The Context Principle: Designing the AI User Experience

User experience (UX) plays a critical role in designing systems that enable workflows. Small inefficiencies can compound over time and create unnecessary drag on human outcomes. In a complex processing environment such as asset management operations, giving access to the right data views and the right features to modify or act on data is already a critical design principle.

From that design perspective, adding AI capabilities becomes another set of user actions and interfaces to manage.

The cost of leaving context

Trade breaks can help illustrate this point. Imagine you are an analyst tasked with resolving a set of breaks. You have to open internal records in your own system, pull incoming data from counterparties in a separate view, compare both sides of each break, identify the cause of the mismatch, and then move to another system to run a diagnostic or resolve it. Doing so means you are frequently switching context as you go.

Each transition is a context break. You have to keep in mind what you understand from each step as you go from system to system. When you’re familiar with the task, context switching becomes second nature. You notice it when you bump into additional challenges or when there’s an improvement that makes it easier.

There’s a problem with context switching, however. It adds time and risks accuracy. Each context switch introduces opportunities for data to be misread, miscopied, or stripped of the surrounding detail that made it interpretable in the first place. Researcher Gloria Mark found that it can take over 23 minutes to return to interrupted work.ⁱ In separate research, even brief mental blocks created by shifting between tasks can cost as much as 40% of someone’s productive time.ii

AI can either add more layers or streamline your work.

In a worst-case scenario, you’d have to grab all of that data, paste it into an AI-based tool, get the diagnosis and resolution, and go back to your core operating platform. Added steps aside, when you move a screenshot of counterparty data into a separate AI platform, the AI doesn’t see the workflow state, the prior exceptions on that account, or the broader pattern you were already tracking or encountered before.

The alternative is AI that operates where the data already lives, that is, inside the reconciliation workflow itself. Rather than asking you to provide context to the tool, it can fetch those screenshots or the underlying data on your behalf, then provide a diagnosis with a clear explanation. Essentially, you’re tasking it to do all that context switching for you and offer you a defensible result.

What “human in control” actually requires

That scenario shows how defensibility is an element of user experience. When AI surfaces a recommended resolution for a trade break, the ops analyst can review the rationale, evaluate it against what they know about the position and the counterparty, and then execute the data change.

When you design it so that the AI handles pattern recognition and data assembly, and the human applies judgment and takes consequential action, you’ve delivered a unified experience. It eliminates friction instead of adding to it.

The analyst needs to see why the AI proposed a specific resolution, what data it relied on, and how confident it is in its decision. That rationale should appear inline, in the same workspace where the analyst is already reviewing the break. If it takes additional context switching to understand the recommendation via a separate audit log or a different interface, the benefits of integrated AI start to erode. It’s integrated in principle, but not for the end user’s actual work.

Auditability adds another business rationale and is a natural byproduct of inline transparency. When the reasoning is recorded alongside the action, the audit trail builds itself. Six months later, a compliance review can trace exactly what was recommended, what was accepted or overridden, and why. And it doesn’t take complex forensics with system logs to do so.

Where rules end and judgment begins

Unified UX also helps raise the bar for what it means to have a human in the loop. Asset management operations rely on human judgment, and treating analysts as downstream consumers or validators of AI outputs does both people and your firm a disservice. 

In a confirmation-only model, the operator clicks approve and moves on. In a context-rich model, the operator sees enough to challenge, adjust, or reject. AI extends what an experienced analyst can accomplish within a given time window. They drive the loop without AI substituting for the judgment that makes their decisions defensible.

Some AI-enabled validation is straightforward. When a model ingests an agent notice for a private loan position via OCR, and the extracted value is off by a factor of 10, a threshold rule catches the error immediately. A dropped zero falls well outside any plausible tolerance band. You define that boundary condition once and enforce it consistently.

Other errors resist that approach. An OCR that confuses an “8” for a “9” could result in a value that falls comfortably within the normal range. A rules engine won’t flag it because the output looks plausible. And AI-powered OCR can introduce these same character-level errors, which means the tool itself is capable of generating the mistakes you’re relying on it to catch.

The operational design response is a dashboard that organizes exceptions by severity and confidence level, providing users a place to review and take action. Clear violations can route to automated handling, while ambiguous cases — where the value is plausible, but the confidence score is low — get surfaced for human review. Then the analyst can see what falls outside a defined threshold or fails a specific test condition to make the call.

Two design principles emerge from this approach to AI. First, provide inline recommendations, but don’t allow the AI to execute data changes without human review. Second, define context-aware validation rules, but recognize when they become complex enough to require human discretion.

Context as the bridge between probabilistic and deterministic

AI handles unstructured questions well. An analyst can ask a loosely phrased question about a position, a counterparty, or an exception pattern, and a well-built conversational interface will accurately interpret the intent.

But interpretation and operational reliability are different problems. The answer to an operational question needs to be consistent with the data model and grounded in the business rules that govern the underlying records. It also needs to be correct.

That precision depends on business context rather than interface context. Specifically, it depends on the AI having access to the schema, the ontology, and the business logic that define what the data means and how it relates to other data. That goes beyond the basic concept of an LLM. Without that structured context, an LLM defaults to probabilistic inference, generating plausible-sounding answers that may or may not reflect the actual state of a portfolio, fund, or counterparty relationship.

That aspect of AI can lead to misconceptions about whether AI can be trusted for operational work. It’s not an either/or choice, however. AI can use LLM capabilities and a well-defined ontology with a clear set of business rules at the same time. The combination produces answers that are deterministic where they need to be and auditable throughout.

Context is what resolves the either/or. Strip it away, and you get hallucinations. Provide it, and the outputs become something an operations team can act on with confidence.

AI as interface, not replacement

In investment operations, AI needs to be thought of as an interface layer that demands careful design, just like the end-user interface that operators use every day. AI surfaces relevant data, recommends actions, and accelerates workflows that would otherwise consume analyst hours in manual comparison and diagnosis. The data governance, the structured context, and the human judgment underneath that interface are what make it trustworthy.

AI that preserves context across the complexity of data across asset classes, counterparties, and stakeholders can take friction out of workflows for both the user working inside the workflow and the system reasoning over the data. But it takes thoughtful design. Otherwise, if AI breaks context, it becomes another source of operational risk and end-user burden.