Most AI tools in wealth management fail for a boring reason. They arrive with plenty of intelligence and almost no working context.
A note-taker can survive that. A real workflow cannot.
This is where the market is getting confused. Leaders keep treating AI performance as a model question, a vendor question, or a procurement question. In live operations, the bigger issue is usually simpler. The system does not know enough about the job it has just been asked to do.
Charles Schwab's 2026 RIA and AI research found 63% of advisers are already using AI tools, with standalone tools and administrative use cases leading adoption. EY's 2025 wealth and asset management survey found 95% of firms had scaled GenAI to multiple use cases, yet only 27% reported substantial business impact. Activity is rising. Dependable workflow value is harder to find.
That gap is rarely about raw model quality.
It is a context gap.
The system is not one thing
Most firms still talk about "using AI" as if they are buying a very smart employee.
The better analogy is a very fast junior dropped into the middle of annual review season with no clean client file, no house view on what good looks like, and no idea who signs off when something looks odd.
You will get output. You will get speed in pockets. You will also get hesitation, rework, and the kind of polished uncertainty that makes people lose trust quickly.
That is the pattern.
The tool is not missing intelligence. The workflow is missing context.
This is the part many teams still underinvest in because the demos hide it. The demo has tidy records, a narrow prompt, a known task, and no awkward exception halfway through. Production has none of those courtesies.
We have seen this in advice, wealth, mortgage and adjacent financial services work. The first pilot works because humans around it are quietly supplying the missing context themselves. The second or third use case stalls because the workflow now needs that context to be explicit.
That is where the Context Stack matters.
The Context Stack
Most AI workflows that hold up in production have four layers of context operating together.
Miss one and the others start compensating badly.
1. Record context
Record context is the factual base the system is allowed to treat as real enough to work from.
Client details. Risk profile. Product holdings. Prior advice. Review date. Fee consent status. Source documents. The ordinary records a human team uses every day, usually with more workarounds than anybody likes to admit.
This is why your CRM becomes AI infrastructure. The moment an AI system reads from those records, messy administration data turns into a production risk.
Record context is about more than completeness. It is about authority.
Which system wins if two values conflict. Which field is current. Which document outranks the CRM note. Which records are usable as evidence and which ones are merely hints for a human to follow up.
Without those decisions, the model fills the gaps with fluent guesswork.
That is how a workflow starts looking smart while getting dumber.
2. Rule context
Rule context is the boundary around what good, safe, acceptable work looks like in your firm.
This includes compliance obligations, yes. It also includes house style, product rules, licensee settings, review standards, escalation thresholds, approval limits, and the little bits of judgement that teams stop noticing because they have internalised them over years.
Most AI tools arrive with generic competence. Financial services workflows do not run on generic competence.
They run on specific rules applied consistently.
ASIC's Report 798, released on 29 October 2024 after reviewing 23 licensees and 624 AI use cases, warned that a governance gap could emerge as adoption runs ahead of oversight. That warning lands hardest when teams assume policy can be bolted on after the tool is chosen.
By then the workflow has already learned the wrong habits.
Rule context is what keeps an AI system from confusing a plausible draft with a defensible one.
3. Workflow context
Workflow context is where the system sits in the actual sequence of work.
This sounds obvious. It gets missed constantly.
"Draft the review summary" is a task. It is not workflow context.
Workflow context answers the questions underneath the task. What happened one step before this. What is meant to happen next. Which pieces are mandatory. Which ones are optional. What standard the output has to meet before the hand-off. Which hand-off matters.
This is where many pilot projects flatten out.
The tool performs one isolated task well enough to impress the room. Then the business tries to place that task back into the real operating sequence and discovers the surrounding workflow is full of exceptions, dependencies, and silent assumptions nobody had written down.
That is why the first gain from AI often shows up in back-office or controlled functions. EY's survey found early cost savings showing up most clearly in compliance, risk management and IT infrastructure. Those environments tend to have tighter workflow context. The boundaries are clearer. The hand-offs are more legible. The standards are easier to name.
Wealth firms that want front-office value need to do the same work for client-facing workflows, even if it feels less exciting than the demo.
4. Escalation context
Escalation context defines what the system should do when reality stops matching the happy path.
Conflicting records. Missing file notes. Advice language drifting too close to personal recommendation. A client instruction that changes the task halfway through. A transaction pattern that feels off. A workflow step the system cannot complete with confidence.
Good teams decide this before the tool goes live.
The system stops here. It routes there. This person reviews it. That evidence is attached. The clock starts now.
Poorly designed AI workflows treat escalation as a failure state.
Mature ones treat it as normal operating infrastructure.
That distinction matters because human trust depends less on whether the system is perfect and more on whether the system knows when to stop.
AUSTRAC's updated regulator statement of expectations on 21 May 2026 made the broader operating principle very clear. Regulated businesses are expected to document risks, document controls, and show sustained progress against implementation plans where reforms are still being worked through. That is a useful standard for AI workflows as well. The business does not need magic. It needs a visible route through uncertainty.
What the assembled system changes
Once these four layers are in place, the whole workflow changes character.
The model matters less. The organisation matters more.
Suddenly the tool does not need to be brilliant in the abstract. It needs to be reliable inside a defined lane. That is a much easier standard to meet, and a much more valuable one.
This is also why firms sometimes buy an impressive AI product and still feel underwhelmed six months later. They bought intelligence. They did not supply the operating context that would let intelligence turn into dependable work.
The first draft looks fine. The tenth exception exposes the gap.
This sits underneath the clearance problem as well. Clearance answers what the system is allowed to do. The context stack answers what the system needs to know in order to do it properly.
Both matter.
This is the free consulting bit
Take one workflow you already care about. Annual review preparation is usually a good candidate because the pain is obvious, the repetition is high, and the inputs cut across multiple systems.
Then test it against the four layers.
- Record context: Which systems and fields count as the source of truth for this workflow?
- Rule context: Which policy, practice standard, approval rule, or house view changes what "good" looks like here?
- Workflow context: What is the step before this one, the step after it, and the exact hand-off standard in between?
- Escalation context: What are the three most likely exceptions, and where does each one go?
If the answers live in people's heads, the workflow is not ready for AI that does real work.
If the answers are split across five documents and two Slack threads, the workflow is also not ready.
Write them down in one place.
Do it for one workflow only. Do not start with a firm-wide AI policy refresh. Start with the workflow that keeps making your team late, tired, or inconsistent.
Most firms do not need a larger AI ambition. They need one use case with enough context around it to survive contact with reality.
That is the work.
The uncomfortable implication
There is a reason so many AI programmes keep circling around note-takers, meeting summaries and draft emails.
Those tools can create visible value while demanding very little context from the business.
The harder, more valuable workflows ask the firm to reveal how much of its operating model still exists as tribal knowledge, loose judgement and half-settled exceptions. AI did not create that mess. It does expose it quickly.
That is why context work feels slow right up until the moment it starts compounding.
Once record authority is clear, rules are explicit, the workflow is mapped, and escalation is designed, the business can swap tools, test models, and automate more confidently because the operating context stays with the firm.
That is the asset.
Models will keep improving. Vendors will keep multiplying. Product demos will keep looking smoother than production.
Context is the part that turns software into work.