What to automate first in an advice firm, and what to never automate.
Most automation projects fail at the choosing, not the building. Here is the test we use, which takes about ten minutes and rules out the expensive mistakes before anyone writes anything.
Published 6 min read Sources checked 8 September 2026
Filed under Process automation, Financial advice operations, AI adoption
The short answer
Score each task on two axes: how reversible is a mistake, and how easily would you spot one. Automate what is reversible and checkable. Never automate what is irreversible and hard to check: that quadrant contains the recommendation itself, and no efficiency gain is worth entering it.
On this page · 8 sections
The usual question is “what can AI do for us?” It is the wrong one, because the answer is “quite a lot”, and that tells you nothing about where to start.
The better question has two halves, and you can answer both about any task in under a minute.
Why reversibility?
Because it sets the cost of being wrong, and every automation will be wrong sometimes.
A mis-tagged email costs a moment. A payment sent to the wrong account costs a relationship, a complaint, and possibly a report. Same error rate, entirely different exposure. Start where the blast radius is small enough to learn in.
Why checkability?
Because this is the axis people forget, and it is the one that hurts.
A completeness check either produced the list or it did not: you can see it. A suitability narrative reads perfectly whether or not it is right about the fund. The second kind of task hides its failures inside fluent prose, and fluency is precisely what these systems are best at.
If you cannot describe how you would notice the mistake, you have not found an automation candidate. You have found a way to make errors quieter.
The middle two quadrants
| Quadrant | Example | What to do |
|---|---|---|
| Reversible, easy to spot | Re-keying, chasing, tagging | Automate. Start Monday. |
| Reversible, hard to spot | Drafting a narrative section | Automate with a fact-level check. The prose is not the check |
| Irreversible, easy to spot | Booking into a live diary | Automate with a confirmation step. Cheap insurance |
| Irreversible, hard to spot | The recommendation | Do not. Not with review, not with guardrails |
The order to do them in
- Pick the highest-volume task in the top-left quadrant. Volume, not glamour. The dull one that happens forty times a week.
- Write down how you will know it worked: a number, measured now, before anything is built.
- Build only that. One task, one leg, one handover to a named person.
- Run it for a month and check the number. If it did not move, stop. Do not build the second thing on top of a first thing that failed.
- Then take the next one. Not a platform. The next task.
Why not do it all at once?
Because long, multi-step, cross-application work is exactly where agents fall over: on OSWorld 2.0’s long-horizon tasks the best agent completes 31.4% on its own, against 86.0% for the best model on OSWorld’s bounded desktop tasks, and the best checked system on Carnegie Mellon’s simulated company resolves 42.9%.
Doing it one leg at a time is not caution. It is buying the 86% number instead of the 30% one.
What this does not tell you
Where any given task sits. Reversibility and checkability are judgements, and the same task lands in different quadrants in different firms: a draft letter is checkable where a second person reads it and not where nobody does. The two axes are our framework, marked as opinion in the sources, not an industry standard.
Volume is missing from it too. A reversible, checkable task that happens twice a year is not worth automating at all, and the matrix will happily put it in the first quadrant.
Questions people actually ask
What should a financial advice firm automate first?
The highest-volume task where a mistake is both reversible and easy to spot: typically re-keying between systems, completeness checks before submission, and chasing clients or providers for documents.
What should never be automated in an advice firm?
Anything where a mistake is both irreversible and hard to notice: the recommendation itself, attitude-to-risk conclusions, vulnerability assessments, anything that moves money, and anything that reaches a client unreviewed.
Why automate one task at a time?
Because agent reliability collapses on long, multi-application work. On OSWorld 2.0’s long-horizon tasks the best agent completes 31.4% on its own, against 86.0% on OSWorld’s bounded desktop tasks. Cutting work into bounded legs buys the higher number.
Sources
- 01 OSWorld-Verified leaderboard, read 24 September 2026: best model 86.0% on 369 bounded desktop tasks. OSWorld 2.0, same day: best agent 31.4% on 108 long-horizon tasks. TheAgentCompany, Carnegie Mellon, arXiv:2412.14161: 30% at publication, 42.9% best checked on its leaderboard. Until 24 September this note gave 86.1% and 30%. osworld-v1.xlang.ai(opens in a new tab)
Weight: primary Re-checked: every 30 days Last read: 24 September 2026
- 02 The two-axis framework is Noxia’s own and is not a regulatory classification.
Weight: our opinion, labelled Re-checked: never: it cannot go out of date
Checked 8 September 2026. Next scheduled check 24 October 2026. Numbers that move (leaderboards, live indices) are re-checked every 30 days; annual datasets and rules in force every six months; dated research once a year. If something here has gone stale before we got to it, tell us and we will correct it and say what changed.
Cite this note
Noxia, “What to automate first in an advice firm, and what to never automate”, Field notes, 8 September 2026; sources checked 8 September 2026. https://www.noxia.co.uk/field-notes/what-to-automate-first-and-never
Send us your five most repetitive tasks.
We will score them on both axes and tell you which one to do first, including if the answer is that none of them are worth automating yet.