Clarista White Paper · Part Two
Four layers, and the order you build them in. Most firms aim at the layer that sounds most exciting. The firms that arrive start at the one that makes the rest possible.
Picture a demo that goes well. The data is consolidated. The assistant is fluent, it reads documents accurately, it drafts a decent follow-up email. The room is impressed.
Then somebody asks the question they actually came with. Given everything we know about this family, what should we be doing that we are not?
The answer arrives in four seconds, reads well, and would be equally true of any family in America.
Part One explained why that happens. The system was grounded in records of what happened, your records of what happened look like everyone else's, and the reasoning that makes your firm different was never written down anywhere it could be read.
Firms agreed, and then asked the obvious next question. Where do we actually start?
That deserves a better answer than a picture of the finished state.
Before the layers, the thing that causes most of the confusion.
Part One said to start with the question that matters, work back to the decisions behind it, build the applications around those decisions, and land the data last, so it arrives with a job to do. That is still right, and it is a rule about design.
Construction runs the other way. You cannot record judgement on top of numbers your firm argues about, and you cannot capture a process that still lives in a spreadsheet. Building runs bottom up.
Both rules are true at once because they answer different questions. The question tells you which slice of each layer you need. Then you build that slice from the bottom. A firm that designs bottom up ends up with a warehouse that sits unused, which is the mistake of the last twenty years. A firm that builds top down ends up with signals people stop reading by March, which is the mistake being made right now.
Four questions. The first one you answer with a straight no is the layer you are on. Read that section first.
Can you answer any question about a household from one place, and does the answer agree with itself across reports?
Does the work happen inside your systems rather than in spreadsheets beside them?
For a decision made eighteen months ago, can you show the reason without asking the advisor who made it?
Does your firm know where its advisors decide the same question differently?
Most firms we meet answer no at question one or two, and arrive wanting to talk about question four.
Each section below has the same shape, so you can walk your own firm through it. What the layer is. What it looks like when it is missing. What AI does here. And a plain test for when you are done.
Every vendor sells this layer, and most firms believe they have finished it. (What a finished one actually looks like: a governed data fabric that serves one trusted view without copying sources.) Aggregating sources is table stakes, and several firms will do it for you competently.
What makes the layer load bearing is the work after the aggregation. A household has to be one thing rather than nine account records that happen to share a surname. Your firm has to agree what each term means, so two reports stop disagreeing about the same number. Values have to be checked against their sources rather than assumed. And something has to notice the moment a feed breaks, rather than a client noticing first.
What it looks like when it is missing
People reconcile by hand before every meeting. Two reports give two answers, and the firm has learned which one to trust for which purpose. A question that crosses two systems takes a day, so it stops being asked. The answer to what this family actually owns lives in somebody's working file.
What AI does here
Four things, and they are the reason this layer takes months now rather than the years it used to take.
This is the layer of the work itself. Meetings, trades, onboarding, reviews, approvals. The record of what the firm did.
Every firm has systems for this, which is why the layer gets skipped in planning. The problem is rarely that the record is absent. It is that the systems were built for somebody else's process, so your firm has spent years fitting itself to them. A field used for something it was never designed to hold. A workaround everyone knows about. A step that happens in a spreadsheet because the system has no place for it.
The work here is bringing the technology to your process rather than the other way round. Fewer systems, each closer to how you actually work, able to change at the speed your business changes rather than at the speed of a vendor's release cycle.
What it looks like when it is missing
The real process lives in spreadsheets beside the system. New people learn the method by sitting next to someone. A task that should take an hour takes a day because three systems have to be opened and reconciled. And the firm cannot change how it works without asking a vendor first.
That firm is worth describing, without naming it. The spreadsheets held the pieces: which accounts were managed and which were held under oversight, what the family needed in cash and when, the asset allocation at household level and again at account level, and the risk the family said they would accept against the risk actually sitting in what they owned. Every piece existed. The connections between them lived in the people who maintained the files.
Connecting those pieces into one view per household removed the reconciliation. The more interesting consequence came next. Trading at that firm had been an entirely manual process, and the connected household view turned out to be exactly what a trading process needs as its input. It became a one click generation followed by a review, which is a different job from the one the team had been doing.
What AI does here
One caution, because it is the honest version. Start in the gaps between the systems you own, where your method lives and where no vendor is selling you anything. The boundary moves outward on its own: a decision system usually turns out to need a piece of record keeping to work properly, and a trading application needs positions at tax lot level through the day rather than at the close, so the bookkeeping comes with it. Extend when that happens, with evidence. Setting out to replace everything on day one is a different project, and a worse one.
Here the paper stops describing plumbing and starts describing the part worth paying for.
This layer holds two things that are easy to confuse. The plan is the intended approach for a family, built from their situation and the environment around them. Most firms have some version of that already, sitting in a planning tool. The judgement is the running record of what the firm decided against that plan, and why. Almost no firm has that, and it is the thing Part One was about.
In working form it looks like this. The system watches the plan against the world and raises a signal when something requires a change or suggests one. The advisor decides. And whatever they decide, the reason is recorded.
The three answers matter more than they look. A system that offers accept or dismiss throws away the most common case of all, which is the signal that was correct and where the right answer was still to do nothing. An advisor who leaves a concentrated position alone because it is already earmarked for a gift in the fourth quarter has made a real decision for a real reason, and that reason exists in no system in the firm.
Delay is the third, and it is the one firms forget to design for. Much of advice is correct and badly timed. Recording that something was right but early, with the condition that makes it ripe, is how a firm stops losing good ideas to a busy week.
What it looks like when it is missing
A client of fourteen years asks why the firm did something in 2019, and the room assembles an answer on the spot. A departing advisor's households go to a colleague along with the outcomes and none of the thinking. An examiner asks the basis for a recommendation and the file holds the recommendation.
What AI does here
The fourth layer works across the other three. It reads the transactions, the plan and the judgement together, and does on your behalf what a very good analyst would do with time to read everything about every family every night.
The important property is that it learns, and it learns at three levels.
At family level it knows what this household has decided before and why, so it stops repeating itself and starts noticing when a stated reason has expired. At advisor level it learns how a particular advisor decides, where they move quickly and where things sit. At firm level it finds the patterns across everybody: where six advisors decided the same question nine different ways, and where a reason appears often enough that it should stop being a judgement and become firm policy.
That last one is the point of the whole exercise. A firm that can see its own practice written down can choose which parts to make standard, and teach the rest to the people who joined last year.
What AI does here
Everything here is model work, and it is also the layer that cannot be bought ahead of the others, because the material it runs on is produced by layer three. A vendor offering it on day one is describing something that cannot exist yet.
Layers three and four are easy to read as two more steps, and they are not. They are the two halves of one loop. Layer three writes the record. Layer four reads it. You cannot read before you have written, which is why three comes first and why four cannot be bought early. Once both are turning, every pass makes both better.
That is the same shape Part One described in somebody else's words. Ray Dalio wrote down the criteria before the trade, reflected on how they had worked once it closed, turned what he learned into rules, and ran the rules alongside his own judgement. The writing is layer three. The reflecting is layer four. He ran them as a loop for fifty years, and the loop is what produced the system, rather than either half on its own.
What comes back from layer four is specific, and worth naming rather than calling it a feedback loop.
The one worth saying out loud in a meeting is what that adds up to for an advisor.
That is the opposite of what every prospect braces for when a vendor says the word signals, and it is the honest description of what compounding feels like from an advisor's chair.
There is a second circuit, and commercially it may be the more useful one.
Your record of reasoning is the best backlog your firm will ever have for layers one and two. When an advisor declines a signal because of a fact the system does not hold, that is a gap in layer one, reported by the person who felt it rather than guessed at in a workshop. When the same manual workaround turns up in reasons across twenty decisions, that is the next application to build in layer two, identified by evidence.
So the picture is a loop between three and four, with spokes running down into one and two. A firm two years in stops guessing what to build next, because its own advisors have been telling it, one decision at a time.
Read the four layers as a staircase your whole firm has to climb and the project becomes a three year programme with nothing to show for eighteen months of it. Few firms will start that, and they are right.
The layers are a path, and the thing that travels the path is a process rather than a firm.
In practice that means picking one process where your method differs from what your platform assumes. You build the slice of layer one that process needs, which is a fraction of your data rather than all of it. You build the application in layer two, with the place to record why built into the work. Layer three starts producing a record from the day it goes live. And layer four has something to learn from on that process long before the rest of the firm has caught up.
Then you do the next process, and the second one is cheaper, because half the layer one work is already done and the application pattern already exists.
A firm two years into this does not have four finished layers. It has five or six processes running deep and the rest running shallow, and it knows exactly which is which. That is a healthier picture than a maturity score, and it is the one we see at firms that are actually getting somewhere.
The failure is worth describing, because it is the common one and it looks sensible on the way in.
A firm reads an argument like Part One, agrees with it, and decides to start capturing reasoning. They buy or build something that raises signals and asks advisors for a reason. The data underneath is the data they already had.
Within a month, advisors are arguing with the signals, because the numbers in them disagree with the numbers in the report they trust. A signal argued with twice stops being read. The reason box asks for a second job in a second place, which advisors will do for about six weeks. By the end of the quarter the capture rate is low enough that the record is unusable, and the firm concludes that thinking systems do not work.
The firm in Figure 2 did it the other way round, deliberately. They took layers one and two first, because those are the foundation for capturing reasoning, and reasoning is what drives judgement. That was speed rather than caution. Every piece of layer two they built started producing the record layer three needed, from the day it went live.
Four deliverables, two of them platform and two of them service, mapped onto the four layers.
The data platform carries layer one. Aggregation is table stakes, so the work is linking entities into households, defining what your firm means by each term, verifying values against their sources and monitoring for breakage. It lands in a data lake you own, on your own infrastructure.
Reasoning systems and custom apps carry layers two and three. One process at a time, where your method differs from what the standard tool assumes, with the place to record why built into the work rather than bolted onto it. You own the applications.
Agents carry layer four, and they come in two kinds (how an agent inherits a named person’s permissions is covered in our paper on permission-aware data access). Productivity agents work from the data and arrive first: meeting preparation, portfolio analysis, tax analysis. Every firm will have those within a couple of years. Reasoning agents work from the data and the recorded judgement together, and they exist only at a firm that has been recording.
Hosting and monitoring runs underneath all of it, with access management, observability, and a record of what each agent did and on whose authority.
Run the four questions with your operating committee and find the first no. Then pick one process on that layer where your method differs from what your platform assumes, and take that one process down through all four layers, with the reasons captured from the first day it runs.
One process built on a foundation you trust produces more in a year than four built on one you do not.
Sometimes, and the test is specific rather than general. If the facts a given signal needs are reliable and agreed, build that signal while the rest of layer one is still being finished. What fails is recording reasoning across the whole firm on data the firm argues about, because the arguments land on the signals and people stop reading them.
For one process rather than a whole firm, layer one and the first application of layer two usually run in parallel over a matter of months. Layer three produces a usable record as soon as that application is live. Layer four says little useful in the first year on that process, and becomes the reason you did the rest by the third.
Start by assuming you keep them, and build in the space between them. The boundary moves outward on its own as decision systems turn out to need record keeping, and that is a decision to take later with evidence rather than at the start on principle.
An alerting system tells you something and records whether you dismissed it. This records why. The decline and the delay are where a firm's method actually sits, and they are what make the fourth layer possible.
It has to be good enough for the one process you picked, and agreed by the people who will use it. Perfect data for the whole firm is the goal that keeps firms at layer one for three years.
The client owns the data lake and the applications. Hosting, monitoring and the agent configuration run on the Clarista platform.
An AI maturity model describes the stages an organization moves through as AI goes from experiment to core capability. In this framework the stages are four layers: a governed data layer, transaction capture, planning and judgement capture, and finally a thinking system that can apply the firm's own method. Each layer produces the material the next one needs, which is why skipping ahead fails.
A practical adoption framework for a wealth or advisory firm is: design backwards from the question that matters to the business, build forwards from the data foundation, and take one process down through all four layers rather than moving the whole firm up one layer at a time. A firm can be at layer four on liquidity while still at layer one on everything else, and that is the correct shape of progress.
Assess each layer separately. Layer one: can your systems serve one consolidated, governed view of a client household? Layer two: are transactions and their context captured where a system can read them? Layer three: is the reasoning behind decisions written down at the moment of decision? Layer four only exists if the first three do. Most firms discover they are at layer one, and that is the honest starting point.
The data layer, always. It is the least exciting layer and the one that makes every other layer possible. One firm in this paper replaced roughly 700 spreadsheets with a single connected view of each household, and that unglamorous consolidation became the input to automating a trading process that had been entirely manual.
Choose which categories of cookies and tracking technologies you allow. Strictly necessary cookies cannot be disabled. Read our full Cookie Policy.
Required for security, authentication, fraud prevention, load balancing, and storing your consent preferences. Always on.
Remember choices you make to provide a personalized experience (language, region, theme, saved filters).
Help us understand how the site is used (Google Analytics, Vercel Analytics, Microsoft Clarity, etc.).
Used for conversion tracking, retargeting, and personalized ads (LinkedIn, Meta, Google Ads, Bing, etc.).