The business case for operational transformation

Most AI business cases fail at the same place: they put several different kinds of value into one number. Cash recovered, hours released, a cost avoided and a working-capital effect are different things. Finance treats them differently, and a board that discovers they were added together will discount the whole case. This article gives a CFO a way to keep them apart, and gives a COO a way to establish the baseline before anything is redesigned or scaled. It leaves out what the program costs. That is covered in the cost article, and the numbers from its synthetic scenario are reused here only as inputs.

Five quantities, never one number

1. Gross value. What the change produces over a stated horizon, in categories that are never summed blindly.
2. Net value. Gross value minus every cost over the same horizon. This is the decision number.
3. Released capacity. Hours that stop being spent on work that has been removed or reduced. Real, useful, but not cash.
4. Cash and balance sheet. One-off releases of working capital, which are not profit and can reverse.
5. Horizon. The period you are measuring. A case that is negative at 12 months and positive at 36 is a different decision from one that is positive at 12.

We solve bottlenecks worth $1M or more. That is a fit criterion. It is not a promised saving, an annualized return or a booking threshold, and no multiple is promised. What matters is whether the net case holds over a horizon you choose, under assumptions you can inspect.

The value categories

Two rules prevent most double counting. First, do not count the same effort twice, for example once as hours saved and once as lower rework cost. Second, never multiply a straight-through rate and a time-saved figure over the same volume. Use the volume that is actually expected to pass through the new workflow, and subtract what review, new exceptions and shifted work take back.

Baseline first: how to measure an operating bottleneck

You cannot build a net case without a baseline, and you cannot redesign an operation sensibly without one. Microsoft’s guidance for AI agent adoption says to record current performance of the KPIs the work must improve before development begins and to use business metrics as go/no-go decision gates. Its planning guidance adds that proof-of-concept results should refine estimates. A baseline is what turns those instructions from slogans into something checkable.

Write the definitions down

A baseline is a set of definitions, not a set of averages. Before collecting any data, agree these:

• Unit of work. What counts as one case? A deduction, an invoice, a revision, an RFI. If two teams count differently, the numbers will never reconcile.
• Boundary. Where does the process start and end, and which teams, systems and decisions does it cross? Many bottlenecks live at the handoffs, not inside a team.
• Eligible population. Which cases are in scope and which are excluded, and why.
• Cycle time. Elapsed time between two events, stated as clock time or working time, with the start and end events named.
• Touch time and wait time. Minutes of hands-on work versus time spent waiting. Most delay is wait time, which an automation of the touch step will not reduce.
• First-pass yield and rework. The share of cases completed without being returned, corrected or reopened.
• Exception taxonomy. The reasons work leaves the standard path, with an owner for each reason and a count.
• Backlog age. Open cases by age band, not only the total.
• Money attached. Value in flow, value written off, expedite costs, credits.
• Period, source and owner. The dates covered, the system or sample each number came from, and who is accountable for it.
• Confidence. Observed from an event log, sampled with a stopwatch, or estimated. Say which.

Choose the measurement period deliberately

The period has to include a normal cycle of the process, including peaks and month-end effects, or the baseline will flatter or punish the future result. As a working starting point, which is Sembium’s rule of thumb and not a standard, 8 to 12 weeks of volumes plus a representative sample of cases, including failures and edge cases, is often a workable minimum. Slow-moving processes need longer. Record seasonality and workflow changes that happened during the period so the later comparison can allow for them.

Separate observed from estimated

Event logs give you timestamps and counts. Samples and stopwatch studies give you effort. Interviews give you reasons and context. Each has different reliability, and a baseline should say which number came from where. Sembium’s rule is simple: never invent a loss figure. An unknown stays unknown and goes onto the list of things to measure.

A synthetic baseline

The table below is synthetic and illustrative. It describes an invented B2B company’s customer-to-cash dispute handling and no Sembium client.

The last row illustrates a typical weakness: reasons are often not coded consistently, so that baseline is low confidence and the first piece of work is to code them. A baseline that lists what it does not know is more useful than one that looks complete.

A synthetic business case over three horizons

Now the value model. These numbers continue the synthetic company above and are assumptions, not forecasts.

Value assumptions:

• Invalid value per year: $9.0m x 35% = $3.15m. At baseline, 30% is recovered.
• Recovery rises by 6, 12 or 20 percentage points in the low, base and high cases, which gives an extra $189,000, $378,000 or $630,000 recovered per year at steady state.
• Open-dispute age falls from 55 days by 8, 15 or 22 days. One-off working-capital release = $9.0m x days / 365 = about $197,000, $370,000 or $542,000. This is a balance-sheet effect, not profit.
• Financing effect on that release at an assumed 6% cost of funds: about $12,000, $22,000 or $33,000 per year.
• Cash cost removed that finance would recognize (avoided overtime and temporary cover): $0, $20,000 or $40,000 per year.
• Steady-state annual benefit in the P&L sense (recovery + financing + cash cost): about $201,000, $420,000 or $703,000.
• Ramp: half a year of steady-state benefit in the first 12 months, full benefit afterwards. This is a modeling assumption, not a schedule.
• Capacity, shown separately: 4,800 adopted cases (80% of 6,000) x (50 min baseline - 14 min residual review - 6 min new exception work - 4 min effort shifted elsewhere) / 60 = 2,080 hours per year, worth $99,840 at $48 loaded. It is not included in the net figures below.
• Costs: taken from the synthetic scenario in the cost article: $701,000, $831,000 and $961,000 cumulative at 12, 24 and 36 months.

What a CFO should take from the table:

• Gross value looks comfortable; net value does not. At 36 months even the low case produces a sizeable gross value, yet its net value is negative. A gross-value figure is not the same as a net case.
• The base case only just turns positive at 36 months. A decision made at 12 or 24 months would be negative. This is the horizon question stated in numbers.
• The one-off working-capital release ($370,000 in the base case) is kept off the table. If you include it as cash position, the base case at 36 months is about $459,000 better off on cash, but that release is not profit and can reverse if payment terms or behavior revert.
• Capacity is real but separate. 2,080 released hours per year are worth something only if they remove a cost or produce measured additional output.
• The sensitivity is mostly the recovery rate. That is the assumption to test hardest in the baseline and in a pilot on the hardest stage.
• Synthetic sensitivity: what break-even by month 24 would require. Cumulative cost at month 24 is $831k. The base case reaches $630k, so it falls $201k short. If months 13-24 stay at the base steady state of $420k, months 1-12 would have to deliver $831k - $420k = $411k of gross value, against $210k assumed: about 1.96 times as much, or roughly $34k per month instead of $17.5k. The assumed steady state is $420k / 12 = $35k per month, so break-even by month 24 would need almost the full steady-state benefit from the first month. That is the same ramp assumption seen from the other side; it is illustrative arithmetic on invented numbers, not a forecast or a target.

A vendor case, read carefully

HighRadius publishes a case about HP Hood, a USD 3.4 billion family-owned dairy company with 12 plants. The page reports USD 5 million of invalid deductions recovered annually, up from under USD 1 million, settlement time down from 60 to 12 days and about USD 15 million of working capital freed. It also describes governance changes made alongside the software: a USD 25 tolerance, 16 standardized reason codes and accounts receivable reorganized into three pods. The page quotes 85% in one place and 72% in another for the reduction in open deductions.

It is useful because it shows the categories above in a real setting: recovered cash, a working-capital release and process and governance change together. What it does not show is equally important. It is a vendor-published account of one company, with no independent audit, no stated counterfactual and no program cost that we found on the page. It is not a typical result, and it says nothing about what Sembium or any other provider would achieve in a different operation. Use it as a prompt for what to measure, not as a benchmark to promise.

What this does not prove

• The scenario figures are invented. They show a method, not a result, and the low case is negative on purpose.
• The baseline table describes a fictional company. Real baselines are messier, and the first finding is often that the data is worse than assumed.
• A positive model is not a forecast. The real value depends on adoption, exception handling, upstream behavior and decisions that people make.
• The vendor case is one reported example and cannot be generalized.
• Nothing here promises savings, a return multiple, a timeline or a ranking in any search result.

A practical decision rule

Choose the horizon first, with finance. Write the baseline definitions with the process owner. Compute the net case in three scenarios with the assumptions beside them. Then agree what would stop the work: Microsoft’s guidance describes treating decision gates seriously and pausing when a pilot does not meet agreed results, and a justified no-go is a legitimate outcome of an honest case.

Next step

Got a bottleneck that costs you $1M or more, and you want it solved? Book the 30-minute bottleneck call. Bring your baseline if you have one. You leave with one practical recommendation and a clear answer on fit, and no confidential data is needed.