A Threadneedle Labs book
Beyondthe Model
The New Architecture of Artificial Intelligence
By Zia Karim and Faris Karim
The model was the breakthrough.
What comes next is the architecture.
AI spent years making the model smarter. The next challenge is everything around it.
Beyond the Model argues for a new unit of AI: the intelligence system. Its architecture determines how intelligence reasons, knows, acts, adapts and remains governable.
Five parts Eighteen chapters One connected argument
First edition · October 2026
Model/ The opening case
The model may be excellent. The decision still needs an architecture.
Imagine a bank reviewing a proposed change to a lending policy. An AI system reads the old policy, summarises the proposed text, checks several regulations, compares previous decisions and drafts a recommendation. The model that writes the recommendation may be excellent. Yet almost every consequential question concerns something outside it.
Six questions the architecture has to answer
Which version of the policy did it read?
Taken up in the part on KnowledgeWere the regulations current?
Taken up in the part on KnowledgeDid the system distinguish a legal obligation from an internal preference?
Taken up in the part on KnowledgeWhat happened when two sources disagreed?
Taken up in the part on KnowledgeWho authorised the recommendation to become a change in practice?
Taken up in the part on AgencyCould an auditor reconstruct the decision six months later?
Taken up in the part on Institution
None of these questions is answered by a better model. Every one is answered by the architecture around it.
Better models matter enormously. They make more of this work possible. They do not supply the evidence, authority and institutional responsibility the work requires.
The widening unit of intelligence
The model stays. The argument widens.
Each step asks what must surround capability for it to become useful, accountable intelligence. Architecture includes evidence, adaptation, authority, human judgement and institutional control.
These layers are connected obligations. A system’s freedom to act depends on what it knows, who authorised it and whether someone can intervene.
- Model
- Intelligence System
- Adaptive / Empirical Architecture
- Epistemic Architecture
- Agency & Authority
- Human / Organisational Intelligence
- Computational Governance
- Institutional Sovereignty
Intelligence/ Parts I & II
Intelligence is becoming a system.
A useful intelligence system chooses among models, tools, data sources, evaluators and people. The task determines which functions it needs and how they work together.
The unit of AI
The intelligence system: a loop of observe, retrieve, represent, reason, select, act, evaluate and adapt, with the model as one component at its centre.
The model as one component inside a larger architecture.
A repertoire of functions, not a compulsory sequence.
Chapters 3 & 4 · Allocation and Adaptation
Choose the method the work deserves.
A simple request may need a direct answer. A disputed claim may need retrieval, a calculation or a specialist. Adaptive intelligence selects the method, representation and reasoning budget as the evidence changes.
Chapter 6 · Architecture as Experiment
Test the arrangement against a simpler one.
A new route, agent or verifier is a hypothesis about better work. Compare outcomes and costs. Retain, revise or remove it. The ability to propose a change does not confer the right to deploy it.
“We stop treating a diagram as an argument and start treating it as a hypothesis.”
The book’s signature doctrines
Five propositions that hold the argument together.
Complexity must earn its keep.
Every added model, agent, memory layer, verifier or workflow component must improve the work enough to justify its cost and fragility.
Chapter 5
Evidence before confidence.
A fluent answer is not evidence. Confidence in a claim should rest on its sources, provenance, scope and limitations.
Chapter 8
Capability does not confer authority.
A system may be technically able to perform an action without being entitled to perform it.
Chapter 11
The right to interrupt.
A mature intelligent system must be able to stop when evidence conflicts, authority is unclear or its operating assumptions no longer hold.
Chapter 14
Institutions must retain sovereignty.
Organisations can buy external intelligence while retaining control over evidence, policy, memory and permissions, and the ability to challenge or replace the system.
Chapter 18
Knowledge/ Part III
A fluent answer does not establish what the system knows.
The wording tells you nothing about whether a claim was observed, reported, inferred or invented. The system has to preserve that distinction. Confidence should follow the support a claim has earned.
Try it · Chapter 7
The sentence stays the same. Its basis changes.
An operations model says: The shipment will arrive on Thursday.
Choose a basis to see what action the same sentence can bear.
Epistemic statusObserved
- Can bear:Customer promise
- Can bear:Qualified estimate
- Can bear:Contingency planning
“A live scan may support a customer promise.”
Epistemic statusSource statement
- Cannot bear:Customer promise
- Can bear:Qualified estimate
- Can bear:Contingency planning
“A timetable may support a qualified estimate.”
Epistemic statusInferred
- Cannot bear:Customer promise
- Cannot bear:Qualified estimate
- Can bear:Contingency planning
“A pattern may justify contingency planning.”
Epistemic statusAssumed
- Cannot bear:Customer promise
- Cannot bear:Qualified estimate
- Cannot bear:Contingency planning
“An invented date should support neither.”
“Epistemic status determines what action the claim can bear.”
Try it · Chapter 8
Three links. How many sources?
Imagine an AI assistant preparing a board note on a supplier’s financial condition. It finds three articles reporting a fall in revenue. All three quote the same company press release.
Apparent sources: 3Independent sources: 1
“They are three links but one source of the underlying claim.”
“The crucial move is not to retrieve more text; it is to understand what kind of support each text provides.”
Chapter 9 · Reasoning Provenance
Preserve the path to the decision.
Provenance is an external, inspectable dependency and decision history: sources, versions, tools, assumptions, rules and approvals. A reviewer can follow those connections and challenge them. A plausible explanation does not supply that record.
Chapter 10 · Unknown Unknowns
The frame can be wrong before the answer is.
Every retrieved fact may fit while a decisive fact is missing from the question. More reasoning alone does not establish that the frame is complete. The architecture needs ways to notice exceptions and invite a different account of the problem.
Five Parts · Eighteen Chapters
These are not isolated principles. They form one architecture.
Beyond the Model develops the argument across five movements: from model capability, through adaptive systems and machine knowledge, to agency, and finally to the human and institutional structures in which intelligence operates.
The examples here open the questions. The book works through their dependencies, trade-offs and limits.
Explore the full argument- Part I
After the Model
Capability, systems, and the economics of thought.
- Part II
Intelligence That Can Reason About Itself
Adaptive intelligence, complexity, and architecture as experiment.
- Part III
What Does the Machine Know?
Evidence, epistemic status, provenance, and unknown unknowns.
- Part IV
From Answers to Action
Agency, constitutions, bounded autonomy, and interruption.
- Part V
Intelligence Inside Institutions
Humans, organisational intelligence, computational governance, and sovereignty.
Agency/ Part IV
Intelligence creates possibilities. Agency creates consequences.
“A system that drafts an email is different from one that sends it. A system that suggests a code change is different from one that merges it. An agent can have a good understanding of a task and still lack the right to act.”
Chapter 13 · Bounded Autonomy
Autonomy belongs to actions, not to systems.
“A system may freely sort incoming tickets, prepare a refund, require approval to issue it and be prohibited from changing the refund policy.”
- Sort incoming tickets: Act with broad discretion
- Prepare a refund: Prepare
- Issue the refund: Act after approval
- Change the refund policy: Prohibited
- Observe
- Recommend
- Prepare
- Act after approval
- Act inside limits
- Act with broad discretion
“These are positions on a spectrum, not a ladder every application should climb.”
Try it · Chapter 13
Every step valid. The pattern is not.
“A single £40 refund may be routine. Twenty such refunds in an hour to the same account may be suspicious.”
Illustrative sequence limit: 20 per hour
- Each refund, alone
- Within limits
- The sequence
- Routine
“A series of individually valid steps can amount to an unauthorised strategy.”
Delegation is bounded by the action and its consequences: evidence, reversibility, novelty, observability and mandate all matter. A low transaction value alone does not establish permission to act.
Chapter 14 · The Right to Interrupt
Knowing when to stop is part of intelligence.
“An agent working through a set of account changes finds that the customer’s name matches but the account number differs from the signed instruction. It has enough information to continue filling fields and may even guess which number is correct. The valuable act is to stop.”
How the system stops matters as much as whether it stops. Compare two interruptions from the same chapter.
“I cannot proceed.”
“The payment details in the submitted invoice differ from the approved supplier record; confirmation from the authorised owner is needed before release.”
“The agent’s right to interrupt is ultimately a limit on the institution’s appetite for smoothness.”
Human + machine/ Part V
Humans are part of the architecture.
“A human reviewer receives a machine-generated decision with a green button marked ‘Approve’. The underlying evidence is hidden behind several clicks. The reviewer has thirty seconds before the next case arrives.”
Chapter 15
The question is not whether a person is in the loop. It is whether the loop gives them anything to judge with. Try being that reviewer.
Machine recommendation
Update the customer’s account details
- Model confidence
- High
- Customer name
- Matches
The account number differs from the signed instruction.
“The person is present in the workflow but absent from the judgement it claims to preserve.”
“A signature at the end of a process does not supply judgement retroactively.”
“A fast hierarchy with a distorted view is not an intelligent one.”
Chapter 17 · Computational Governance
Rules must reach the point of action.
A policy requires two approvals for a large payment. The software allows one person to release it. The institution’s declared rule and its operational authority have parted company.
“The policy exists in a document; the operational authority exists in software. When the two diverge, the software often wins.”
Computational governance makes appropriate institutional rules executable and observable: approval requirements, evidence checks, permissions and routes for escalation.
People still decide their purpose, judge exceptions and challenge their effects. The code must remain answerable to that process.
Institution/ Part V
Institutions must retain sovereignty.
“Institutional sovereignty is the practical capacity to set the terms on which intelligence affects an organisation, inspect those effects, revise the terms and continue functioning when a particular provider or component changes.”
Chapter 18
An institution can use powerful external models and tools while keeping its governing functions under its own control. Ownership alone does not establish sovereignty. The test is whether it can inspect, direct, challenge and replace the intelligence it relies on.
Inside the book
The full argument, chapter by chapter.
Each Part supplies the conditions the next one depends on. Browse the chapters, from the capability trap to institutional sovereignty.
Part I · Capability & cost
After the Model
Capability, systems, and the economics of thought.
- Chapter 01
The Capability Trap
“Capability is treated as a universal solvent, expected to dissolve poor context, ambiguous authority, missing data and weak evaluation. It cannot.”
- Chapter 02
Intelligence Is Becoming a System
“Adding a box does not make a system more intelligent. It may simply give failure another place to hide.”
- Chapter 03
The Economics of Thought
“A system that escalates everything to people has not solved the allocation problem. It has exported it.”
Part II · Adaptation
Intelligence That Can Reason About Itself
Adaptive intelligence, complexity, and architecture as experiment.
- Chapter 04
Adaptive Intelligence
“This also suggests a governance principle for self-improving systems: the ability to propose a change need not include the right to deploy it.”
- Chapter 05
Complexity Must Earn Its Keep
“The presence of several agents does not establish independent judgement any more than several copies of the same document establish independent evidence.”
- Chapter 06
The Architecture Can Become the Experiment
“We stop treating a diagram as an argument and start treating it as a hypothesis.”
Part III · Knowledge
What Does the Machine Know?
Evidence, epistemic status, provenance, and unknown unknowns.
- Chapter 07
The Epistemic Problem
“A model answers a question about a patient’s medication with a sentence that is grammatically perfect and factually wrong. It gives the same sentence structure when it is right.”
- Chapter 08
Evidence Before Confidence
“A model that counts sources without examining their dependence mistakes volume for corroboration.”
- Chapter 09
Reasoning Provenance
“The system should never have to invent a plausible history after the decision is made.”
- Chapter 10
The Unknown Unknown Problem
“The system may appear most confident when the frame is most defective: everything it can see fits neatly together.”
Part IV · Authority
From Answers to Action
Agency, constitutions, bounded autonomy, and interruption.
- Chapter 11
Intelligence Becomes Agency
“Better capability can justify wider delegation within a mandate. It cannot by itself supply the mandate.”
- Chapter 12
Agents Need Constitutions
“Several agents can contribute information, but they cannot vote themselves a new mandate.”
- Chapter 13
Bounded Autonomy
“A series of individually valid steps can amount to an unauthorised strategy.”
- Chapter 14
The Right to Interrupt
“The agent’s right to interrupt is ultimately a limit on the institution’s appetite for smoothness.”
Part V · Institutions
Intelligence Inside Institutions
Humans, organisational intelligence, computational governance, and sovereignty.
- Chapter 15
Humans Are Part of the Architecture
“The person is present in the workflow but absent from the judgement it claims to preserve.”
- Chapter 16
Organisational Intelligence
“A fast hierarchy with a distorted view is not an intelligent one.”
- Chapter 17
Computational Governance
“The policy exists in a document; the operational authority exists in software. When the two diverge, the software often wins.”
- Chapter 18
Institutions Must Retain Sovereignty
“The institution has bought intelligence and lost part of its ability to govern itself.”
Opens with Introduction: The Model Is No Longer the Whole Problem
Closes with Conclusion: The Frontier Moves Outward
Written for
The people who build, direct and question machine intelligence.
People building intelligent systems
AI architects, engineers, researchers, technical founders.
People responsible for what those systems do
CTOs, CIOs, senior leaders, governance and risk owners.
People thinking about institutions in an age of machine intelligence
Policymakers, investors, organisational thinkers.
A Threadneedle Labs book
Beyond the Model
The New Architecture of Artificial Intelligence
Zia Karim and Faris Karim
The model is becoming one component of intelligence. The architecture around it determines what that intelligence can know, how it may act and who remains in control.
First edition · October 2026
Five Parts · Eighteen Chapters · One connected argument