This is an illustrative scenario, not an AIPIVT client engagement. The firm, the team, and every figure below are made up to show the method.

Picture a financial services firm of about 400 people. Its operations team of six handles trade confirmations, settlement reconciliation, client reports, and a monthly regulatory filing. Everyone agrees the work matters. Everyone also agrees that too much of it is people copying fields from one screen into another and checking that two numbers match.

The COO wants AI in the back office. That is too big to measure, so the first job is to make it smaller.

Pick one workflow

A good first workflow has high volume, mostly rule-based steps, structured inputs, a clear end state, and its data already sitting in a system of record. It also needs one person who will own it afterward.

In this firm, confirmation matching fits. Every trade produces a counterparty confirmation, and someone has to match it against the firm's own booking before settlement. About 6,000 arrive each month. The regulatory filing is the opposite case: low volume, heavy judgment, and a costly mistake. It can wait until the team has one system it trusts.

Measure the baseline from the system, not from memory

Ask people how long a match takes and they will describe a good day. The matching queue keeps timestamps, so use them. Where it does not, time a sample of a few hundred items across a full month, including month-end.

The baseline for this scenario:

  • Volume: 6,000 confirmations a month.
  • 80% match cleanly and take about 4 minutes each: 4,800 items, 320 hours.
  • 20% are exceptions and take about 20 minutes each: 1,200 items, 400 hours.
  • Total: 720 hours a month. At 150 productive hours per person per month, that is 4.8 FTE of time, spread across six people who also do other work.
  • Fully loaded cost: $90,000 per FTE a year, or $50 an hour. Annual handling cost: 720 × 12 × $50 = $432,000. Cost per item: $6.00.
  • Errors: about 30 items a month (0.5%) reach settlement wrong, at roughly $150 each to fix. That is $54,000 a year.
  • Cycle time: a median of 26 hours from receipt to match.

Write these on one page. Every later claim gets checked against it.

Set the target before the build

A target needs a named line and a unit. Here the line is handling cost per item: people's hours plus the new system's running cost, divided by the items handled. Error cost is tracked as a second number, on its own.

The figures AIPIVT plans against on a high-volume workflow are 60–80% lower cost and about 5x throughput, with first measurable impact in 4–8 weeks. Those are methodology targets set against a baseline, not client results. Applied to this scenario, the target is a handling cost of $1.20–$2.40 per item, down from $6.00, and about 42 items per human hour, up from 8.3.

Agree those numbers with finance before anyone writes code. If the target cannot be stated this plainly, the workflow is not ready.

Fix the process, then automate it

Mapping the steps turns up work that should stop whether or not anything gets built.

A check that catches nothing. The process has two reconciliation checkpoints, added years apart after different incidents. The error log shows the second one has not caught a single error in 18 months. Removing it is a process decision, and it costs nothing.

Bad data at intake. Many exceptions start upstream: a missing reference, a date in the wrong format. Validating confirmations when they arrive, and rejecting malformed ones back to the sender, is cheaper than cleaning them later. The model below still keeps the exception count at 1,200, because the effect of the intake fix has not been measured yet.

Exceptions designed for a person. The exceptions that stay human should arrive with the trade, the confirmation, the difference, and the likely cause on one screen. Today the handler opens four systems to build that picture.

A fixed milestone, on the system of record

The scenario runs one eight-week milestone. The matching runs against the books the firm already has, reads from the existing confirmation feed, and writes its result back into the same queue. Nothing is migrated.

  • Weeks 1–2: confirm the baseline, remove the dead checkpoint, add intake validation, agree the target.
  • Weeks 3–6: build the matching and the exception screen. Assemble a test set of a few hundred past confirmations with known answers.
  • Weeks 7–8: parallel run. The system and the team match the same items, and every disagreement gets read by a person. Cutover happens only when the disagreements are understood.

What the model says

Illustrative model, not a measured result. Clean matches are handled by the system, and a person reviews a 5% random sample at 2 minutes each: 240 items, 8 hours. The 1,200 exceptions arrive with context and take 10 minutes each: 200 hours. Human time falls from 720 to 208 hours a month. Labor: 208 × $50 = $10,400 a month. Assumed running cost of the system: $3,000 a month. Handling cost: $13,400 a month, or $2.23 per item, 63% below the $6.00 baseline. Throughput: 6,000 ÷ 208 = 28.8 items per human hour, about 3.5x.

So the model lands inside the cost target and short of the throughput figure. That is the usual shape when exceptions stay with people, and it tells you where the next milestone should go: exception handling time, not more automation of the clean matches. Error cost stays at $54,000 in the model. Nobody has measured the new error rate yet, and a benefit you have not measured does not belong in the business case.

At full rate the saving is $36,000 − $13,400 = $22,600 a month, or $271,200 a year. Payback in months is the one-time build cost divided by $22,600, and the ramp in the first months makes it longer. Calculating AI ROI walks through the full cash flow and NPV.

Hand it to a named owner

The owner here is the operations lead who ran the old queue, named in the milestone plan before week one. At handover that person should have:

  • a runbook written for the team, not for the builders
  • the test set, so any change can be rerun against known answers
  • a weekly report with the two numbers: handling cost per item and errors reaching settlement
  • a switch that sends every item back to the manual queue if the match rate drops
  • the accounts, keys, and access, held by the firm

If nobody inside the firm can change the system or switch it off, the milestone is not finished.

The headcount question

The model returns 512 hours a month, about 3.4 FTE of time. That is not 3.4 people. It is hours spread across six people's weeks, and it only becomes money when someone decides what to do with it.

There are three honest options. The team can absorb growth: at 9,000 confirmations a month, the old process needs 7.2 FTE of time and the new one about 2.1. The hours can go to work that has been waiting, such as the client-report backlog or fixing the root causes behind the exceptions. Or roles are not refilled as people leave.

Pick one before the build and write it into the business case. Tell the team early, because they will work out the arithmetic for themselves. The system is run by your people and one of them owns it. What you do with the hours that come back is your decision.

Is one of your workflows like this?

If a workflow in your back office has the volume, the rules, and the data in a system you already run, start with the baseline. The process cost worksheet turns one workflow into a cost per item and a cycle time on a single page.

To test it against one specific workflow, book a 30-minute scoping call. We use it to name the workflow and to check that you can grant access to the process and the data a baseline needs.