Working Paper
March 2026

From Outputs to Outcomes

Context Is Not Enough: Domain Operating Frames as the Missing Layer Between Personal AI and Business Outcomes

Frontier AI models are rapidly becoming a shared utility. As access to top-end models diffuses, advantage shifts from raw model intelligence to the quality of the grounding around the model.

MA
Mohamed Anis

Founder, Human Edge

30 minResearch • Context Engineering • Architecture

Abstract

Frontier AI models are rapidly becoming a shared utility. As access to top-end models diffuses, advantage shifts from raw model intelligence to the quality of the grounding around the model. Recent analysis argues that when companies access the same models and tools, organisational context becomes the differentiator [1]. Meanwhile, Anthropic's own data shows that real-world AI deployment covers only a fraction of theoretically addressable work, and that the most complex tasks continue to reward complementary human expertise [2][3].

This paper argues that personal context alone — knowing who the user is — is necessary but insufficient for business outcomes. The missing layer is the Domain Operating Frame: a structured, queryable representation of the job being performed, including its ontology, workflow state, constraints, objective function, and domain playbooks. Personal context tells the system who is acting. The operating frame tells the system what game is being played.

We propose a bottom-up architecture that begins with one human, binds a personal context graph to a job-specific operating graph, and closes the loop with execution telemetry and feedback. We use fundraising as the first proving ground. The near-term thesis is not enterprise context. It is individual context plus one operating frame, which together can turn generic model outputs into aligned decisions and, eventually, measurable outcomes.

The shortest version of the idea: you are not missing a better prompt. You are missing a second graph. One graph for the human. One graph for the job.

01

The Equation

Every interaction between a human and an AI model has two sides: the human's contribution and the model's contribution. The quality of the result is bounded by the weaker of the two.

Weak Human
Weak AI

Noise

Weak Human
Expert AI

Great Output, No Context

Expert Human
Weak AI

Bad Output, Great Context

Expert Human
Expert AI

Outcomes

This equation reveals a structural asymmetry in the current AI industry. The entire venture capital ecosystem, the research community, and the technology sector are focused on the right side: making AI more expert. Anthropic, OpenAI, Google DeepMind, and others have raised over $50 billion to advance model intelligence. Their progress has been extraordinary.

The left side — making the human more expert — has received almost no systematic investment. There is no infrastructure for it. There is no venture category for it. There is no research agenda around it.

This paper argues that the left side is not merely important — it is the binding constraint. Model intelligence is advancing on a predictable trajectory. Human expertise in using these models is not advancing at all. Most users interact with frontier AI the same way they interacted with GPT-3.5: with shallow prompts, no context, and no domain-specific grounding. They get polished outputs. They do not get outcomes.

The bottleneck is no longer the model. It is everything that happens before the model receives its input.

02

The Blind Spot

The foundation-model layer is rapidly commoditising. In February 2026, Harvard Business Review argued that when firms can access the same AI models, the same AI-enabled tools, and the same vendor ecosystem, organisational context becomes the differentiator. Their definition of context is not a slogan; it is demonstrated execution: the workflows teams actually follow, the signals they respond to, the exceptions that trigger action, and the judgment calls that repeat across real work [1].

The Evidence from Anthropic

Anthropic's March 2026 labour market paper reinforces a complementary point. It distinguishes between theoretical capability and observed usage, reporting that actual AI coverage remains a fraction of what is theoretically feasible. In its data, Claude currently covers just 33% of tasks in the Computer and Math category even though the theoretical addressable share is much larger [2]. That gap matters. It implies that business value does not arrive automatically when models become more powerful. Something else is bottlenecking adoption.

Anthropic's January 2026 Economic Index report sharpens the diagnosis. It finds that Claude generally succeeds on many tasks, but success rates fall on more complex tasks, and the value of complementary human expertise remains high when interpreting and validating the work [3]. Put bluntly: model capability is advancing fast, but real business outcomes still depend on better grounding, judgment, and context management around the model.

The Measurement Problem

The labour market report measures AI's impact on jobs through task-level exposure metrics — evaluating how much of each role's tasks can be performed or augmented by AI. This is methodologically sound but fundamentally incomplete.

Tasks are the visible outputs of work. But the value of most knowledge work is not in the task itself — it is in the contextual intelligence that shapes the task. When a senior marketing director writes a campaign brief, the task is "writing." The value is the fifteen years of pattern recognition, competitive awareness, stakeholder dynamics, and institutional memory that determine what the brief says and how it is framed. AI can perform the task. It cannot supply the context.

The industry is optimising for task automation while the actual business value resides in context preservation and amplification. Every unfilled role is a context gap. Every automated task without captured context is intelligence lost.

03

Why Context Alone Is Not Enough

Most current discussion of AI personalisation stops too early. It assumes that once a model knows who the user is — their voice, goals, preferences, background, and relationships — the system will become meaningfully useful. That is only partly true. Personal context improves relevance. It does not, by itself, guarantee business outcomes.

Consider a founder using AI to raise a Series A round. Assume the system has a comprehensive digital twin of that founder: traction metrics, team composition, communication style, investor network, company narrative, and strategic vision. Assume the model is the most capable reasoning engine available.

With these two components alone, the output will be technically excellent and personally authentic. But it will not necessarily produce fundraising outcomes.

Because the system does not know how fundraising actually works.

It does not know that the sequencing of investor engagement follows a specific protocol: warm introduction, then deck, then data room, then term sheet. It does not know that different investor archetypes — venture capital, corporate strategic, sovereign wealth, family office — require structurally different approaches. It does not know that Series A evaluation criteria differ fundamentally from seed criteria. It does not know that the timing of a fundraise relative to market cycles and portfolio construction windows materially affects outcomes. It does not know the unwritten social norms of follow-up cadence, information sharing, and partner dynamics.

This knowledge belongs to neither the individual nor the model. It belongs to the domain. It is the accumulated operational intelligence of fundraising as a discipline.

Without it, even a perfect digital twin and a perfect model produce what we call "polished nothing" — content that sounds right but misses operational reality.

Without the domain operating frame, the system risks becoming a very sophisticated autobiography engine. Without the personal layer, it becomes just another generic vertical copilot. Without execution and feedback, it becomes a recommendation engine. All three must be present.

04

The Domain Operating Frame

The missing piece is not "better prompting." It is not "niche" in a vague sense. It is a structured operating frame for the job being performed.

Definition: A Domain Operating Frame is a structured, queryable model of a job that answers five questions: what entities matter, where are we in the workflow, what constraints apply, what does winning look like, and what playbooks govern the next move.

Personal context tells the system who the human is. The Domain Operating Frame tells the system what game is being played, what stage it is in, what constraints apply, and what winning looks like.

Why Existing Terms Are Inadequate

"Niche" — a market label, not an executable system.

"Workflow" — too narrow; captures steps but not judgment, trade-offs, or objective functions.

"Domain context" — too muddy, because "context" already refers to the personal layer in our framework.

"Playbook" — too static; implies fixed instructions rather than a living, evolving intelligence layer.

"Prompting" — an interface technique, not an operating model.

The Domain Operating Frame is broader and more exact: it is the formal structure of the job itself.

The Ontological Structure

1. Workflow Ontology

The sequenced stages, decision gates, and branching paths that define how the job proceeds from initiation to completion. In fundraising: identification, qualification, warm introduction, first meeting, due diligence, term sheet, close.

2. Actor Taxonomy

The roles, archetypes, and stakeholder categories involved in the domain, and how they differ in expectations, criteria, and communication norms. In fundraising: angel investors, seed funds, Series A VCs, corporate strategics, sovereign wealth funds, family offices.

3. Operational Norms

The unwritten rules, social conventions, timing expectations, and professional customs that govern effective execution. In fundraising: warm introductions carry more weight than cold outreach. Founders should not send decks before establishing interest.

4. Evaluation Criteria

The benchmarks, pattern-matching heuristics, and success/failure indicators used by domain participants to assess quality. In fundraising: revenue run rate, growth velocity, unit economics, team pedigree, market size methodology.

5. Temporal Dynamics

The cycles, windows, seasons, and timing-dependent patterns that affect when and how actions should be taken. In fundraising: VC deployment cycles, end-of-fund dynamics, market sentiment windows, competitive fundraise timing.

6. Risk Topology

The common failure modes, red flags, and error patterns that experienced practitioners know to avoid. In fundraising: signalling risk, valuation anchoring errors, over-optimisation for terms versus partner quality.

This structure is domain-specific but architecturally consistent. Fundraising, enterprise sales, executive recruiting, content strategy, and wealth management each have their own workflow ontology, actor taxonomy, and operational norms. The infrastructure that captures and delivers the Operating Frame can be generalised. The knowledge within it cannot.

05

The Revised Equation

The original equation was binary: Expert Human + Expert AI = Outcomes. Our research reveals this was directionally correct but structurally incomplete. The path from input to outcome has distinct stages, and each stage requires a different layer of infrastructure.

1.
Output

AI generates technically competent text. Required: Model capability alone

2.
Contextualised Output

AI generates text grounded in who the user is. Required: Model + Personal Context

3.
Aligned Decision

AI generates a recommendation grounded in both the person and the job. Required: Model + Personal Context + Domain Operating Frame

4.
Outcome

The recommendation is executed, results are observed, and the system learns. Required: All of the above + Execution Loop + Feedback Loop

Most AI products today operate at Stage 1 or Stage 2. They produce outputs — sometimes impressive ones — that are disconnected from the operational reality of the job being performed. This is consistent with Anthropic's finding that real-world AI coverage remains far below theoretical capability [2]. The gap between theoretical and actual is, in large part, the missing operating frame.

Deep Personal Context + Frontier Model = Better Output

Deep Personal Context + Domain Operating Frame + Frontier Model = Aligned Decisions

Aligned Decisions + Execution Loop + Feedback Loop = Business Outcomes

This is the equation that governs everything Human-Edge.AI builds.

Two Halves of the AI Industry

Model Infrastructure
  • Makes AI smarter
  • Commodity (multiple capable providers)
  • Capital-intensive, research-driven
  • Value from compute and algorithms
  • Converging (models becoming similar)
  • Scale moat (training costs)
Context Infrastructure
  • Makes the human smarter
  • Proprietary (unique to each user and domain)
  • Data-intensive, relationship-driven
  • Value from accumulated context and expertise
  • Diverging (context becomes more unique over time)
  • Dual moat (retention + network effect)

Model Infrastructure is the domain of Anthropic, OpenAI, and Google DeepMind. Context Infrastructure — the combination of Personal Context, Domain Operating Frames, and Execution Loops — is the domain we are building. Neither half works alone. Together, they produce outcomes.

06

The Automatic Prompt

The Problem With Prompt Engineering

The prevailing industry narrative holds that AI's value is unlocked through prompt engineering — the human skill of crafting instructions that elicit optimal model responses. The premise is wrong. Or rather, it is a transitional truth — necessary today because the infrastructure does not yet exist, but destined to be absorbed by the infrastructure layer as it matures.

When context is absent, the user must supply it manually through the prompt. When Domain Intelligence is absent, the user must encode it manually through the prompt. The prompt becomes a lossy compression format for knowledge that should exist in the system, not in the user's head.

A common objection is that Domain Intelligence can be supplied through better prompts. An expert fundraiser could simply instruct the model: "Write this investor email following Series A conventions, using warm introduction norms, and calibrated for a venture capital audience." This objection proves our point. The ability to write that prompt requires the user to already possess the Domain Intelligence. If using AI effectively requires the user to already be a domain expert, the tool has failed its purpose.

The Three Generations of AI Interaction

Generation 1: Raw Prompting

The user types a request. The model responds. No context, no domain knowledge, no personalisation. Output quality is entirely dependent on the user's ability to articulate what they want. This is how most people use AI today.

Generation 2: Prompt Engineering

The user learns to craft structured prompts with role assignments, few-shot examples, chain-of-thought instructions, and output formatting. Output quality improves but remains dependent on user skill. This approach scales poorly — it requires every user to become an expert in a new technical discipline.

Generation 3: Context-Driven Interaction

The system assembles the prompt automatically by combining Personal Context (who the user is), the Domain Operating Frame (what the job requires), and task parameters (what needs to happen right now). The user provides a simple, natural-language instruction. The system generates an expert-grade prompt internally.

How It Works

When a user says "help me raise my Series A," the Context Infrastructure performs three operations:

Step 1 — Personal Context Retrieval

The system traverses the user's knowledge graph to assemble relevant contextual information: company traction, financial position, team composition, prior investor relationships, communication style, strategic narrative, and current stage. This is a targeted graph query, not a document dump — only context relevant to fundraising is retrieved.

Step 2 — Domain Operating Frame Activation

The system activates the fundraising operating frame: workflow stages, investor archetypes, evaluation criteria, sequencing norms, timing dynamics, and risk patterns relevant to the user's specific fundraising stage. This provides the operational expertise that the user may not possess.

Step 3 — Prompt Synthesis

The system combines personal context and the domain operating frame into a structured, optimised prompt sent to the model. The prompt contains the user's authentic voice, factual context, strategic positioning, and domain-appropriate framing — all assembled without requiring the user to understand prompt engineering, fundraising conventions, or AI interaction design.

The user types six words. The system delivers expert-grade output. That is the product.

If this analysis is correct, prompt engineering is a transitional discipline. The companies that build context infrastructure will capture the value currently dissipated across millions of individual prompting efforts.

07

Why the Correct Wedge Starts With One Human

Human-Edge is explicit that it is not yet solving company-wide context. That honesty is strategically valuable. Investors have heard too many AI stories that start with the enterprise because the total addressable market sounds larger. The more credible wedge is smaller and more original: begin with the individual human, then bind that human to one job-specific operating frame.

The Architectural Argument

There are structural reasons for this sequencing:

Every higher layer of context is composed of individuals. Team norms, company culture, and organisational memory are emergent properties of the people who create them. A company's decision-making pattern is the emergent product of its members' judgment, communication styles, and expertise. You cannot faithfully represent a company's context if you have not first represented the humans within it.

Individual context is the hardest to fake. A company can document process. It is much harder to reconstruct how a particular founder thinks, what trade-offs they tolerate, how they communicate under pressure, and who they can actually reach through their network. Enterprise context can be approximated through process documentation, workflow analysis, and organisational charts. Individual context cannot.

Individual context is portable. A person's context travels with them across teams, companies, and careers. Enterprise context is locked inside organisational boundaries. If Human-Edge can become the trusted system that knows one person deeply, it can later expand into multiple jobs for that individual, then to interactions between individuals, and only after that toward team and enterprise context.

Bottom-Up vs. Top-Down

The history of technology suggests that the most durable architectures are built from the bottom up. The internet was not designed by starting with enterprise applications. It was designed by building a protocol that allowed any two computers to communicate. Social networks were not designed by modelling organisational behaviour. They were designed by modelling individual identity and connections, then letting group dynamics emerge.

Top-down context is brittle. When you build a model of how an organisation works by instrumenting systems of record — CRMs, ERPs, ticketing systems — you capture process outcomes, not the reasoning that produced them. The model looks correct until something changes: a key person leaves, a team is reorganised, a market shifts.

Bottom-up context is resilient. When context is built from individuals, the system can adapt to organisational change naturally. When a team is restructured, the individuals' contexts persist. When someone changes roles, their contextual layer travels with them.

08

Architecture: Two Graphs on a Memory Substrate

The Right Substrate

The architecture requires a memory substrate that combines relational storage for document provenance, vector storage for semantic similarity, and graph storage for entities and relationships. The ideal substrate treats structured units of knowledge as first-class objects that can become both embeddings in vector search and nodes or edges in the graph — allowing semantic recall and structural reasoning to operate over the same underlying memory objects.

The broader research literature supports this direction. The recent GraphRAG survey argues that graphs are especially powerful for retrieval because they capture heterogeneous and relational information that plain embedding spaces do not represent cleanly [10]. Microsoft's GraphRAG paper demonstrates that graph-based approaches can outperform conventional RAG on global sensemaking questions over large private corpora [11]. Mem0 reports that persistent memory beats full-context baselines while sharply reducing latency and token cost [12]. The ACE paper shows that evolving contexts updated with natural execution feedback can outperform strong baselines on agent and finance tasks [13].

The implication is clear: structured memory matters. But structured memory alone is still not the same thing as an operating frame. A memory substrate gives us storage, connection, enrichment, and retrieval primitives. It does not define what success means inside fundraising, what sequence of moves is appropriate at each stage of a raise, or how a founder should trade off speed, valuation, investor fit, and dilution. Those are not generic memory features. They are job-specific operating intelligence. That is where Human-Edge must add something original on top.

Two Graphs

Graph 1: The Human Graph (Personal Context)

Captures who the user is. Primary entity types:

  • Identity nodes: voice signature, communication patterns, psychological traits, expressed values
  • Expertise nodes: skills, knowledge domains, professional history, demonstrated capabilities
  • Network nodes: relationships, connection strength, interaction patterns, mutual contacts
  • Goal nodes: stated objectives, current priorities, stage of progression
  • Content nodes: articles, posts, insights, anecdotes, prior work products
  • Artifact nodes: documents, decks, memos, and other work products with versioning

Edges are typed (HAS_VOICE, BELIEVES, FOCUSES_ON, CONNECTED_TO), weighted by confidence and recency, and subject to temporal decay. Information older than 90 days loses confidence weight. Frequently validated information gains weight.

Graph 2: The Job Graph (Domain Operating Frame)

Captures how the job works. Its structure follows the six ontological components defined in Section 4: workflow ontology, actor taxonomy, operational norms, evaluation criteria, temporal dynamics, and risk topology. For fundraising, this means stage nodes, investor archetype nodes, artifact readiness nodes, constraint nodes, and norm nodes — all connected by typed relationships that encode sequencing, dependencies, and decision logic.

The Job Graph is seeded by domain experts and refined by every user operating within the domain. It is not static documentation. It is a living ontology that evolves as more users operate within it.

The Coupling: Just-In-Time Context Assembly

The power of the architecture is in the intersection. When the system traverses the Human Graph AND the Job Graph simultaneously, it produces output that is both authentically the user and operationally sound for the domain. Neither graph alone gets you there.

For each task, the system:

  1. Identifies the relevant domain operating frame
  2. Determines the user's current position within the frame (what stage, what constraints, what next actions)
  3. Traverses the Human Graph for context relevant to the current task
  4. Traverses the Job Graph for domain intelligence relevant to the current stage
  5. Assembles a concentrated context payload that is both personal and operational
  6. Generates the optimised prompt internally
  7. Sends to the model

This prevents the "context dumping" failure mode where too much information degrades model performance, and the "context starvation" failure mode where too little information produces generic output. Response times target sub-400ms for standard queries.

The Human-Edge Stack

1.

Personal Context Graph

voice, identity, goals, preferences, trajectory, relationships, artifacts, history

2.

Domain Operating Frame

ontology, state, constraints, objective function, playbooks

3.

Just-in-Time Context Assembly

retrieves the minimum sufficient slice of both graphs for the current decision

4.

Agent and Tool Runtime

specialised agents execute next-best actions

5.

Execution Telemetry and Feedback

outcomes flow back into both graphs

09

Fundraising as Proving Ground

Fundraising is the right initial domain because it concentrates exactly the kinds of work generic AI handles badly: relationship-sensitive sequencing, narrative adaptation, asymmetric information, evolving state, and repeated judgment under time pressure. It is a job with clear artifacts — deck, memo, target list, warm paths, data room, diligence materials, meeting notes, follow-ups, and term-sheet trade-offs — but those artifacts only make sense inside a live operating frame.

The Fundraising Operating Frame

Round Configuration

Round type (pre-seed, seed, Series A, Series B), target raise amount and acceptable range, valuation expectations and dilution constraints, timeline and urgency factors, use-of-funds narrative.

Investor Intelligence

Investor archetypes (VC, corporate strategic, sovereign wealth, family office, angel) with per-archetype evaluation criteria, thesis alignment, typical check sizes, stage and sector and geography fit filters.

Relationship Mapping

Warm introduction paths (traversing the Human Graph's network nodes), introduction quality scoring (first-degree vs. second-degree, relationship strength), mutual connection identification, prior interaction history.

Outreach Sequencing

Warm intro before deck transmission, deck before data room access, first meeting protocol (associate screen vs. partner meeting), follow-up cadence norms (timing, tone, content).

Readiness State

Deck quality and completeness, data room artifact checklist, financial model robustness, team narrative coherence, reference and customer proof availability.

Objective Function

Optimisation target (speed, valuation, strategic value, dilution minimisation, founder time cost), trade-off framework when objectives conflict, walk-away criteria.

The Difference in Practice

Generic Assistant Response

What should I do next to accelerate this raise?

Refine the deck. Target the right investors. Follow up promptly. Network more actively.

Operating-Frame-Aware Response

What should I do next to accelerate this raise?

Revise the deck's traction section before broad outreach — two investors in the last batch flagged it as unclear. Prioritise the five investors with warm paths already present in your relationship graph. Pause cold outreach to two geographies where the current story is weak. Prepare a one-page diligence memo addressing the churn objection surfaced in the last two calls. Schedule the three highest-fit meetings within the next 10 days to build parallel momentum before the end-of-quarter deployment window.

That is not merely better output. It is a materially better next decision.

Measurement

Once the system can repeatedly produce this level of specificity, measurement becomes possible. Investors should not ask whether the AI sounds smart. They should ask whether the system increases qualified introductions, improves reply rates, compresses time-to-close, raises diligence conversion, and reduces founder hours spent on low-probability activity. That is the bridge from personalisation to business outcome.

10

Execution and Feedback Loop

Even with the operating frame, outcomes do not appear automatically. The system must act, observe results, and recalibrate. This is the bridge from aligned decisions to business outcomes.

Action

The system generates a recommendation or produces an artifact (investor email, deck revision, follow-up message). The user approves, modifies, or rejects it.

Observation

The system observes what happens. Was the email opened? Did the investor respond? Was a meeting scheduled? Did the meeting progress to the next stage? Each observation is a signal.

Calibration

Signals feed back into both graphs:

  • Human Graph calibration: Edit patterns refine voice profiles, content preferences, and communication style. If the user consistently softens aggressive language, the system learns.
  • Job Graph calibration: Aggregate signals across users refine the domain operating frame. If 80% of successful Series A raises use a specific outreach sequence, that pattern strengthens in the ontology.
Adaptation

The system adjusts its next recommendation based on accumulated signals. Confidence scores update. Stale patterns decay. Proven patterns strengthen.

Why This Matters

Without the feedback loop, the system is a one-shot recommendation engine — better than no context, but not improving over time. With the feedback loop, every interaction compounds:

  • The Human Graph gets richer
  • (more calibration data, more refined voice, more accurate relationship mapping)

  • The Job Graph gets smarter
  • (more pattern validation, more norm refinement, more nuanced archetype models)

  • The coupling between them gets tighter
  • (better JIT assembly, more precise prompt generation, higher alignment between output and operational reality)

This compounding mechanism converts a software tool into a learning system, and a learning system into a competitive moat.

11

The Compounding Flywheel

The economic argument for Context Infrastructure rests on the interaction between its two graph layers.

Two Moats

Personal Context = Retention Moat

The Human Graph is private, personal, and becomes more valuable with use. Voice signatures, relationship maps, communication patterns, goal structures, and accumulated calibration data represent months or years of investment. This context does not transfer to a competitor. The richer it gets, the more irreplaceable the system becomes.

Domain Operating Frame = Network Effect Moat

The Job Graph is shared across users operating within the same domain. Every user's interactions generate calibration signals that refine the operating frame. When 500 founders use the system for fundraising, the 501st founder receives better fundraising intelligence on day one than any individual user could build alone.

The Flywheel

1.

A user builds Personal Context

— digital twin, voice signature, network graph, goals.

2.

The user operates within a domain

Every interaction generates calibration signals that refine the Domain Operating Frame.

3.

Improved Domain Intelligence attracts more users

Better fundraising intelligence draws more founders.

4.

More users generate more signals

More users generate more Personal Context and more calibration signals, further refining the Domain Operating Frame. The cycle compounds.

Defensibility

A competitor who builds only Personal Context has a retention moat but no network effect — the product does not improve for new users. A competitor who builds only Domain Intelligence has a network effect but no retention — users have no personal investment and can switch freely. The combination creates a defensibility profile that neither approach achieves alone.

The moat is not the base model. It is the compounding interaction of memory, job structure, and outcome telemetry.

12

What We Are Building: Now, Next, and Later

The most credible investor narrative is not that Human-Edge has already solved the full stack of context. It is that the company has identified the right sequence in which to solve it.

The Build Sequence

Now

Deep personal context graph plus the first Domain Operating Frame, with fundraising as the proving ground. The goal is to demonstrate that binding individual context to one operating frame produces measurably better next-step decisions than a generic assistant.

Next

Extend the same Human Graph into adjacent operating frames — enterprise sales, executive recruiting, business development. The Human Graph is shared across domains. A founder's Personal Context serves them whether they are fundraising, selling, hiring, or building their public profile. Build the Human Graph once, activate it across many jobs.

Later

Model interactions between context-aware users. When individuals within the same team use the platform, Team Context (Layer 2) emerges naturally from interaction patterns between individual knowledge graphs. Only after individual and team context are proven do we attempt department and enterprise context, grounded in actual human behaviour rather than inferred from system logs.

The Seven Layers of Scale

1

Individual

Voice, identity, expertise, goals, patterns — Building now

2

Team

Shared norms, decision dynamics, role complementarity — Next

3

Department

Cross-team coordination, institutional memory — Future

4

Company

Culture, strategy, risk tolerance, execution philosophy — Future

5

Industry

Regulatory norms, competitive dynamics, market cycles — Horizon

6

Region

Cultural norms, legal frameworks, economic conditions — Horizon

7

Global

Shared human knowledge, scientific consensus — Horizon

Most AI memory systems jump straight to Layer 4 because that is where enterprise revenue exists. This produces brittle architectures. We begin at Layer 1 because individual context is the hardest to fake, the most authentic, and the most portable.

Relationship to Enterprise Memory Companies

In the near term, Human-Edge is not trying to out-execute platforms that instrument organisation-wide workflows. It is building the lower layer they often skip: the individual context graph bound to a real operating frame. Over time these approaches are complementary rather than contradictory. Enterprise context built from the bottom up tells you what the organisation actually is. Enterprise context built from the top down tells you what the organisation looks like.

13

What Investors Should Measure

The right diligence question is not whether the demo is impressive. It is whether the system is learning the job well enough to move operating metrics.

Graph Recall Fidelity

When the system retrieves context for a fundraising task, how accurately does it surface the relevant personal and domain information? Measured by relevance precision on a held-out set of human-graded queries.

State Inference Accuracy

Can the system correctly identify where a user is in the fundraising workflow — what stage, what has been completed, what is pending — without being told explicitly?

Action Acceptance Rate

When the system recommends a next step, how often does the user accept it without modification? This is the most direct measure of decision alignment.

Workflow Compression

Does the system measurably reduce the time from fundraise initiation to close, or the number of low-value activities performed during the process?

Outcome Lift

For users who engage with the operating-frame-aware system: are qualified introductions higher, reply rates better, diligence conversion improved, and founder hours on low-probability activity reduced?

Extensibility

Can the architecture support a second operating frame (e.g., enterprise sales) without fundamental redesign? This validates the claim that the infrastructure generalises while the domain knowledge specialises.

14

Risks and Open Questions

The obvious risks do not weaken the thesis. They sharpen it. Most are direct consequences of choosing to work on the hard layer rather than the decorative one.

Cold Start

The Human Graph requires data before it is useful. A new user with no ingested context receives limited benefit. Mitigation: design the onboarding to capture high-value context quickly (voice recordings, LinkedIn import, document ingestion) and make the Domain Operating Frame immediately valuable even before the personal graph is deep.

Privacy and Consent

Individual context is deeply personal. The system must earn trust through transparency about what is stored, how it is used, who can access it, and how it can be deleted. This is a feature, not a bug — trust is the moat that competitors will find hardest to replicate.

Evaluation Difficulty

Measuring "better decisions" is harder than measuring "better text." Task-level quality metrics (BLEU, perplexity, human preference) do not capture decision quality. The system needs outcome-level metrics that take time to accumulate. Fundraising is a good first domain precisely because outcomes are measurable (meetings booked, term sheets received, rounds closed).

Frame Brittleness

An operating frame that is too rigid will fail when real-world fundraising deviates from the modelled workflow. The frame must be adaptive — learning from exceptions, not just standard patterns. The feedback loop is the mechanism for this.

Execution Liability

When AI moves from generating text to recommending next actions in high-stakes domains like fundraising, the consequences of bad recommendations are material. The system must be explicit about confidence levels and preserve human judgment authority at every decision point.

These are the right risks to face. The wrong risk is making another generic assistant sound persuasive in a demo.

15

The Investment Thesis

Observation 1: Models Are Converging Toward Commodity

Multiple labs produce models of comparable capability. Claude, GPT, and Gemini will continue to improve, but the gap between them narrows with each generation. When every company has access to the same model intelligence, the model provides no competitive advantage.

Observation 2: Value Accrues to the Input Layer

In every technology wave, value migrates from the component layer to the layer that controls the user relationship. In cloud computing, value migrated from hardware to the platform (AWS, Azure). In mobile, from devices to the app ecosystem. In AI, value will migrate from the model layer to the context layer — the system that determines what the model receives as input and therefore what it produces as output.

Observation 3: Context Infrastructure Has Two Compounding Moats

The Personal Context layer creates switching costs that deepen with use. The Domain Operating Frame layer creates network effects that strengthen with adoption. Together, they produce the compounding dynamic required for platform economics — not just tool economics.

Observation 4: The Adoption Gap Is an Infrastructure Problem

Anthropic's own data shows that real-world AI usage covers a fraction of theoretical capability [2][3]. The gap is not model intelligence. The gap is that users lack the context and domain grounding to extract real value from these models. Closing that gap is an infrastructure problem, and infrastructure problems create large, durable companies.

The Opportunity

The AI industry has invested over $50 billion in Model Infrastructure. It has invested nearly nothing in Context Infrastructure. Human-Edge.AI is building the missing half — beginning with the individual, grounded in one domain, and expanding through seven layers of complexity.

The race to build smarter models is being won. The race to build smarter context has not yet started. That is the opportunity.

16

Conclusion

Deep Personal Context + Frontier Model = Better Output

Deep Personal Context + Domain Operating Frame + Frontier Model = Aligned Decisions

Aligned Decisions + Execution Loop + Feedback Loop = Business Outcomes

The market is not short of models. It is short of systems that know enough about a person and the job they are performing to produce decisions that survive contact with reality.

Human-Edge's opportunity is to build the missing layer between personal AI and business outcomes: the Domain Operating Frame. The company describes itself as building a bottom-up context system that starts with one human, binds that human to one job-specific operating graph, and improves through execution and feedback.

This is a more credible story than claiming enterprise omniscience. It is also a stronger one. Start with one human. Ground the system in one job. Learn from real outcomes. Then expand the library of operating frames and, only after that, climb toward team and company context.

That is not a retreat from the larger vision. It is the only honest way to build it.

One graph for the human. One graph for the job. One model that receives both. That is how AI produces outcomes.

17

Appendix A: Technical Architecture

A.1 The Three-Store Architecture

Knowledge Graph Store

Entities — people, companies, skills, beliefs, investors, deal stages — are modelled as nodes. Connections between them are typed, weighted edges. The graph captures what no other format can: the relationships between facts, which are the actual carrier of context.

Vector Store

Semantic embeddings enable similarity-based retrieval across the entire context corpus. When an agent needs anecdotes, prior experiences, or thematically related content, the vector store provides sub-second retrieval based on meaning rather than keywords. This powers the Semantic Vault.

Relational Store

Document provenance, chunk lineage, processing metadata, and temporal records are maintained in a relational database. This provides the audit trail and temporal dimension — tracking not just what is known, but when it was learned.

A.2 The Processing Pipeline

Extract-Cognify-Load pattern:

Extract

Raw data — voice recordings, social posts, documents, professional profiles, interaction logs — is classified, chunked semantically (200–500 words per chunk), and prepared for processing.

Cognify

An LLM extracts entities and relationships from each chunk, generating triplets (subject-relation-object) committed to the graph store. Simultaneously, embeddings are generated and stored in the vector store. Ontologies ground extracted entities to canonical concepts, preventing duplication and ensuring consistency.

Memify

Over time, the system prunes stale connections, strengthens frequently-accessed pathways, reweights edges based on usage signals, and adds derived insights. Context is not static storage — it is an evolving structure that adapts based on interaction patterns, calibration signals, and temporal decay.

A.3 Node Scoping and Multi-User Isolation

NodeSets tag subsets of the graph into logical containers — one user's personal context is isolated from another's, while shared domain intelligence can be accessed by all users within a domain. This enables the dual-layer architecture: private Human Graphs with shared Job Graphs.

A.4 Retrieval Architecture

  • Graph traversal:
  • For structured queries requiring relationship-based reasoning ("who in my network connects to this investor?")

  • Vector similarity:
  • For semantic queries requiring meaning-based matching ("find anecdotes about overcoming product-market fit challenges")

  • Hybrid:
  • For complex queries requiring both structural relationships and semantic relevance (assembling a complete fundraising context payload)

18

Appendix B: The Four Axes

The Scale Axis

The seven layers: individual, team, department, company, industry, region, global. This describes the organisational scope of context. We are building at Layer 1.

The Value Axis

The eight dimensions of human progression: knowledge, social, financial, physical, time, spiritual, legacy, purposeful. This describes the domains of human life that the system ultimately serves. It is the long-term vision for individual context depth.

The Job Axis

The domain operating frames: fundraising, sales, recruiting, content strategy, wealth management. This describes the professional activities where context infrastructure produces measurable outcomes. We are starting with fundraising.

The Execution Axis

The operational pipeline: discovery, processing, delivery, feedback. This describes how the system converts context into action and action into learning.

These four axes are orthogonal. The Scale Axis determines how wide the context extends. The Value Axis determines how deep the individual context goes. The Job Axis determines what domain the system operates within. The Execution Axis determines how the system acts and learns.

The current paper addresses one position on each axis: one individual (Layer 1), building personal context across relevant dimensions, operating within the fundraising domain, through a full discovery-to-feedback execution pipeline. Everything else is the expansion path.

References

[1] Rohan Narayana Murty and Ravi Kumar S., "When Every Company Can Use the Same AI Models, Context Becomes a Competitive Advantage," Harvard Business Review, February 18, 2026.

[2] Anthropic, "Labor market impacts of AI: A new measure and early evidence," March 5, 2026.

[3] Anthropic, "Anthropic Economic Index report: economic primitives," January 15, 2026.

[10] Haoyu Han et al., "Retrieval-Augmented Generation with Graphs (GraphRAG)," arXiv:2501.00309, 2025.

[11] Darren Edge et al., "From Local to Global: A Graph RAG Approach to Query-Focused Summarization," arXiv:2404.16130, revised 2025.

[12] Prateek Chhikara et al., "Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory," arXiv:2504.19413, 2025.

[13] Qizheng Zhang et al., "Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models," arXiv:2510.04618, revised 2026.

Keep Reading

Explore more essays on strategy and AI

Read: The Arena Is the Product