What 123 million AI tokens teach us about workflow design: Build systems that make reasoning visible

After 123 million tokens, the biggest lesson is not “use more AI.” It is “build a better system around it.”

A modeled agentic workflow running 2,000 completed jobs per month can reach 123 million tokens before failed jobs, observability overhead, embeddings, external services, and engineering effort are counted. That number is useful because it strips away the novelty. At this scale, weak workflow design becomes expensive, difficult to inspect, and surprisingly hard to improve.

For a strategy consultant, the problem is rarely a shortage of generated text. The real problem is that context gets scattered, assumptions disappear between prompts, competing outputs are hard to compare, and the final recommendation loses its connection to the evidence that produced it.

The productive question is therefore not, “How do we generate more?” It is, “How do we make every stage of the reasoning visible, reviewable, and reusable?”

Why does more prompting fail to produce better thinking?

More prompting can increase volume without improving the quality of the decision. It often creates a larger pile of plausible material while leaving the underlying reasoning process unchanged.

Research on long-context language models has shown that simply providing more context does not guarantee that every relevant detail will be used consistently. Performance can decline when important information is buried in the middle of a long input. That matters in professional work because the missing detail is often not obvious. The answer may still sound polished.

Three failure patterns appear repeatedly:

1. Context dilution. Important constraints sit beside background material, copied notes, and prior outputs. The system receives more information but a weaker signal.
2. Output drift. Each new prompt subtly changes the framing. By the fifth iteration, the response may be solving a different problem from the one the client originally raised.
3. Reasoning loss. The final recommendation survives, but the assumptions, rejected alternatives, and unresolved questions that shaped it do not.

This is why token volume is a poor proxy for intelligence. Tokens measure activity. They do not measure whether the workflow retained the right context, challenged the right assumptions, or produced an inspectable chain of judgment.

What should an AI workflow preserve?

A useful AI workflow preserves more than the final answer. It retains the working structure that lets a professional review how the answer was formed.

For 250 years, consequential ideas have depended on people who could structure complexity, challenge assumptions and make the path forward visible.

That long-standing decision discipline still applies. The tools have changed; the need to organize competing concerns has not.

A durable workflow should preserve six things:

Workflow layer	What it must retain	Why it matters
Context	Source documents, data, prompts, constraints, and definitions	Prevents later stages from operating on an incomplete brief
Task definition	The exact question, intended audience, and required output	Keeps generation aligned with the professional objective
Routing	Which reasoning path, visual format, or model setup handled each task	Makes the workflow understandable instead of mysterious
Comparison	Differences, agreements, omissions, and conflicting assumptions	Prevents the first fluent answer from becoming the default
Verification	Claims checked, evidence reviewed, and uncertainty marked	Separates plausible language from defensible analysis
Retained reasoning	Decisions, rejected paths, dependencies, and next questions	Makes the work reusable in later meetings and revisions

Verification deserves its own stage. Research on chain-of-verification methods found that asking a model to generate and answer targeted verification questions can reduce factual errors compared with accepting the first response. The broader lesson is simple: verification should be designed into the workflow, not added as a nervous final check.

How should tasks be routed according to reasoning needs?

Not every task deserves the same AI setup. A workflow becomes clearer when each stage is routed according to what the work actually needs.

A first-pass summary may need one concise reasoning path. A strategic choice may benefit from multiple perspectives. A process problem belongs in a flowchart. A decision with competing criteria belongs in a matrix. A large body of notes may need a mind map before anyone can see the themes.

For strategy work, a practical routing model looks like this:

• Use a flowchart when sequence, handoffs, gates, or dependencies determine the outcome.
• Use a matrix when alternatives must be compared against explicit criteria.
• Use a mind map when the first challenge is understanding the breadth and relationships of a problem.
• Use a diagram when the work depends on systems, actors, information flows, or structural relationships.
• Use multiple reasoning models when the cost of a blind spot is higher than the cost of comparison.
• Use web-grounded research when the recommendation depends on information that may have changed.
• Use document or data analysis when the evidence already exists in files and should remain connected to the output.

Visual formats are not decoration here. Research on visualization and decision-making describes diagrams as a way to externalize a decision problem into explicit options and tasks. Once the structure is outside a person’s head, it becomes easier to inspect, discuss, and revise.

This is where the Jeda.ai visual intelligence workspace fits the workflow. It can turn prompts, documents, data, sticky notes, and web research into editable matrices, mind maps, flowcharts, diagrams, and infographics on a shared canvas. The professional still owns the judgment. The workspace makes the reasoning easier to see.

How-To 1: Build a reusable AI workflow with the AI Menu

The AI Menu method works best when the task benefits from a guided structure. It reduces the chance that every consultant on a team invents a different prompt sequence for the same type of engagement.

Step 1: Define the decision before choosing the recipe

Write the decision in one sentence. Include the intended audience and the form the recommendation must take. For example: “Determine which operating approach should be tested first and show the trade-offs in a decision matrix.”

Step 2: Open the AI Menu and choose the visual structure

Select a category that matches the reasoning need. Use Matrix for criteria-based comparison, Diagram for relationships, Flowchart for sequence, or Infographic for a concise visual summary.

Step 3: Complete the guided context fields

Add the project, audience, objective, constraints, internal factors, external factors, and any additional context requested by the recipe. The exact fields vary, but the discipline is consistent: define the work before generation begins.

Step 4: Attach the evidence where relevant

Use document or data analysis options when the recommendation should be grounded in uploaded material. Turn on Web Search when current external information is necessary. Keep source material attached to the same workspace so the visual output remains connected to its evidence.

Step 5: Choose the reasoning setup

Use a single reasoning model for straightforward work. For a decision with meaningful ambiguity, enable Multi-LLM and select up to three reasoning models. When two or more are selected, use an aggregation model to combine or evaluate the outputs.

Step 6: Generate, inspect, and annotate

Review the output cell by cell or node by node. Mark assumptions, missing evidence, contradictions, and questions that require professional judgment. Do not treat completeness of layout as proof of correctness.

Step 7: Retain the workspace as a reusable method

Keep the completed visual, source material, notes, and decision trail together. Duplicate the workspace for the next comparable engagement instead of rebuilding the method from scratch.

The AI Whiteboard capability page documents the editable canvas, visual commands, file-based analysis, collaboration, and transformation features that support this kind of reusable workflow.

How-To 2: Build a custom AI workflow from the Prompt Bar

The Prompt Bar method is better when the workflow is specific, experimental, or not well represented by an existing recipe. It gives the consultant direct control over the task definition and output format.

Step 1: Create a compact context block

State the situation, audience, objective, constraints, evidence available, and what must not be assumed. Keep this block stable across iterations. Changing the context every time makes output comparison meaningless.

Step 2: Select the command before writing the full prompt

Choose Flowchart, Matrix, Mindmap, Diagram, or another command based on the decision structure. The output format should match the reasoning task, not merely the preferred visual style.

Step 3: Write an explicit output contract

Specify the required stages, fields, criteria, labels, and unresolved questions. An output contract makes it possible to compare results because each response is solving the same defined task.

Step 4: Add the appropriate grounding

Attach documents or data when those materials are the evidence base. Turn Web Search to Auto or On when current information is required. Separate sourced facts from working assumptions in the prompt.

Step 5: Run one or several reasoning paths

Use a single model for low-ambiguity work. Use Multi-LLM when you need competing interpretations, alternative structures, or a stronger challenge to the initial framing. Compare the outputs before synthesis.

Step 6: Convert the best reasoning into the right visual

Use Vision Transform when the first output reveals that another format would be more useful—for example, converting a mind map of issues into a matrix of choices or a set of sticky notes into a flowchart.

Step 7: Extend without losing the existing context

AI+ can extend or deepen selected visual content while keeping the current structure in play. It is most useful after the core workflow is already visible, not as a substitute for defining the original task.

The release notes for Web Search and AI+ describe how live grounding and context-preserving extension operate inside the workspace.

Example prompt for an inspectable workflow

The prompt below is deliberately structured. It defines the decision, the workflow stages, the required artifacts, and the visual output. That makes it easier to compare responses and harder for the work to drift.

Create an inspectable AI workflow for evaluating a new team collaboration feature. Use a horizontal flowchart with these stages: context intake, task routing, evidence collection, multi-model analysis, comparison and verification, and final recommendation. For every stage, include the required inputs, owner, decision criteria, failure risks, and retained artifact. Flag assumptions and unresolved questions. After the flowchart, generate a matrix comparing three implementation paths by user value, delivery complexity, dependency risk, reversibility, and evidence strength. Keep sourced facts separate from working assumptions.

What makes this prompt useful is not its length. It defines a contract. The workflow can now be reviewed against explicit expectations rather than judged by whether the output feels intelligent.

Compare outputs instead of accepting the first answer

Multi-model generation only helps when the workflow defines how outputs will be compared. Otherwise, it creates several polished answers and transfers the synthesis burden back to the consultant.

A comparison layer should ask:

• Which assumptions appear in every output?
• Where do the outputs disagree on facts, priorities, or causal logic?
• Which response introduces a useful alternative that others ignore?
• Which claims are supported by evidence, and which are merely plausible?
• What would need to be true for each recommendation to succeed?
• Which choice is easiest to reverse if the evidence changes?

The goal is not automatic consensus. Agreement can hide a shared blind spot. The goal is structured contrast.

A matrix makes that contrast visible. A flowchart shows where verification occurs. A diagram exposes dependencies. A reusable workspace preserves the discussion so the next revision begins with accumulated reasoning instead of a blank prompt.

Retained reasoning is the real productivity gain

The most valuable output of an AI workflow is not always the final paragraph or recommendation. It is the visible record of how the team moved from evidence to judgment.

That record reduces repeated explanation. It helps a new collaborator understand why one path was rejected. It makes assumptions easier to update when conditions change. It also preserves professional agency: the AI contributes analysis, but the consultant can still identify where human judgment entered the process.

This is the practical lesson behind the 123-million-token extreme. When usage grows, every hidden weakness scales with it. Missing context becomes repeated rework. Unverified claims become repeated risk. Lost reasoning becomes repeated debate.

A better workflow does the opposite. It turns context into structure, structure into comparison, comparison into verification, and verification into a decision-ready visual that can still be edited.