseline and the comparison method and signs off amounts as realized. A target whose owner declines it is not counted in the case. The benefits realization entry covers owners and tests in more detail, and the decision rights entry covers who may release, change, or stop the money. In the case, the executive committee holds that right, the platform lead proposes, and finance verifies.

The board does not need a single number, and the six-region comparison cannot support one. Six regions per group is a small sample, so assume, for illustration, that the revenue effect lies between 0.30 and 1.30 per customer-week, with 0.80 as the best estimate. The Green Book asks appraisers to communicate the sources of uncertainty and to calculate switching values. The table shows what each case implies when only the revenue channel is counted, using the 1.2 million annual net run cost.

Revenue effect per customer-week	Revenue a year (million)	Gross profit a year (million)	After net run cost (million)	Years to recover 9.6 million
0.30	4.68	1.521	0.321	29.9
0.80	12.48	4.056	2.856	3.4
1.30	20.28	6.591	5.391	1.8

Two different decisions are hidden in the CFO’s question. The first is whether to keep operating the platform. The 9.6 million already spent cannot be recovered and should not drive that choice, a point the budget article makes about past spending. The forward question is whether gross profit exceeds the 1.2 million net run cost. The switching value is 1.2 million divided by (0.325 × 52 × 300,000), which is 0.24 per customer-week. Even the low case of 0.30 clears it, so the answer to the first question holds across the range. The second decision is whether the program should recover its cost within three years, which needs 9.6 / 3 + 1.2 = 4.4 million of gross profit a year, or 0.87 per customer-week. The best estimate of 0.80 falls 8% short. Retention would have to close a gap of 0.344 million a year, which is about 328 customers at month 12, or 0.11 percentage points of the 300,000. The tested mechanism alone is worth 2,960 customers, so the case probably holds if the mechanism does. A difference as small as 0.11 points may be too small for a twelve-region comparison to separate from noise, so waiting for statistical proof of the retention effect may never settle the question.

A written rule settles it better than another debate. Before asking for expansion money, the committee records, in the manner of a pre-registered decision gate, what happens at the time twelve-month cohort data exist. If the retention gap between feed and comparison regions is at least 0.11 points and the late-delivery check still holds, expansion funding is released. If the gap is zero or negative, the expansion stops while operation of the existing platform continues. Any result in between extends the review by one quarter. The operating decision is made now, the expansion decision is tied to evidence that will exist by a known date, and nobody can choose the metric after seeing the result.

At the next quarterly review, the CFO’s answer has four parts. Revenue: a measured effect of 0.80 per customer-week against a comparison, worth about 4.1 million of annual gross profit, within a range of 1.5 to 6.6 million. Margin: the platform raised annual run cost by 1.2 million net of retired invoices, and the margin percentage depends on where the run cost is booked. Retention: the tested mechanism supports about 3.1 million of annual run-rate and the full measured gap would be worth 9.9 million, but no twelve-month cohort has yet completed. Growth: one launch moved forward eight weeks because data was the binding constraint, and two others were not credited because other constraints set their dates. Four lines, each labeled by the kind of evidence behind it, replace the slide that said 41 data products.

References:

• HM Treasury, The Green Book (2026): business as usual, deadweight and additionality, sensitivity analysis, and switching values
• HM Treasury, Magenta Book Annex A: Analytical methods for use within an evaluation (PDF), sections A1.4 on contribution analysis and A2.7 on difference-in-difference
• Infrastructure and Projects Authority, Guide for Effective Benefits Management in Major Projects, 2017 (PDF)
• Google web.dev, Vodafone: 31% improvement in LCP increases sales by 8%, 2021