Key facts
- When a system drafts part of advisory work, reconstructing it later depends on three things captured at the time: what the system saw, what it produced, and what the person decided.
- A note an advisor composed and a note an advisor approved read the same on the page, so the finished artifact alone cannot show where judgment was applied.
- The knowledge of what a system was working from exists only while the system runs; unless it is captured at the time, it is never stored in anyone's memory.
- A firm can test its own memory by asking someone uninvolved to answer the three reconstruction questions from the record alone for one recent piece of AI-assisted work.
- Whether a given capture is a required record under Advisers Act Rule 204-2 is a determination for the firm's compliance lead, and the continuity argument does not depend on the answer.
Picture a client file eighteen months from now. The review note is thorough. The follow-up email was sent the same afternoon. The plan reflects the discussion. By every visible measure, the work was done well.
Part of it was done by software. The note began as a generated draft. The email started from a suggested reply. The advisor read both, changed a few things, approved them, and moved on. That is a reasonable way to work, and it is becoming the normal way.
Now put a different person in front of that file. A planner joining the relationship. A teammate covering during a leave. A successor, years on, trying to understand why the strategy took the shape it did. That person needs to do something specific: reconstruct how the work happened. And the file, complete as it looks, may not support it.
A firm's institutional memory used to be two things: what its people wrote down and what its people remembered. When a system does part of the work, memory acquires a third component, and it is the one nobody is in the habit of keeping. It is the record of what the system saw, what it produced, and what the person decided about it.
A firm that captures those three things can reconstruct its own work. A firm that captures only the finished artifact cannot, and no amount of asking around will recover the difference.
The finished artifact hides its own history
A polished paragraph does not announce where it came from. A note the advisor composed and a note the advisor approved read the same on the page. For most day-to-day purposes that is fine. The purposes where it stops being fine are exactly the ones continuity exists for.
Consider what happens when someone later asks a modest question: why does the note emphasize the education funding goal and barely mention the concentrated stock position the client raised in the same meeting?
If a person wrote the note, there is a person to ask, and human memory, for all its flaws, degrades gracefully. The advisor may not recall the sentence, but they recall the client, the season of the relationship, the reason a topic felt settled. They can usually rebuild the thinking well enough to be useful.
If a system drafted the note, the accurate answer to the same question may be that nobody chose the emphasis at all. The draft arrived that way, it read well, and it was approved. The advisor remembers approving it. What the advisor cannot remember, because they never knew it, is what the system was working from when it produced the draft. Which notes it drew on. Which prior meeting it summarized. Whether it saw the updated account information or last quarter's.
That knowledge did exist, for a moment, inside a running system. Unless it was captured at the time, it is not stored in anyone's memory, anywhere. Human recollection fades. This never forms.
Three questions a reconstruction has to answer
When another professional needs to retrace AI-assisted work, the reconstruction rests on three questions. They are worth stating plainly, because they are the specification for what the firm's record needs to hold.
What did the system see? Not in forensic detail, but in substance: which client information the draft was based on, and as of when. A summary built from a transcript is a different object from a summary built from the CRM history. If the record cannot say what the system was shown, the output floats free of its inputs, and its emphasis and omissions can never be explained, only observed.
What did the system produce? The suggestion as it stood before a person touched it. This matters because the gap between the draft and the final version is where the professional judgment lives. If only the final version survives, the human contribution and the machine contribution have been blended into one voice, and the next reader cannot tell which sentences carry an advisor's intent and which carry a model's fluency.
What did the person decide? Approved as written. Corrected the beneficiary detail. Rejected the framing and rewrote it. Struck the third paragraph because the client had asked to keep that discussion informal. These decisions are the human thread through the work, and they are usually the least documented part of it because they feel too small to write down. Later, they are the only evidence that judgment was applied, and the clearest guide to what the professional actually thought.
None of this requires a person to narrate their day. Most of it can be a byproduct of working, provided the tools and habits are set up to keep it. The failure mode is not effort. It is that nobody decided the trail was worth having until someone needed it.
Why reconstructable AI work is a continuity problem
It is tempting to file all of this under technology risk. The better frame is the ordinary one this publication keeps returning to: someone else, later, has to continue the work.
Firms already know what a fragile handoff looks like when the missing ingredient is human context. The record shows what happened but not why, and the team compensates by asking whoever was there. That compensation has always been the quiet backbone of continuity, and it has a known weakness: it fails when the person leaves.
AI-assisted work has the same weakness with a shorter fuse. The author of a generated draft is a version of a system, and versions change without ceremony. The tool that wrote half of this year's meeting notes may behave differently next year, or be replaced entirely. It cannot be asked what it was thinking. The only version of "asking the person who was there" that works for machine-assisted work is reading what was captured at the time.
There is also a subtler shift. When people did all the work, the firm's memory and its people's memories overlapped, so a thin written record was survivable; the firm still, in practice, remembered. As systems take on more of the middle, that overlap shrinks. More of what happened was never in anyone's head to begin with. The written record stops being a backup for institutional memory and becomes the institutional memory.
What a firm should start keeping
A firm does not need to rebuild its stack to act on this. It needs a few deliberate habits, applied where the work is consequential.
Mark the provenance where it matters. For material client work, the record should distinguish what was machine-drafted from what was human-authored. Not as a disclaimer, but as ordinary care with sources, the way a careful analyst distinguishes a source document from their own inference.
Keep the decision beside the output. A one-line trace is usually enough: reviewed and approved, or corrected in these respects, or rejected for this reason. The shorter the note, the more likely it gets written. What matters is that the human decision is recorded as an event, not implied by silence.
Note what the system was working from when the stakes warrant it. For routine drafting, this can be light. For work that shapes a client decision, the record should say, at minimum, which materials the system was given and as of when. If that question cannot be answered even in principle, that is worth knowing about the tool before it becomes worth explaining about the work.
Treat the tool's own history as convenience, not memory. Activity logs inside a tool are useful day to day, but they are not designed to be a firm's long-term memory. Whatever the firm would need in order to reconstruct the work later belongs in the firm's own record, in the systems the firm has approved for client information. Which of these captures count as books and records under Rule 204-2 is a determination for the firm's compliance lead, and this article does not make it. The continuity argument does not depend on the answer.
Run the reconstruction test. Take one recent piece of AI-assisted client work and ask someone who was not involved to answer the three questions from the record alone. What did the system see, what did it produce, what did the person decide. Ten minutes of this teaches more about the firm's actual memory than any policy discussion.
The practical implication: the default answer is that AI-assisted work is not reconstructable
None of this argues for slowing down. The drafting help is real, the time it returns is real, and advisors should take it. The argument is about what the firm keeps while it takes it.
Every firm is currently deciding, mostly by default, whether AI-assisted work will be reconstructable. The default answer is no, because the finished artifact is the only thing the workflow naturally preserves. Changing the answer costs little at the moment of work and is impossible afterward.
Institutional memory was never really about storage. It was about a firm's ability to stand in front of its own past work and explain it. People provided that ability for as long as people did all the work. Now part of the work has no person inside it, and the explanation has to be built at the time or not at all.
The firm remembers what it can reconstruct. Everything else, it merely stores.
Questions
What does a firm need to keep so AI-assisted work can be reconstructed?
Three things, captured at the time: what the system was shown and as of when, what it produced before a person touched it, and what the person decided, whether approved as written, corrected, or rejected. Most of it can be a byproduct of working if the tools and habits are set up to keep it.
Why is AI-assisted work a continuity problem rather than a technology problem?
Because someone else, later, has to continue the work. When people did all of it, the firm could ask whoever was there. A generated draft's author is a version of a system that changes without ceremony and cannot be asked what it was thinking. The only way to ask the person who was there is to read what was captured.
Does an AI tool's own activity log count as the firm's memory?
Treat it as a convenience, not as memory. Activity logs inside a tool are useful day to day but are not designed to be a firm's long-term record. Whatever the firm would need to reconstruct the work belongs in the firm's own record, inside the systems the firm has approved for client information.
What is the reconstruction test?
Take one recent piece of AI-assisted client work and ask someone who was not involved to answer three questions from the record alone: what did the system see, what did it produce, and what did the person decide. Ten minutes of that shows more about the firm's actual memory than any policy discussion.
Advisor Insights provides general professional information, not individualized investment, legal, or compliance advice. Records, supervision, and workflow conclusions are specific to each firm and its approved systems. For client-related information, use only systems and workflows approved by your firm.
Primary sources
General information from ValaisOS LLC, not legal, compliance, tax, or investment advice. Confirm requirements for your firm with counsel. See Terms of Use.