---
title: "Four Architectures That Make AI Work"
url: "https://atlan.com/context-and-chaos/issue/four-architectures-that-make-ai-work/"
description: "Decisions, meaning, proof, people. The four systems that decide whether it delivers."
keywords: "Enterprise AI, Context Engineering, AI Readiness, Data Governance"
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

A Context & Chaos essay by Colin Hardie, Chief Data Architect at SEFE (London), published September 21, 2026 (20 min read). His argument: when AI initiatives fail, "the model isn't the problem, everything else is." A model is a component; systems produce value. Four architectures decide whether AI delivers: decisions, meaning, proof, and people.

**Author.** [Colin Hardie](https://www.linkedin.com/in/colinhardie/) works on enterprise data and AI strategy: the decision architecture, semantic foundations, and operating models that determine whether AI initiatives reach production.

## The setup

- [Gartner's 2026 CIO and Technology Executive Survey](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai): 17% of organisations have deployed AI agents; more than 60% expect to within two years.
- Agentic AI echoes [Dan Ariely's](https://x.com/danariely/status/287952257926971392) line on Big Data: everyone talks about it, nobody knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are.
- Many will join the [42% who shelved their initiatives](https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/).
- AI exposes and magnifies existing capability. It removes the hiding places: the analyst compensating for ambiguous definitions, the workaround over a broken process, the human patience absorbing what the system could not. An agent pointed at a business that disagrees with itself about what its words mean "will scale the disagreement at machine speed."
- The better board question than "are we ready for AI" is whether the organisation itself works. Architecture here means the deliberate design and arrangement of parts and their relationships in service of a purpose.

## The Architecture of Decisions

The deliberate design of the path from strategic objective to operational outcome: a closed loop connecting objectives with the decisions and actions where they are won or lost, and the capabilities that support them.

- Hans Peter Luhn of IBM coined the modern usage of business intelligence in 1958 as information that guides action towards a goal; the industry then spent 60+ years producing artefacts (warehouses, marts, lakes, dashboards) instead of closed loops.
- Technology selections masquerade as strategy ("We need a lakehouse," "Let's implement governance"). Real strategy asks where objectives are won or lost and what the decisions are at that point.
- **Hardie's framework** (diagram): strategic objectives decompose into decisions and actions (which customers do we intervene with, which trades do we approve, which suppliers do we renew), then a prioritised portfolio of initiatives (each gated by feasibility, impact, and risk, each naming the decision it serves), then foundational capabilities built incrementally. Measurement runs as a vertical spine through all four layers and the outcome feeds back to the objective. Some problems need AI; some need a report, a cleaner process, or a conversation.
- Process redesign is inherent in naming the decision. [McKinsey's 2025 State of AI survey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value) tested 25 organisational attributes; fundamental workflow redesign had the largest effect on EBIT impact from AI, and high performers are roughly three times as likely to have redesigned workflows.
- "An agent dropped into an undesigned process is not a better dashboard. It is a dashboard that can act, which is worse."
- **Retail banking example.** Bolt-on: a churn model emails the retention team a ranked list every Monday that mostly goes unread. Redesigned: start from "which customers do we intervene with this week, and with what offer?", push them into the agent's work queue, hold back a control group, feed results into next week's model. Same data, team, and technology; one produces an artefact, the other retained customers.

## The Architecture of Meaning

The deliberate design of shared understanding: a governed path ensuring the system and the organisation agree on what words mean, how they are computed, and where to go when the answer is ambiguous.

- Self-service analytics failed on context, not access: which table is trusted, which revenue definition applies, is the fraud filter included. A human analyst compensates or asks; an agent picks the most likely interpretation and returns an answer that is "plausible, well-formatted, and wrong."
- Anthropic's own analytics agent never exceeded 21% accuracy without structured context; after encoding analytical workflows and business context as governed, reusable skills, [aggregate accuracy exceeded 95%](https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude). Same model, same data.
- The minimum context is narrower than the full enterprise context layer (mapped by Prukalpa in [What an Enterprise Context Layer Actually Is](https://atlan.com/context-and-chaos/issue/what-an-enterprise-context-layer-actually-is/)). Three layers answer three questions:
  1. **Definition**: what do we mean when we say this word?
  2. **Computation**: how do we compute it?
  3. **Procedural guidance**: where do I go to answer this question, and what do I do when it is ambiguous?
- Diagram: a question ("what was EMEA revenue") passes through all three layers before reaching data; the direct route to raw data is blocked. A dashboard, a notebook, and an agent asking for the same metric get the same number.
- Structural enforcement matters as much as content: if agents can bypass a layer and query raw tables, they will. Architecture does not rely on someone reading documentation; the wrong path is made harder than the right one.
- **Starting small:** a glossary of thirty to fifty terms for one domain in markdown, each linked to its physical table or column; shared metric definitions specifying exact logic, kept alongside data transformations (one canonical "EMEA revenue," not seven spreadsheets); a per-domain reference document telling the agent which definitions to use, which tables are trusted, and the gotchas a senior analyst carries in their head.
- Things that must change together must live together: at the markdown stage, same directory and same pull request; later, version-controlled schemas deployed through CI so a data model change still triggers a definition review.
- Controlled vocabularies, taxonomies, ontologies, and knowledge graphs are not sequential rungs; they solve different representation problems and are added when a use case demands them (see Jessica Talisman's [Ontologies, Context Graphs, and Semantic Layers: What AI Actually Needs in 2026](https://atlan.com/context-and-chaos/issue/ontologies-context-graphs-and-semantic-layers-what-ai-needs-in-2026/)). The question is when semantic precision becomes the bottleneck.

## The Architecture of Proof

Proof is a continuous system property with three structures.

- **Evidence.** The measurement spine extends into production: "did it deliver?" [IBM's 2026 study](https://newsroom.ibm.com/2026-06-08-new-ibm-study-finds-cios-and-ctos-face-growing-ai-control-gap-as-enterprise-deployment-scales) of 2,000 technology executives: 85% lack full visibility into real-time AI spend; 84% have not operationalised AI financial management. Define the metric, capture the baseline, and agree the target before writing code, so a leader can say "this initiative delivered a 14% reduction in customer churn."
- **Governance.** Much governance is theatre that drives shadow systems. The route to production must be the easiest route: risk-differentiated tracks with pre-approved patterns so most use cases never reach a committee and most stage gates become thirty-minute conversations instead of six-week reviews. Lightweight controls for last month's sales dashboard; comprehensive review for a model that autonomously approves insurance claims. Match the cost of oversight to the cost of being wrong.
- **Integrity.** Everything an AI system depends on is perishable; models drift. Provenance is the minimum defence: every answer should carry where the data came from, how current it is, and which definition was applied. Post-launch decay is covered in Ankit Anand's [Production AI & The False Finish Line](https://atlan.com/context-and-chaos/issue/production-ai-the-false-finish-line/).

## The Architecture of People

The deliberate design of the organisational forces that determine whether the other three survive reality. Four forces (diagram), three from system design and one human:

- **Coupling**: how tightly teams are bound. Too tight and nothing moves; too loose and three business units build three customer models with three definitions.
- **Cohesion**: whether each team has a clear, internally consistent purpose.
- **Interfaces**: defined contracts with agreed accountability, or informal relationships that collapse when someone leaves.
- **Incentives**: determine whether people do what the operating model says or what their performance review rewards, and whether the other three function as designed. Platforms launch with branded mugs and gather dust when no KPI mentions collaboration ("the mugs outlast the initiative"). See Gaurav Ramesh's [The Human Elements of the AI Foundations](https://atlan.com/context-and-chaos/issue/the-human-elements-of-the-ai-foundations/).

Conway's Law is a diagnostic: a data estate fragmented across incompatible platforms is an organisational problem expressed in technology. Every operating model choice is a power allocation decision; "somewhere in the middle" is not a design until you specify what is centralised, what is distributed, and the interface contract. A data function without a seat when strategy is discussed is an order-taker. The question is whether you are in the room when it matters.

## The Work

- None of the four architectures is new, and no vendor can sell all of it. Platforms provide infrastructure; they cannot choose your decisions, settle what your words mean, redesign workflows, or realign incentives. When a true idea becomes purchasable, organisations buy it and skip the work.
- The model sometimes is the binding constraint (a task at the edge of current model reliability); a narrow, low-risk use case does not need all four at equal maturity. Judging whether the model is good enough is the first architectural decision.
- Models improve for competitors at the same rate, often from the same vendor (Prukalpa: [If Intelligence Is Abundant, What Is the Moat?](https://atlan.com/context-and-chaos/issue/if-intelligence-is-abundant-what-is-the-moat/)). What differs is which decisions you chose, what your data means, whether you proved it works, and whether your people are enabled. "The model was never the thing being tested. You were."

Contributions reflect their authors' views; Context & Chaos curates and vets submissions, with no paid promotions. All issues: [https://atlan.com/context-and-chaos/](https://atlan.com/context-and-chaos/).