Our method · The Kyroga OIS

Start with the problem. Technology comes last.

Most AI projects pick a tool and go looking for something to point it at. We start with the capability you want to improve, work out the decisions behind it, and choose technology only when we know what it has to do.

The Kyroga Organizational Intelligence System (OIS) is the standard every BrainBuild is built to.

The method in brief

The whole method, in eight lines.

What it is
The Kyroga Organizational Intelligence System (OIS) is a method for improving a business capability with AI, starting from the decisions behind it rather than from technology.
Design sequence
Capability, decisions, knowledge, reasoning, assurance, technology. In that order.
Architecture
Four layers: FIND, KNOW, DECIDE and VERIFY, with a management and improvement plane, and a loop that feeds what happens back into knowledge.
Doctrine
Default simple. Design for evolution. Escalate on evidence.
Ownership
Canonical knowledge is defined once and held independently of any database, model or app.
Maturity
Each layer is assessed on its own, for each capability, on levels 1 to 4. Never collapsed into one score.
Authority
Seeing, deciding and acting are separate permissions. Every agent works inside a mandate with a named human owner.
Test
Every engagement starts with an impact statement. No impact, no engagement.
The doctrine

Default simple. Design for evolution. Escalate on evidence.

Complexity has to earn its place. A higher level of capability means more cost and more to run, so it is never better by default. We build the least that works, and leave a clear path to more.

The design sequence

Six questions, in order.

Every engagement works through the same sequence. Skipping ahead to technology is how AI projects drift.

  1. 01 · CAPABILITY

    What are we improving?

    One capability, with an owner.

    The business capability to improve, and the measure that will show it worked.

  2. 02 · DECISIONS

    Which decisions drive it?

    Who decides, on what.

    Each decision's owner, trigger, inputs, options, and the cost of getting it wrong.

  3. 03 · KNOWLEDGE

    What must you know?

    To decide well.

    A map of meaning, sources and owners, including the gaps and the places people disagree.

  4. 04 · REASONING

    What reasoning does it need?

    Rules where rules work.

    Rules for what can be settled, signals for patterns, language models for ambiguity, people for judgement.

  5. 05 · ASSURANCE

    How much trust is needed?

    Scaled to consequence.

    Evidence, authority and audit matched to what each decision puts at risk.

  6. 06 · TECHNOLOGY

    What is the lowest build that works?

    Only now.

    The minimum sufficient implementation, with a credible path to grow.

Every technology choice must trace back: technology ← gap ← requirement ← decision ← capability. If it can't, we don't recommend it.

The architecture

Four jobs. One loop.

Each layer has one job. No layer quietly does another's.

Applications sit above

Workbenches · portals · agents · APIs. They use the system. They are not part of it.

FIND

Discovery

Reliably find what you already have

What might be relevant?

Finds candidates across documents, email, systems and past decisions, designed around the questions your people actually ask.

KNOW

Canonical knowledge

One agreed meaning

What do we know, what does it mean, why believe it?

Holds agreed definitions, relationships and provenance. Conflicting sources stay visible instead of being flattened.

DECIDE

Decision intelligence

Reasoning you can inspect

What should we do, and why?

Rules for what rules can settle, models for judgement calls. Every recommendation leaves a decision record.

VERIFY

Assurance

Trust, authority, audit

Can we trust it, and is it allowed?

Checks evidence, authority and policy before anything happens: approve, approve with conditions, escalate or reject.

Management and improvement plane

The people who own each layer can view, query, review, update and approve changes to it. It cuts across the layers; it is not a fifth one.

ACT→OBSERVE→LEARN→back into KNOW

What happens becomes what you know. Learning is validated before it's promoted, because something observed is not yet something true.

  • If a rule can decide it, don't ask a model to guess.
  • Claims carry receipts.
  • Similarity is not meaning.
  • Seeing, deciding and acting are different permissions.
Ownership

Knowledge you own.

The lasting asset isn't the app, the model or the database. It's your canonical knowledge: your concepts, definitions, policies and accumulated experience, held independently of any vendor.

Knowledge is defined once. Everything else is a projection.

  • Change the database, the knowledge survives.
  • Change the model provider, the knowledge survives.
  • Replace the app, the knowledge survives.
Maturity

Earned, layer by layer.

We assess each layer on its own, for each capability. A decision can need rich knowledge and light retrieval. A target is not a maximum.

F1 / K2 / D1 / V2F2 / K3 / D3 / V2

An illustrative profile: where one capability might sit today, and where evidence says it needs to be. Not a client result.

FINDDiscoveryKNOWKnowledgeDECIDEDecisionsVERIFYAssurance
1Question-led integrated retrievalStructured canonical knowledgeAssisted decision supportOperational assurance
2Enhanced, benchmarked retrievalIntegrated knowledge modelStructured reasoning harnessEnterprise assurance
3Specialised retrieval and relationship traversalKnowledge graph and formal semanticsFormal decision intelligenceRegulatory assurance
4Bespoke retrieval fabricAdvanced knowledge substrateAdvanced cognitive architectureHigh assurance
  • Never collapsed into one score.
  • Never set by company size or budget.
  • Unassessed until there's evidence.
  • Simple stays valid indefinitely.
Autonomy

Every agent has an owner.

Being able to see something, deciding on it, and acting on it are three different permissions. We keep them apart.

Knowledge access

What a person or system is allowed to see.

Decision authority

What they're allowed to recommend or approve.

Action authority

What they're allowed to actually do.

  • A named human owner

    Accountable for the outcome.

  • A clear objective and scope

    What it is for, and what it is not.

  • Operating limits

    Enforced where the action happens, not just requested.

  • Stopping conditions

    When it must stop and hand back.

  • An escalation path

    Who it goes to when it's unsure.

  • Inspect, interrupt, revoke, recover

    Owners can see active work and shut it down.

The engagement

How a BrainBuild runs.

Five phases, with the evidence labelled at every step. BrainBuild delivers it in Diagnose, Design and Implement; ClearBrain keeps it sharp afterwards.

  1. 01Prepare

    Context, goals, constraints, owners and sources.

  2. 02Discover

    Gather evidence, stakeholder language and how work really happens.

  3. 03Synthesise

    Form the capability, decision, knowledge, reasoning and assurance models.

  4. 04Validate

    Test assumptions with evidence and the people accountable.

  5. 05Evolve

    Improve the system through controlled learning.

  • Evidence

    Backed by a source.

  • Observation

    Seen, not yet generalised.

  • Inference

    A conclusion, with its reasoning.

  • Assumption

    A belief we still have to test.

  • Proposal

    A recommendation, traced to a need.

We keep what we know apart from what we think, and we label which is which.

One method, many applications

Different experiences. One system underneath.

  • A food manufacturer

    Recipe development workbench

    Finds what worked before, applies hard rules on prohibited ingredients, recommends an adaptation, and shows its evidence.

  • A security-solutions firm

    Complex quoting workbench

    Scopes, prices and routes bespoke quotes for approval across regions, using what the firm has learned from past work.

  • A coaching product

    AI mentor experience

    Guides people through a methodology, with progress and lessons that build on each other.

The impact test

We don't start with tools. If we can't state the impact before we begin, and measure it afterwards, we won't take the work.

Every engagement begins with an impact statement. Its evidence status travels with it, so a target is never presented as a result.

Impact statement · format

Illustrative, not a client result

Capability
The capability being improved
Decisions
The decisions that drive it
Baseline
Measured before we start
Impact
Stated in business terms, before we begin
Measure
How we will know, and who checks
Evidence status
  • Unassessed
  • Target
  • Baseline captured
  • Measured
Knowledge deposited
What the organisation owns afterwards
Questions

Common questions.

  • What is the Kyroga OIS?

    The Kyroga Organizational Intelligence System is the standard every BrainBuild is built to. It is a method for improving a business capability with AI, starting from the decisions behind the capability and choosing technology last.

  • How is this different from starting with a vector database and RAG?

    Those are projections of knowledge, not the starting point. We begin with the capability and the decisions behind it, then work out what must be known, what reasoning is needed and how much trust is required. Every technology choice has to trace back to a decision.

  • Do I need a knowledge graph or an agent platform to begin?

    No. The default is the simplest build that works, with a clear path to more. Graph capability and heavier reasoning arrive when the evidence says they have earned their place.

  • Which vendor or database does it require?

    None. Canonical knowledge is defined once and held independently, so databases, vector indexes, graphs, APIs and apps are projections of it that can change without losing what you know.

  • How are agents kept under control?

    Seeing, deciding and acting are separate permissions. Every agent works inside a mandate with a named human owner, a scope, operating limits, stopping conditions and an escalation path, and owners can inspect, interrupt, revoke and recover.

  • What does Kyroga claim about results?

    Nothing until it is measured. Every engagement starts with an impact statement, and its evidence status shows whether a figure is a target, a captured baseline or a measured result.

Find the problem worth solving first.

No impact, no engagement.