Arquin Framework Protocol

An evolving methodology for applying AI transformation within an organisation. It sets out our language, approach and what to expect from working with us.

LAST VERSION PUBLISHED: 14 Oct 2026
|
FIRST PUBLISHED: 24 Aug 2026
|
Public specification / v1.1 / AFP
|
DREW POPPLE
|

This document sets out how Arquin believes artificial intelligence should be introduced into a business, and why.

Most major organisations now have an opinion on artificial intelligence. Some are willing to be early-movers, while others operate a more cautious “wait and see” approach. Some have introduced general chatbots like Claude or ChatGPT, others have opted for narrow vertical tools, and of those who report integrating agents, only a handful have built an operating layer they can actually rely on. It’s fair to say, that most organisations are only beginning to scratch the surface of what artificial intelligence can do for an organisation. We believe that leveraging the power of advanced AI capability will be achieved through careful orchestration and rewired, embedded workflows. Or, more simply, through structured AI transformation. 

AI transformation is an emerging discipline. This document outlines Arquin’s structured AI transformation framework. It also covers AI maturity, the shape of an organisation’s internal AI team and the tenets on which Arquin is built. 

The term AI transformation is used loosely across the industry, to describe a workshop, a chatbot, a platform licence and a genuine change to how a business operates, often within the same conversation. This document exists to remove that ambiguity for anyone reading it. It sets out our thinking plainly.

To understand how AI can transform an organisation, it helps to understand how AI actually works. At a fundamental level, AI is a prediction machine. The better the data, in quality and in quantity, the more certain the outcome. That is also where AI differs from most of what came before it. Earlier waves of enterprise technology were deterministic*. AI is not. Every subtlety and every edge case matters, and governance and guardrails become essential rather than a nice to have.

Effective transformation also depends on the organisation, not just the technology. It needs alignment across senior teams, and clear ownership. A sponsor, a project lead, a pioneer team, and someone accountable for outcomes. Arquin works alongside those people from the initial concept and roadmap through to governed, embedded workflows that measurably improve how the business operates.

We stay agnostic throughout. Whether a workflow needs an open weight model, a frontier model, or a narrow vertical tool is a decision made on evidence, not preference, and it is where our organisational thinking, systems thinking, and technical judgement come together. Indeed, we help define what parts of the business AI shouldn’t touch or where standard automation is a better option. 

This document is written for organisations at any point on that journey, from an early interest in artificial intelligence through to a business that already has AI involved in its core processes. Later sections describe stages of maturity and there is no expectation that a reader will recognise their organisation at the final one. Recognising an earlier stage is equally the purpose of the document.

A note on how to read what follows. This is our thesis, our approach and the standards we hold ourselves to and it is written to stand on its own. It does not include the specific mechanics we apply inside a live engagement, the scoring, the thresholds or the exact deliverable formats, since that is the part a client engages us for. What follows is intended to give any reader, whether or not they ever work with us, a clear and honest account of how we believe this should be done.

* Deterministic

In the context of AI, deterministic means a system that produces the same output every time it’s given the same input, with no variation. Most traditional software works this way. Generative AI doesn’t. It produces probable outputs rather than guaranteed ones, which is why governance and guardrails matter more here than they did with earlier technology.

WHERE YOU STAND

AI maturity specturm

Every business sits somewhere on the AI maturity spectrum. In early stages Arquin can work in an advisory capacity and in the latter stages we work with our strategic partners to design and build full AI transformation programmes.

Some organisations may only reach stages 3 or 4 while more sophisticated, innovative and ambitious businesses will progress through to stage 5 and 6. 

  1. Elementary An organisation has taken no or limited steps to adopt AI in their daily operations. This can be due to a number of factors including: “wait and see” approach, anti-AI sentiment, no buy-in from leadership, problem choosing an approach, AI initiative stalled, budgetary restrictions.
  2. Limited Sporadic use of AI tools and chatbots across individual teams. Generally layered on-top of the organisation and poorly integrated with limited governance, scalability, documentation and repeatability. Individual enthusiasm and scattered tools with no shared operating layer yet.
  3. Experimental AI use exists beyond scattered individual tools. At least one workflow, or part of one, has been deliberately mapped, built and put into use. The work sits apart from core operations rather than inside them, run more as a proof of concept than as something the business relies on.
  4. Partial integration Vertical tools and enterprise platforms are wired into the organisation’s existing tech stack. AI agents operate inside these platforms, independently alongside them, or both. AI now touches real operational data, not a sandbox, though its reach across the business is still uneven.
  5. Embedded AI has a dedicated internal team with clear roles and accountability. Workflow candidates are assessed against governance, reliability, compliance and data readiness as standard practice, not case by case. Multiple AI-embedded workflows, some coordinated as multi-agent systems*, run in production, with errors logged and reviewed rather than absorbed silently.
  6. Post-transformation AI-native workflows emerge from within the organisation itself, rather than being introduced from outside. Existing systems are maintained and optimised as a matter of course. Capability and intelligence compound over time. The organisation is AI enabled at a high level of maturity.

* Multi-agent systems

This is a simplification of how this happens in practice. During the Blueprint and Specification stage we make decisions on which parts of the workflow need AI agents (including type of model), which need standard automation and where AI shouldn’t be involved or touch certain data.

KEY TERMS

Transformation language

A handful of terms recur throughout this document. Each describes something specific we have built, learned, or committed to. Defining them once here avoids re-explaining them every time they appear.

Ambient AI

The state almost every business is in before an engagement starts. Scattered tool adoption, individual enthusiasm, no shared operating layer. Ambient AI produces energy and anecdotes, never a system anyone can point to and say “that’s what runs this.”

Deterministic core

The part of a workflow that must behave the same way every time, regardless of what a model outputs. We decide explicitly where this applies, where a model should augment judgement, and where a human decision is required. Blurring these boundaries is where most unreliable systems come from.

Legacy workflow

Any workflow that predates AI-based judgement, however well it currently runs. Legacy here isn’t a comment on quality, in the way legacy code isn’t necessarily bad code, it just means the architecture was built before this generation of tools existed. The Blueprint’s job is deciding which legacy workflows get rewired, and in what order, not assuming all of them need to be.

Judgement debt

Risk that accumulates inside a system that looks like it is working, because nobody has yet been forced to ask what happens when it is wrong. Every ungoverned pilot carries judgement debt, whether or not it is ever called in.

Integration line

The point where AI stops being a named initiative and disappears into how the business runs. Crossing it is the actual definition of success. An initiative still being talked about as an initiative has not finished yet.

Tacit capture

The task of drawing out judgement decisions a person can’t fully explain even when asked directly. They know which request is genuinely urgent because years on the job taught them a rule they’ve never put into writing. This can’t be found by reading a process document, it has to be drawn out directly, through structured conversation or watching the work happen. 

Undigitised knowledge

Information that already exists in a usable, written or spoken form, a note, a marked-up document, an email, a verbal agreement from a meeting, but sits scattered across paper, inboxes and memory rather than a single digital record.

Throughline

The one line of accountability that runs across a build, from architecture through delivery to operation. Splitting a build across vendors is fine, but losing the throughline between them is not.

The trust gate

The point in a system where unverified output becomes verified fact. Nothing crosses it without a check, whether that check is a rule, a person, or both. It is usually the single most important design decision in any system we build.

Living system

A system that’s live and doesn’t stay static. Models improve, regulation shifts and the business changes shape around it. Optimisation and maintenance alongside periodic reviews are designed to reduce future drift. 

THE FRAMEWORK

Introduction to transformation

Any team approaching AI transformation needs to have a clear roadmap, programme and work within a structured framework. The Arquin framework is broken down into four clear phases.

  • Discovery: acquire an understanding of how the organisation operates, where data is kept and what type of data it is. It includes mapping existing workflows and identifying candidates for initial AI embedding.
  • Blueprint: Rewire the workflow/workflows with AI embedded. Sense check across organisation and key team leads.
  • Specification: Liaise with tech/implementation team to transform organisational AI workflow(s) into schematic and tech spec. Compliance and cybersecurity evaluation.
  • Orchestration: Workflow(s) built as functioning and documented system. Benchmarks set, performance monitored and system tweaked and adjusted to improve performance and reliability. User testing and further adjustments. Optimisation of system and start of double run with live data. Once approved system goes live. This builds foundation for pause and evaluate and/or scale up to further workflows. 
THE COMMERCIAL CASE

Business case, ROI and enablement

ROI on an AI initiative shouldn’t be measured the way you’d measure a software purchase, but should drive some of the decision making during Discovery. 

Usage figures and adoption dashboards aren’t value, they’re activity. The real measure sits against the specific decision or process the system was built to change. Time to resolution, error rate, consistency, cost avoided.

If we can’t say in advance what a workflow needs to do differently once it’s live, it isn’t specified well enough to build yet.

The KPIs get defined during the Blueprint, before a line of the system is built, not retrofitted afterwards to justify what’s already been spent.

What you get

A finished engagement leaves you with more than a working system. It leaves an internal team able to run and extend what’s been built, a documented trail for every decision the system makes and workflows designed to scale rather than needing to be rebuilt when volume changes. The measure of a well-specified system is that judgement debt has been paid down at the point of build, not left to surface later.

What it costs

The shape of investment varies with the number of workflows involved and how much complexity sits inside each one, and that scope gets decided together during Discovery rather than quoted before either side understands the work. What stays constant is the sequencing: the business case and its KPIs are established early, so cost is always weighed against a defined outcome, never against activity alone.

TRANSFORMATION TEAM

Your internal AI team roles

Effective transformation depends on the organisation as much as the technology. These are the roles we look for at the start of an engagement, and the gaps we flag when one is missing.

Sponsor

The senior team member able to remove obstacles and connect the work to strategy and to keep it funded. Without a sponsor, there’s a risk that transformation stalls while decisions are made and budget is found. 

Project lead

The single point of contact who keeps the work moving day to day, coordinates between Arquin and the business and makes sure the right people are in the room when a decision is needed. Usually not the most senior person involved, but the one who makes the schedule a reality.

Pioneer team

Generally these are the dynamos of AI in your business. They are innovators from across the business who build adoption of AI across their teams. They are the source of ground truth about what really happens, not what the process document says happens and the first to work with a new system once it’s live. Without them, Discovery is theoretical.

Accountability lead

The named individual who owns outcomes and risk once a system is live, and can explain what it did, why and what happens when an output is incorrect. This role exists from the first specification, rather than appointed after something goes wrong.

Data owner

The person who knows where the relevant data lives, how sensitive the data is, what state it’s in (structured or unstructured) and who else needs to sign off before it can be used. 

Technical liaison

Owns how the system will sit inside existing infrastructure, security and IT policy and works directly with the engineering team who build it. Without this role, integration decisions get made multiple times by the business and IT and the implementation team.

PHASE ONE

Discovery

A structured, current state map of how the business actually operates on a daily basis.

Every workflow has an official sanitised ideal version and a messier real version. The gap between the two is often where risk lives. Discovery exists to surface that gap before anyone starts building. We need to understand what data actually exists, in what state, owned by whom, where decisions genuinely get made versus where they’re formally signed off and which processes are stable enough to automate.

What it reveals

Almost every Discovery session surfaces at least one process that leadership believes is standardised and isn’t, running instead on the accumulated judgement of one or two people who’ve never written it down. That’s not a criticism of the business, it’s simply something that has to be uncovered before anything gets rewired. Embedding AI into an undocumented process usually just amplifies the inconsistency.

Uncovering it usually means two different kinds of work: tacit capture, drawing out judgement someone can’t fully explain even when asked directly, and gathering up undigitised knowledge, information that already exists but was never centralised.

What good looks like

  • A single honest current-state map of existing workflows
  • An honest and transparent understanding of the state of data within the business
  • Alignment on which workflows are best suited to being rewired with AI embedded
  • A clear view of the commercial case for the workflows selected, detailed further in Business case, ROI and enablement.
Discover's expanded output format, scoring criteria and the specific data-readiness checks we run as part of an engagement, are not in this document.
PHASE TWO

Blueprint

Translation of the current state map into an AI embedded map. Usually we start by focusing on one or two workflows while keeping a holistic view of the whole organisation. This then allows us to scale the process should company-wide transformation be required.  

This is the judgement layer, and it’s where most of the risk in AI adoption actually sits. Two workflows that look equally suitable from the outside can carry entirely different risk profiles once you account for data volatility, regulatory exposure, and how much tacit human judgement is embedded in the “simple” version of the task. Part of the Blueprint phase is ranking and sequencing rather than just listing possibilities. Not everything that’s technically a good workflow candidate should be translated first, or at all.

What good looks like

A real Blueprint is unrecognisable from one business to the next, because it’s built entirely from that business’s own Discovery session. An adept blueprint is not just a 1:1 adaption of a legacy workflow, but an entirely rewritten map that is suitable for technical development and agentic orchestration.  

PHASE THREE

Specification

The point at which a prioritised idea becomes a buildable, accountable schematic with defined inputs, outputs, failure modes and an understanding of who owns the outcome.

This is the phase that separates a serious system from an enthusiast’s project. Anyone can produce working code quickly now; model-assisted development has made that trivial. What it hasn’t made trivial is deciding what “working” needs to mean for a specific business-critical process: how the system behaves under bad data, what happens when it’s wrong, who’s accountable when it is and how failure gets logged, surfaced and corrected rather than silently absorbed. This is where judgement debt either gets paid down or quietly compounded.

The Specification is a pivotal point where the Blueprint is translated into tech spec and schematic by the development team and reviewed by compliance and cybersecurity experts to surface risk, compliance and governance issues.  

The test we apply 

Could this specification be handed to a third-party engineering team with no further context, and would they build the same system we would? If not, it isn’t a specification yet, it’s an idea.

PHASE FOUR

Orchestration

Ongoing coordination of the technical build, the organisational rollout and the compliance requirement, held by one party across the full lifecycle rather than split across vendors who each own a slice.

Orchestration is not implementation. Arquin does not sit inside the codebase writing the system; that’s deliberately a separate technical function, brought in and directed rather than performed in-house. Arquin’s role is to provide oversight, which is the thing that tends to get lost when a build is split across a platform vendor, an internal IT team and an external developer. It’s a single throughline of accountability for whether the finished system still matches what the Blueprint and Specification actually called for.

Most AI rollouts don’t fail during the pilot. They fail in the handoff, months later, between the people who scoped it and the people who built it at scale, with no one left holding the original intent. Orchestration is the phase that ensures what is built matches the Blueprint and Specification, is well documented and produces reliable outputs. Before anything goes live, the system enters a pilot: running in parallel with the existing process on live data, so performance and reliability are proven under real conditions rather than assumed from testing. Once the pilot is approved, the system goes live. 

Any change to the build should flow back through the schematic to the blueprint to ensure clear source material and version history. This is also where a workflow crosses the integration line, the point it stops being a visible initiative and simply becomes part of how the business runs.

IN PRACTICE

Use case walkthrough

Every phase above sounds abstract until it’s run against something real. Here’s a short, anonymised pass through a live build. A professional-services network intelligence tool, built to turn spoken voice notes into a trustworthy answer, allowing the partner to ask in natural language “who do I know in finance in Paris?” or “I’m going to Hong Kong again, who should I meet from my contacts?”

Discovery

The stated ask was simple: capture a network just by talking about it, then query it in plain language. Underneath sat two very different products. One that sounds impressive in a demo and one that’s actually trustworthy enough to base business decisions on. The question that mattered wasn’t which model to use. It was what the system should do the moment it’s asked about someone who doesn’t exist in the data or edge cases that needs to be accounted for, a decision most builds never make explicitly and the one this entire engagement turned on.

Blueprint

The obvious, flashy option was pointing a language model straight at the raw notes and letting it answer conversationally. It was ruled out for the same reason a client-facing chatbot gets deprioritised elsewhere. It produces something legible and untrustworthy, confident enough to invent a plausible contact who doesn’t exist. The harder, correct sequence was to build a deterministic core first, and only put a generative layer on top once it’s structurally incapable of fabricating.

Specification

This is where that decision became mechanical rather than aspirational. Every person captured from a voice note lands as unconfirmed and stays invisible to any query until a human reviews and confirms the data is correct (name, organisation, location etc): the trust gate. Every question runs a deterministic search before language gets involved at all. An exact match first, then explicit, defined fallbacks, before ever returning an honest empty result, labelled by exactly how hard it had to reach. A language model sits on each end, parsing the question in and phrasing the answer out, but it’s constrained to narrate only what the underlying data actually returned, forbidden from inventing anyone, the one guarantee most competing products can’t credibly make.

Orchestration

What shipped was the complete, tested loop end to end, including the case that matters most. When asked about someone who definitively isn’t in the network yet, it says so, rather than inventing a name. Just as deliberate is what didn’t ship in this pass, and that it’s written down rather than discovered later: cross-note deduplication, relationship traversal, external enrichment, multi-turn conversation. A named next phase combined with user feedback.

CONTROLLING OUTPUT

Governance

Governance isn’t an approval step we add after a system is built. It’s designed into the workflow from the start, alongside the workflow itself, not layered on top once someone in compliance asks for a review.

Responsibility for an outcome always sits with a named human role, decided before the system goes live, not worked out after something goes wrong.

AI can recommend, automate, or act, but responsibility for the outcome always sits with a named human role. Ownership, escalation paths and authority need to be explicit at every decision point.

A well designed system reduces the load on the person using it day to day. It should also stay completely legible to whoever is accountable for it, able to explain what the system did, why, and what it’s allowed to do next. A system that demands constant interpretation from its users won’t scale. A system that hides its own logic from the people accountable for it is a liability, however well it performs.

Governance is about building control into a non-deterministic system, one that says “one plus one probably equals two”. This process is about ensuring the output provided is reliable and, if it isn’t, making this clear to both the user and flagging it in a log. 

The roles set out earlier, sponsor, project lead, pioneer team, accountability lead, are what a business needs to get an engagement built and launched. Once a system is live, a narrower set of roles takes over.

The reviewer is the person actually performing the check at the trust gate, deciding whether a specific output moves from unverified to verified fact. This is a genuine, ongoing job, not a one-off sign-off, and needs to be resourced as one.

The escalation owner handles the cases the deterministic cascade or the model itself can’t resolve with confidence. Not every low-confidence result belongs with the reviewer, some need a different kind of judgement, and without a named escalation owner those cases tend to sit unresolved rather than get decided.

The calibration lead periodically checks whether the system’s overall behaviour is still accurate, not just whether individual outputs pass review. This is the human side of what a living system actually means in practice, drift gets caught here before it becomes a pattern.

These three can sit with one person in a small deployment, or be genuinely separate roles in a larger one. What matters is that each job is named and owned, rather than assumed to be covered by whoever happens to be accountable overall.

CONSIDERED FROM OUTSET

Compliance

Compliance requirements differ by sector and territory and we map them during the Discovery, Blueprint and Specification phases, not after Orchestration.

That changes the sequencing of a project as much as it changes the paperwork. A workflow that touches regulated data or a protected decision gets specified differently from the start, with the Discovery trail, the retention rules, and the escalation path built in rather than retrofitted.

We don’t treat compliance as a separate workstream running alongside the technical build. The two are specified together, because a system that’s technically elegant and non-compliant isn’t finished. 

THE BUSINESS ASSET

Data

AI is a prediction engine, and prediction is only as good as the data behind it. That’s the plain version.

The harder truth is that data readiness isn’t a single yes or no answer. Most businesses have data that’s excellent in one place and unusable in another and waiting for a perfect data estate before starting anything is one of the most common ways an AI initiative stalls before it begins.

The ideal approach is for a data hygiene process to happen before beginning any AI work. However, we can sequence around what’s usable now and remediate what’s actually blocking the highest value process first, rather than commissioning a data cleanup project as a precondition for everything else. Discovery is where this gets decided, process by process, not as a blanket judgement on the whole business.

FINE TUNING THE SYSTEM

Optimisation

We build first for capability and reliability, then optimise for efficiency. A system needs to prove itself against real production conditions before it’s earned any tuning, which is why optimisation only starts once it’s live and logging real outcomes.

Selecting the right model type and size for each individual task, rather than defaulting to the most capable model available for everything, is one of the clearer hallmarks of a system that’s been carefully and strategically engineered. 

Many tasks inside a workflow don’t need a large, general-purpose model once the task itself has been precisely defined by real usage. Swapping in a smaller, narrower model for a well-understood task makes commercial and operational sense. It costs less to run, it responds faster and for a task that’s already been proven, it’s often just as reliable, sometimes more so, because a narrow model has less room to behave unpredictably on a job it’s been matched to precisely.

This isn’t a one-time pass. As the business, the workflow and the available models keep changing, tuning is something we return to at regular intervals, exactly what a living system means in practice.

THE ARQUIN TENETS

The principles Arquin is built on

Arquin is not a traditional consultancy adapted to shoehorn AI. We are an AI-native consultancy built upon strong principles that drive our approach to transformation.

  1. Independence and agnosticism We design to operate independently of any single vendor, model, or platform. We believe in providing AI solutions that fit the requirements of the business and are entirely independent of any platform or vendor. This means you get a uniquely impartial viewpoint and opinion on AI.
  2. Built to be scaled Every strategy, pilot or workflow we build is designed to be scaled. Whether you ask us to help with vertical tool selection, existing tech stack implementation or build a custom implementation we consider how the architecture can be built upon in the future.
  3. Expert partner network Larger implementations and transformation require multi-disciplinary teams. We are diligent in our selection of expert partners to ensure they work to the highest level and standards expected of Arquin.
  4. Client discretion and confidentiality We ensure complete confidentiality for our clients and always act with discretion. 
  5. Limited volume Arquin works with a small number of businesses at a time and only those who have a shared vision and collaborative style of working.
  6. Integrity If your project, style of working or initiative doesn’t match our expertise and ethos we will be honest about it. We believe in building strong long term working relationships with our clients built on professionalism, mutual respect and integrity.
  7. Hybrid working Arquin works in a hybrid way. In person when necessary and practical and remote when not required. This enables us to expedite progress and ensures budget is allocated to the right places. 
  8. End-to-end involvement Arquin are the architects of your AI implementation but even when we hand the baton over to compliance, cybersecurity or a development team, we keep oversight to ensure consistency in what was planned and what is produced. 
  9. Resilience Geopolitics, regulation and technology are increasingly in flux. We plan any implementation with resilience as a consideration.  
  10. Designed for invisibility Well-designed AI reduces cognitive load for users, integrates seamlessly into existing workflows and should be broadly invisible to the day-to-day function carried out by your organisations. At the same time, its behaviour, constraints, and controls remain legible to leadership and governance functions.
  11. Intuitive design Human interfaces with AI should be intuitive and user friendly. While we cannot ensure this of third party tools and platforms, our bespoke builds are designed by UX professionals with ease of use paramount for inputs, outputs and tracking/logging.
  12. Judgement training While we make efforts to build intuitive systems an important skill is learning where and how to exercise human judgement. This forms the backbone of our post-implementation training. Not just how to use the system, but when and where it’s appropriate.
NEXT STEPS

How to get started

Bringing intelligence into an organisation in an effective way is vastly different from previous eras when building capability was achieved through hiring or buying software/hardware. It’s a serious undertaking which will touch every part of your business and where missteps can be costly. It requires buy-in at leadership level and a clear vision of what your AI enabled organisation looks like.

While we have borrowed from practical experience and best practice across sectors, Arquin was built from the ground-up for AI implementation and transformation.

This document outlines our thesis, ethos and approach to AI work. However we preserve the details of how we do it for our trusted clients.

Whether you are looking for us to work in an independent advisory capacity or ready to move into a methodical approach to AI transformation, we adapt to meet an organisation’s needs.

We recommend any organisation to either start with advisory work or begin a Discovery session. To take the next step you can email me directly at drew@arquin.ai or via the contact page