Your board has probably used the word by now. That is not the interesting part. The interesting part is the question sitting underneath it, which is a great deal older and a great deal less exciting: can this operation's data actually carry what you are about to put on top of it?

At most institutions we walk into, the honest answer is: some of it, in places, provided somebody checks first.

That is not pessimism. It is the ordinary condition of a thirty-year-old investment operation. Positions get reconciled by hand at quarter-end because that was quicker in 2011 and nobody revisited it. A valuation convention gets applied consistently for a decade because one person applies it consistently. A field means one thing in the accounting platform and something subtly adjacent in the reporting warehouse, and everyone downstream has quietly learned to compensate without ever writing the compensation down.

None of that breaks anything while people are in the loop, because people notice. An automated process does not notice. It produces an answer that looks exactly like an answer, at volume, inside a control framework designed to catch a human making an obvious mistake.

Three things worth establishing before you commit budget

Not a technology assessment. A plain inventory of what is true.

1. Whether the same thing is called the same thing everywhere

Multi-asset transaction feeds, derivative structures and corporate action records typically live in three or four back-office systems that were never designed to agree. They differ on naming, on granularity, and on timing. The useful distinction is not which system is right — it is which records are genuinely reconciled and which are merely asserted, and nobody can tell those apart by looking at a screen.

2. Which exceptions are governed, and which are merely handled

Every operation has manual overrides. That is fine, and often correct. The problem is that an override applied by judgment and never recorded leaves behind the decision without the reasoning. Automate on top of that and the process inherits the judgment — inconsistently, and with no way to explain itself when asked.

3. Whether you could trace any given number back to its source

Pick a figure on a client statement at random and follow it home. If the honest answer is "give me a week and I'll ask Linda", you have a lineage problem, and it becomes a much larger problem the moment the number is produced faster than Linda can check it.

The unglamorous conclusion

Run this properly and the finding is usually specific and slightly boring: two of your five candidate processes are ready today, one needs a fortnight of definition work, and two should be fixed rather than automated at all. Occasionally the finding is that you are in better shape than you feared, which has been worth the engagement more than once.

Either way it is considerably cheaper to learn it now than eighteen months into a program with a number already reported.

This is not new work for us. Establishing what is true before anyone builds on top of it is the job we have had since 1993. Only the thing being built on top has changed.