An AI business case is a capital request like any other. It has a cost, a benefit, a timing profile, and a set of assumptions someone has to defend. The difference is that the benefit side often arrives as a percentage with no line item behind it. A CFO's job is to find the line item.

Where the money shows up

Automation in a back-office workflow can move three things. Only one of them is EBITDA in the first year by default.

The OPEX line

This is the direct one: people's hours and contractor spend on the workflow go down, and a new running cost for the system goes up. The case should net the two on the same line. A proposal that shows the labor saving and leaves the model, hosting, and support costs in a separate appendix is overstating the result.

Hours freed are not savings yet. They turn into money only through a decision: a hire you no longer make, a contractor you release, a role that is not refilled, or volume growth handled by the same team. If the case does not say which, the saving is capacity, and capacity should be valued as such.

Cycle time

Faster invoicing and collections reduce days sales outstanding. For a business with $100M in revenue, each day of DSO is about $274,000 of receivables, so five days frees about $1.37M of cash once. At an 8% cost of capital that is worth roughly $110,000 a year.

That benefit is real, and it is not EBITDA. The cash release is a balance-sheet movement and the carrying benefit sits in interest, below EBITDA. A case that adds freed working capital to the EBITDA uplift has counted it in the wrong place.

Error cost

Errors cost rework hours, write-offs, late-payment penalties, and audit adjustments. Some of that is visible in the ledger and some is not. Count only what you can trace to an account or a timesheet. The rest can go in the narrative, not in the number.

Margin points versus percent

A frequent error in AI proposals is mixing up margin points and percentage growth. They are different units. At a 10% margin, the gap between them is a factor of ten.

Illustrative worked example. A company has $50M in revenue and $5M in EBITDA, a 10% margin. One workflow change saves $400,000 a year, net of the system's running cost. EBITDA becomes $5.4M. The margin becomes $5.4M ÷ $50M = 10.8%. That is 0.8 margin points, and 8% EBITDA growth. Both statements describe the same result. "8 points of margin" would overstate it tenfold.

When a proposal quotes "points," ask for the revenue figure it divided by. When it quotes a percentage, ask what it is a percentage of: EBITDA, the workflow's cost, or the department's budget. A 60% cut in a workflow that costs $400,000 a year is $240,000, whatever the percentage looks like on the slide.

Ask too how the case treats the build cost. Expensed, it lands in year-one EBITDA. Capitalized, the treatment depends on the arrangement. Check it with your controller, because it changes the year-one margin even though it does not change the cash.

What a baseline must contain

A target means nothing without the number it is measured from. For each workflow in the case, the baseline should state:

  • volume per month, and the period it was measured over, including a month-end
  • human hours, taken from system timestamps or a timed sample rather than estimates
  • fully loaded cost per hour, and the resulting cost per item
  • cycle time, as a median and a tail, because the tail is where penalties come from
  • error rate, and the cost per error traced to the ledger
  • what the current tools cost to run
  • who owns the workflow today, and who will own it afterward

This fits on one page. The process cost worksheet is that page for a single workflow.

Reading a target against a result

A target is a number agreed before the build, on a named line, measured the same way as the baseline. A result is that same measure after the system has run in production for a stated period. Most of the confusion in AI reporting comes from treating the first as the second.

AIPIVT plans against 60–80% lower cost and about 5x throughput on a high-volume workflow, set against the client's baseline. Those are methodology targets, not client results. Any vendor's figures deserve the same label until they have been measured on your data.

When a result comes back, check that:

  • the denominator is the same as the baseline's (cost per item, not cost per team)
  • the period is comparable, with a month-end in both
  • the system's running cost is included
  • volume has not shifted in a way that flatters the unit cost
  • the ramp months are shown, not averaged away

What to ask a vendor

Is the model built on our data? Industry benchmarks can start a conversation. They are not a business case.

What is the full cost? Licenses, model usage, integration, data preparation, training, maintenance, and a contingency.

Which line does the target sit on, and who measures it? The answer should name a P&L or OPEX line and a unit.

What happens when the system is wrong? Who catches it, how fast, and how does an automation error stay out of the books?

How does it fit our controls? Anything that touches financial controls has to be designed with SOX and your audit trail in mind from the start.

Who owns it after the engagement? A named person inside your company, with the runbook, the test set, and the access to change or switch off the system.

Is any number in this deck a client result? If so, ask for the baseline it was measured against. If not, it is a target, and it should say so.

Where to start

Start with one workflow and its baseline. The process cost worksheet takes you to a cost per item and a cycle time from your own system. If you want to test a specific workflow against it, book a 30-minute scoping call. We use the call to name the workflow and to check that you can grant access to the process and the data a baseline needs.