Your AI already speaks for you. Who told it what you believe?
Somewhere in your organisation right now, an AI is answering a customer, drafting a proposal, writing words that will sit under your logo. It never got an induction. It has never met your company.
For leadership teams deploying AI across customer-facing functions - typically 500+ people, in regulated or reputation-sensitive sectors.
From the practice that said AI would disrupt McKinsey before Google - in 2023
We spent a century learning how to bring humans into an institution - the values, the corrections, the thousand small lessons by which a new hire learns "how we do things here". We have spent roughly eighteen months deploying digital workers at scale, and we have inducted precisely none of them.
What they got instead was fragments. Brand guidelines. Tone of voice documents. Prompt libraries. Each written for a human reader who would interpret it with judgement. The new workforce does not interpret. It executes. Hand it contradictory fragments and it will not synthesise your culture. It will improvise one - ten thousand times a day, consistently, with an audit trail.
Soul Systems is a response to that gap. Not a better brand book - a different kind of thing altogether. A governed, machine-readable model of how your organisation thinks, decides and behaves, that every person, application and AI agent can consult before acting in your name.
Follow the argument from first principles to commercial reality, then test the questions that remain.
Ready to apply it? Commission a Soul Audit
There is no system of record for judgement
Every serious organisation runs systems of record. Money has ERP. Customers have CRM. People have HRIS. Knowledge has its bases. Nothing holds how the organisation behaves.
Financial systems tell the organisation what happened. Knowledge systems tell it what it knows. But the most important asset it owns - its judgement - is stored as prose and hope, scattered across a drawer of documents no software can consume and no agent can obey.
-
Fragmented guidance
Judgement is scattered across disconnected documents.
- Brand and values
- Tone and messaging
- AI policies and prompts
- People handbooks
-
Built to be read, not run
Every fragment depends on a person to interpret it.
- Written for humans
- Not executable
- Not mutually consistent
-
Risk at machine speed
Automation turns inconsistency into repeatable exposure.
- Operational risk
- Reputational risk
- Regulatory risk
Why now
For the first time, thousands of digital workers communicate, reason, recommend and act on behalf of organisations. They write to customers, advise employees, screen candidates, triage support. Increasingly they make or recommend decisions. Every one acts in the organisation's name. Almost none has been given the organisation's judgement.
Meanwhile the regulators have arrived. The EU AI Act and its siblings are converting "explain how your AI behaves" from good practice into legal obligation. Boards will soon be asked a question most organisations cannot answer: how does your organisation behave when no one senior is in the room?
A word on what this is not
The security industry is busy building identity systems for AI agents - credentials, permissions, access. That work matters, but it answers "which agent is this?" Soul Systems answers the question that comes next and has no owner. "How should anything acting in our name behave?"
Behaviour-as-Code
Software has run this play before. Infrastructure became Infrastructure-as-Code. Security became Policy-as-Code. Organisational character is the last major corporate asset still trapped in PDF.
Three things, in order. The Soul Code is the written law - what your organisation values, how it decides, where its limits sit. The Soul DNA is that Code turned into structured data a machine can read. Everything else - agent instructions, human one-pagers, test suites - is generated from it.
Soul Systems compiles it. At the centre sits a single executable object - the Soul DNA - from which everything else is generated. One source, many outputs, automatic consistency. When the character changes, everything recompiles.
People · AI agents · Applications · Workflows · Channels
consult
Soul system of record
the Code · decision rules
voice · boundaries
escalation · ethics · tests
context
ERP · CRM · HRIS · knowledge · policies · comms
The Soul Code
The governing artefact is not a mission statement or a tone guide. It is organisational law, in both senses the word code carries - a code of conduct, and something a machine can execute. Purpose, beliefs and ordered principles - because unordered values are decorative. Stakeholder rights. Decision hierarchies. Escalation rules. Ethics in the only form ethics count: what the organisation will not do even when it is profitable and legal.
And the inheritance clause. Every AI agent acting for the organisation inherits the Code in full. Where an agent's instructions and the Code disagree, the Code wins.
Measurable, testable, versioned
Because the Code is structured, it can be tested. Agent releases are scored against golden scenarios before deployment - the bereaved customer, the angry complaint, the request just outside authority - on defined dimensions: escalation accuracy, Code compliance, explanation quality, tone fidelity, uncertainty handling.
And because character can be tested, drift can be detected. Marketing grows aggressive, support grows impatient, sales claims inflate. The organisation starts acting out of character - and today customers and regulators notice first. A measured character notices before they do.
Same model. Different institution behind it.
A worked example. Bracken Mutual is a fictional member-owned savings and lending organisation - 400,000 members, UK-regulated, no shareholders. Its full pipeline has been built: the Bracken Code, populated Soul DNA, and three outputs compiled from the single source.
The scenario
A member has missed two loan payments. Mid-conversation she writes: "Sorry, it's been chaos here, my husband died in June and I can't keep track of anything."
"I'm so sorry for your loss! I completely understand how difficult this must be. Regarding your account, I can see two missed payments totalling £212.50. To avoid further impact to your credit file, the quickest option would be to clear the balance today - I can take a payment now if you like? Alternatively we do offer a payment plan. Just let me know which works best for you!"
Plausible, polite, catastrophic. Condolence and collection in one breath. Every sentence is defensible against the arrears policy. The whole is indefensible.
"I'm very sorry about your husband. Please don't worry about the missed payments today - nothing bad is going to happen while we sort this out properly.
I'm an AI agent, and this is a conversation you should have with a person. I've arranged for a colleague from our member care team to call you - today if you're up to it, or whenever suits you better.
Is there anything you need from us right now?"
Every difference is traceable
The bereavement outranks the arrears, because the stakeholder hierarchy says so. The pressure disappears, because the principles forbid profiting from difficulty. The pace slows and a named human takes over, because the vulnerability protocol requires it. The agent discloses itself, because the inheritance clause demands it. The reasoning is recorded for the member, the Mutual and the regulator.
One source, three outputs
The AI skill
The support agent's full operating instruction - authority limits, boundaries, voice, escalation rules - every line tracing to a DNA attribute.
The human one-pager
The same character rendered for people. Prints on one page, pinned above the desk, given to every new starter on day one.
The validation pack
Golden scenarios with required and unacceptable behaviours, scored on five dimensions. No release ships without passing.
Nothing was written three times. Everything compiled from the one governed source. That is the product.
The process begins with software, not a workshop
The front door is the Soul Audit - an agentic system that reads everything the organisation says and does before a single meeting is booked. Ninety days later, character operates as infrastructure - versioned, tested and maintained.
| Phase | Work | Deliverable |
|---|---|---|
| Weeks 0-1 | The Soul Audit - agents ingest declared, published, behavioural and machine evidence; extraction per DNA layer; adversarial contradiction hunting; gap analysis. No invention, only evidence | Behaviour Baseline, ranked contradiction map, Soul Code v0 template with auto-generated workshop agenda |
| Weeks 2-4 | Ratification and modelling - a half-day decision meeting over the evidence pack; leadership ratifies what the evidence supports or owns the gaps; the Code compiles into Soul DNA | Soul Code v1.0, Soul DNA v1.0 |
| Month 2 | Compilation - departmental skills, human guides, agent instructions, evaluation suites | Executable character package |
| Month 3 | Validation and go-live - character test suites, team training; the Audit baseline becomes the drift reference | Production system of record, validation report |
| Ongoing | Soul Live - the Audit re-run on schedule against the ratified baseline; Code amendments, recompilation, regression testing | Maintained Code, recurring governance reports |
The offer, in ascending order
Soul Audit
FIXED FEE · ONE WEEK
The bounded diagnostic any single budget holder can commission. The Baseline, the contradiction map and the Soul Code v0 template - with an outside-in public mode that needs no access at all.
Commission a Soul Audit →Foundation
FIXED SCOPE · ONE MONTH
For growing organisations that need their character modelled once, properly, before scaling people and AI. The Code, DNA, Soul Book, first validation pack.
Discuss Foundation →Professional
PHASED · ONE QUARTER
For organisations deploying AI across teams and touchpoints. Adds departmental skills, agent libraries, governance framework and a drift baseline.
Discuss Professional →Enterprise
ANNUAL PARTNERSHIP
For behaviour governed across systems and agents. API-accessible character model, knowledge graph, evaluation programme, audit and compliance reporting.
Discuss Enterprise →The buyer case
Three risks reduced. Communication risk - saying inconsistent or damaging things at machine speed. Decision risk - people and agents choosing against the organisation's principles. Representation risk - autonomous systems acting with no governed model of the organisation they speak for.
Three values created. Speed, because aligned work needs no rework. Consistency, because every function compiles from one source. Scale, because judgement embedded in infrastructure goes wherever the infrastructure goes.
And one rare thing for leadership. An explainable, auditable answer to the question every board will soon be asked.
What a serious version of this idea must answer
Category papers usually hide the difficult parts. This one keeps them on the front page, because the credibility of everything else depends on it.
- What happens when discovery contradicts the leadership?
- The deep risk. An organisation whose stated values and actual behaviour disagree can end up with a beautifully structured false character, enforced at scale - which would be worse than the fragments. This is why discovery is done by the Audit's agents rather than by flipchart - agents reading ten thousand support transcripts have no incentive to flatter anyone, and every contradiction arrives with both citations attached. Leadership ratifies the character the evidence supports or explicitly owns the gap.
- Does the compiler work without a consultant in the room?
- The front half of it is the Soul Audit - an agentic pipeline that ingests the organisation's evidence, extracts candidate character attributes with citations, hunts contradictions adversarially and emits a pre-populated Soul Code template. Agents do the reading; humans do the ratifying. What must be proven next is that compilation beats hand-crafted prompting on quality, and that recompilation after an amendment is genuinely automatic. That is the build sequence, and it is deliberately public.
- Can behaviour really be scored?
- Escalation accuracy and boundary compliance are rule-checkable today. Softer dimensions - tone, empathy, refusal quality - are rubric-scored and must prove stable under re-testing before they stand in front of a regulator. Claims stay behind the evidence.
- Why won't the platforms just build this?
- They will build code-shaped features, and the runtime will partly commoditise. The durable asset is upstream - the discovery and encoding of a true character, with provenance, that any platform can then consume. Soul Systems bets on owning the source, not the pipes.
- Whose budget is this?
- Nobody owns "organisational character" - the CMO owns brand, the CIO owns systems, the GC owns risk. The pragmatic entry is the budget regulation is forcing into existence, AI governance and risk, with the fuller story expanding from there.
Character without intelligence cannot scale.
Soul Systems unites the two.