Situated cognitive guidance: A new interaction pattern for human-in-the-loop workflows

What is situated cognitive guidance (SCG)?

A cognitive interaction pattern for live digital workflows refers to a mode of human–system interaction in which an AI system understands the live operational context of a task (interface, state and workflow) and supports human decision-making by framing actions, interpreting states and sequencing steps, without executing actions on behalf of the user.

The system does not replace human agency. Instead, it offloads cognitive load by absorbing ambiguity, maintaining the procedural model of the task and returning only what is contextually relevant and actionable at each moment, in line with institutional approaches such as those described by the US National Institute of Standards and Technology (NIST) in its AI Risk Management Framework.

Unlike automation systems that replace execution, and unlike copilots that act on behalf of the user, Situated Cognitive Guidance operates at the level of interpretation. It does not optimize processes; it stabilizes understanding. This perspective resonates with the machine-in-the-loop approach described in design and human-centered machine learning, where systems are designed to support human cognition and decision-making rather than automate execution, emphasizing continuous human engagement rather than delegation. Its function is not to accelerate workflows, but to make them cognitively reliable.

Situated Cognitive Guidance is domain-agnostic and technology-independent: It does not rely on a specific interface, platform or AI implementation, but on the presence of a live operational context and a human-in-the-loop task, as defined in industry practice, for example by IBM.

An important aspect of this pattern is that it operates on two complementary surfaces.

First, it functions over external applications: Browsers, platforms and any system where the AI can share visibility of the live interface and its operational state. In these contexts, Situated Cognitive Guidance is expressed as real-time interpretation of workflows, interfaces and procedural ambiguity.

Second, it also operates over the conversational space itself. The chat is not only a communication channel, but an active cognitive surface where understanding is refined, reformulated and stabilized. This creates a feedback loop between action and reflection: What is discovered while operating in external systems reshapes how the dialogue is structured, and what is clarified in the dialogue reshapes how the system is used in practice.

Situated Cognitive Guidance, therefore, does not exist only “in the browser” or only “in the chat”. It emerges from their interaction. One grounds cognition in real operational contexts; the other consolidates and reorganizes that cognition through language. Together, they form a continuous feedback cycle between execution, interpretation and conceptualization.

Minimal architecture of the experience (operational context)

This is intentionally a minimal architecture description. It does not define a technical stack or implementation blueprint, but isolates the minimum conditions required for the cognitive pattern to emerge, independent of specific technologies.

This experience does not occur in “ChatGPT” in isolation, but belongs to the broader field of human-AI interaction and human-centered AI, as explored in research published by ACM SIGCHI, where the focus is placed on designing systems that preserve human agency, cognition and contextual awareness.

It emerges when three elements are combined:

1. Extended‑context AI capability

The system must be able to process extended contextual input beyond isolated prompts, including complete interfaces, long workflows and ambiguous or transitional states.

ChatGPT Plus is one concrete example of this capability, but the pattern itself does not depend on a specific product or plan.

Without extended contextual capability, the system collapses into a prompt–response model and loses any notion of workflow continuity. Guidance becomes fragmented and episodic rather than procedural.

2. Atlas (integrated browser)

Atlas is the environment where the AI shares space with the real web.

Human and system are literally looking at the same operational surface: The same page, the same fields, the same states.

The AI does not interpret descriptions: It observes the active page and its actual state. The AI does not imagine the interface — it sees it.

Without a shared operational surface, cognition becomes speculative: The system reasons about descriptions instead of reality, increasing interpretation errors.

3. Side Chat (contextual chat)

Side Chat keeps the web visible while the AI reasons in parallel.

Each message includes the live context of what is on screen, allowing the system to interpret the current point in the workflow, distinguish real blockers from informational warnings, and understand intermediate states.

The AI does not execute actions or navigate on behalf of the user, but it understands the operational state of the system with precision.

This establishes continuous situational grounding: A persistent alignment between the system’s reasoning and the user’s live position within the workflow.

The result is not automation. Without continuous situational grounding, guidance becomes stale and detached from the real state of the task, breaking alignment.

Operational conclusions of the observed pattern

The following conclusions are derived from direct observation and practical use, not from prior theoretical design. They reflect patterns identified while interacting with real systems and workflows in situ.

Externalization of operational logic (not of action)

The system does not act or execute workflows on behalf of the user.

Its function is to maintain the mental model of the process, interpret intermediate states, and absorb ambiguity, returning only what is actionable.

The human remains the actuator. The AI operates as a situated cognitive guide, not as an executor.

For example, when completing a multi‑step form, the system clarifies which warnings are non‑blocking and which fields can be safely left unchanged, allowing the user to proceed without re‑evaluating the entire workflow.

Highest‑impact case: Parameterized repetition

(Repetition of an action with variable parameters)

The pattern is especially effective when a stable workflow must be repeated multiple times with bounded changes (text, category, date, identifier, attachments).

In these cases, the system:

• Maintains the complete cognitive sequence
• Indicates what remains invariant
• Highlights what changes in each iteration
• Adapts to the real state of the interface

It does not automate the task. It cognitively orchestrates repetition.

Primary value: Elimination of doubt

The benefit is not speed, but clarity. From a cognitive perspective, this aligns with the principles described in Cognitive Load Theory, which explain how limitations in human working memory make poorly structured workflows cognitively fragile and error-prone. By structuring interaction as guidance rather than raw output, SCG reduces unnecessary cognitive load and stabilizes human reasoning.

It avoids:

• Reinterpreting the workflow at each step
• Making errors due to attentional fatigue
• Carrying the entire plan in working memory

Friction does not disappear: It is displaced outside the user’s cognition.

When doubt is removed, action becomes obvious — even if the task itself remains complex.

Difference from classical automation

Traditional automation requires stable workflows, fixed interfaces and pre‑modeled exceptions.

SCG works precisely when:

• The interface changes
• Exceptions appear
• Ambiguity exists
• Human judgment remains necessary

Crucially, SCG does not compete with automation: it operates on a different cognitive layer. Automation replaces execution where the process is known; Situated Cognitive Guidance supports human reasoning where the process must still be interpreted.

This distinction is already visible in contemporary systems such as Microsoft Copilot, which are designed to guide users through complex workflows, clarify decisions and propose next steps without removing human control or responsibility.

It is an adaptive approach, not a rigid one. Automation is designed for environments where meaning is fixed and behavior is predictable. Situated Cognitive Guidance exists precisely where meaning must still be interpreted. Automation assumes clarity; SCG exists because clarity is missing. One replaces execution when the process is known; the other stabilizes cognition when the process must still be understood.

Applicability

The pattern is transversal across domains: administrative platforms, backoffices, e‑commerce, dense professional tools or legacy systems.

Situated Cognitive Guidance is effective wherever workflows exhibit high cognitive density and state ambiguity.

Limits & conclusion

It works best when:

• The objective is clear
• Parameters are defined
• The interface is visible and shared

It does not replace deep strategic decision‑making, open‑ended creativity or tasks without an observable interface.

Some of these ideas build on earlier work I developed in the Threshold framework, where I explored how emerging interaction paradigms reshape human experience, cognition and the way people engage with complex systems.

I have realized that I am not using a single “ChatGPT”, but three cognitive layers simultaneously:

• Side Chat guides my action in real time: it interprets the state of a live interface, clarifies ambiguities and stabilizes decision‑making while I act.
• Canvas structures my thinking: it turns experience into formalized, revisable and reusable knowledge.
• Normal chat reflects on both: it allows me to understand what is happening, why it works, and how to name it.

This is not just a conversational interface. It is an architecture of distributed cognition:

• Action
• Formalization
• Metacognition

…happening in parallel.

The AI not only helps me do things. It helps me observe how I think while doing them.

And that loop — action, understanding, formulation — is where something new emerges: Not productivity, but structural clarity of one’s own thinking.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?