AI on every step. A person's name on every claim.
Assettia is the data and AI programme record: one place where an organisation keeps what its programme owns, what it has decided, and what it has proved.
Most software that uses AI asks you to trust the output. Assettia is built the other way round. Around 150 capabilities across the product draft from the organisation's own evidence, inside frameworks refined over twenty-five years of consulting work, and nothing a model produces becomes a record until a named person has read it and accepted it. That is what makes the AI usable in a boardroom rather than only in a workshop.
Three things, working together
Assettia is not a model with an interface in front of it. It is three things, and the AI is only useful because of the other two.
- Encoded expertise. The scoring criteria, the assessment structures, the question sets and the eight workstreams come from twenty-five years of running data and AI programmes at the world's leading consultancies. A model given no framework returns a plausible paragraph. A model working inside a proven framework returns a scored, comparable, defensible case.
- AI throughout. Drafting at every stage: reading transcripts, proposing use cases, scoring and merging, generating cases, summarising assessments, identifying risks, writing the first version of a policy. The table below is the whole of it.
- Evidence. Every draft lands in a reviewable state. Every acceptance carries a name and a date. Every figure a reader sees is computed from records and names the population it counted.
Together: the judgement of a top consulting team, applied by AI at every step, with a record a board can check.
One pattern, everywhere
AI is woven through every module rather than bolted onto one. Every capability follows the same pattern, and the pattern is built into a component each module uses rather than left to each team to remember.
A model drafts. The draft is shown for review. A person edits it if it needs editing. A person marks it reviewed, and only then does it feed a report or an export. The platform will not treat generated text as final until that has happened, and it warns before overwriting anything a person has already edited.
Where the AI reads a document rather than writing one, the same rule applies in a stricter form. Uploaded documents, transcripts and presentations are read, the relevant passages extracted, and structured entries proposed: a stakeholder, an answer, a value measure, a plan item. Each proposal arrives in a staged queue with the passage it came from attached. A person accepts or rejects each one. Nothing imports directly.
What the AI drafts, by area
Every task below runs in the product today. In each case a model drafts and a person accepts.
| Area | What a model drafts |
|---|---|
| Strategy and board reporting | Board-ready narrative sections, executive summaries and strategy insights, written from the organisation's own recorded objectives, decisions and portfolio position, so a board pack starts as a structured draft rather than a blank page. |
| Discovery and readiness | Turns assessment responses and interview evidence into narrative report sections, cluster groupings, ambition statements and strategic insights, in plain, board-appropriate language. |
| Use cases and value | Suggests data and AI use case ideas from the organisation's objectives, drafts benefits cases and value narratives, proposes the measures a case should track, and matches new ideas against the existing portfolio to avoid duplicates. |
| Interviews and evidence intake | Generates interview guides and question sequences from a brief, then works backwards from the outputs: reads uploaded documents, transcripts and presentations, extracts the relevant passages and proposes structured entries for review. |
| Data estate | Drafts data strategy, data quality and data literacy summaries, suggests how data sources map to use cases, and describes system assets from catalogue metadata. |
| Propositions and markets | Drafts proposition descriptions and value statements, suggests target sectors, summarises market scans and validation evidence, and identifies comparable evidence with a written account of how confident it is. |
| Roadmap and prioritisation | Drafts roadmap narratives, suggests theme phasing, scores and merges initiatives, and rewrites initiative language into plain English. |
| Governance | Identifies candidate AI risks from the organisation's own use cases and data assets, and drafts policy sections for responsible AI, security and acceptable use. |
| Presentations | Drafts slide and deck content, chapter structures and the primary message for a playback session. |
What happens when the evidence is thin
A model asked to summarise nothing will write something. Assettia does not let it. Where there is not enough recorded evidence to produce a confident output, the platform says so and returns nothing, in these words:
The model could not produce a confident output from the available evidence. Add more source material or interview notes, then try again.
An empty result is shown as information, not padded out. It is the same discipline that reports an empty register as empty rather than as zero.
Atlas, the guide inside the workspace
Atlas is the AI assistant inside the product. It behaves like a senior chief of staff rather than a chatbot: it opens in a panel beside the work, it is scoped to one organisation's workspace, and it reads only that organisation's own records. In its own words, Atlas will:
- Answer questions about this business, grounded in its own data
- Point you to the exact page, tab or module you need
- Draft use cases, enrichments and evidence lookups for your review
- Never write to your workspace without your explicit confirmation
The questions it is built for are the ones a leader actually asks. What should I focus on this week. What has changed since my last board report. Which use cases are highest value and most feasible. What is in Now, Next and Later, and why. How ready is the business for AI. Which teams look weakest on decision readiness. What are my top five risks right now.
Every figure it quotes comes from the records it can see. It says plainly when the evidence is thin, and it will not cite a number it cannot find. When asked to produce something, it first shows exactly what it intends to do; only on confirmation does the work run, and the output lands as drafts for review. Each draft can be accepted, rejected or undone, and every action is logged. Where a request is beyond it, it says which part it can help with and which it cannot.
Atlas is in preview. It is enabled for an organisation rather than assumed, and there is a single switch that turns it off everywhere.
How the AI itself is governed
The prompts are not buried in the code. Every one lives in a managed library where its wording can be read, changed, versioned and tested.
- Versioned. Every change to a prompt keeps the previous wording, with who changed it, when, and a version number. An earlier version can be restored.
- Tested. Prompts carry test suites with assertions about what an output must and must not contain, run on a schedule and on request, with the results on a screen rather than in someone's memory.
- Tuned by us, on your behalf. Prompt wording is held by Assettia and is the same for every organisation, so no one at a client can quietly change how the AI behaves, and two organisations get the same discipline.
- Logged. Every AI call is recorded with the model used, the response time, the cost and any error, so usage and cost can be seen per organisation rather than estimated.
Changes to an organisation's own voice settings are recorded in the same way: who changed them, when, and what the previous wording was, with an earlier version restorable.
Your organisation's voice
The wording of the prompts is ours. The voice of the output is yours.
Each organisation sets its own AI voice and context: how it refers to itself, the vocabulary of its sector, terms it does not want used, how formal the writing should be, and a short standing context about the organisation that every draft takes account of. Settings are drafted first and reach the AI only when published, and they apply to that organisation alone.
Changing them is restricted to the organisation's own administrators, so the voice of every draft is not something any user can quietly alter.
The plain-language rules are written into the prompts themselves: no coaching language, no buzzword stacks, British English, hedged where a claim is unproven, and never a figure the platform cannot evidence.
Which models, and where they run
Model requests are made from the server, never from the browser, and routed through a gateway to Google's Gemini models and Anthropic's Claude models. Which model serves which capability is a decision the platform makes and an administrator can change; a client never chooses a model, and no model is exposed directly to a user.
Client data is not used to train any model, under the platform's own policy and the providers' business terms.
Each organisation's records are isolated at the database, and the AI respects that isolation: one organisation's data never feeds another's output.
What the AI never does
Four rules, held by the software rather than by good intentions. They are not a limit on the AI. They are what makes it usable where it matters.
- 1. A model never produces a figure. Every number a reader sees is computed from records by a fixed rule and names the population it counted. No model output is treated as a figure.
- 2. A model never writes to a live record. Output lands in a reviewable state and becomes a record only when a named person accepts it.
- 3. A model never appears on an executive surface as a finding. The Command Centre and the board report show facts derived from records. Where composed narrative appears, it is labelled.
- 4. A model never hides a gap. A register with nothing in it says so. An assessment not yet run is shown as not yet run, never as zero. Where the evidence is too thin, the platform returns nothing and says why.
Why this is a stronger position, not a weaker one
Every AI product reaches the same moment in a boardroom. A director asks whether the number in front of them can be trusted, and whether a person or a model produced it. That question decides whether the AI is used anywhere that matters.
Assettia answers by construction. The figure names what it counted and opens to the list. The narrative beside it carries the name of whoever approved it and the date they did. The model's contribution was the draft, and the draft is on the record too.
An organisation that cannot answer that question will use AI in workshops and nowhere near the board. An organisation that can will use it everywhere.
Questions people ask
Does AI write the board report?
A model may draft parts of it. The report is produced from the records by fixed rules, every figure is computed rather than written, and a named person reviews and approves it before it is filed with its version.
Which models does Assettia use?
Google's Gemini models and Anthropic's Claude models, through a gateway, from the server. A client never chooses a model and no model is exposed directly to a user.
Is our data used to train models?
No. Client data is not used to train any model, under the platform's own policy and the providers' business terms.
Can the AI act on its own?
No. It drafts and it proposes. Every draft waits in a reviewable state, and every proposal from Atlas waits for a confirmation before anything runs. Nothing reaches a live record without a person accepting it.
Can we change how the AI writes?
Your administrators set your organisation's voice: how it refers to itself, your sector's vocabulary, terms you do not want used, how formal the writing is, and a short standing context every draft takes account of. The prompts themselves are held by Assettia and tuned on your behalf, so the discipline is the same for every organisation.
What happens if the model gets something wrong?
The person reviewing it rejects or edits it, and the record is unaffected because nothing was written. Where a draft has been accepted and later found wrong, it can be undone and the change is logged.
Does the AI see other organisations' data?
No. Each organisation's records are isolated at the database, and Atlas is scoped to one workspace. Nothing is pooled and nothing is compared across clients except under explicit agreed rules, anonymised, or not at all.
Is the AI optional?
Atlas is in preview, enabled for an organisation rather than assumed, and can be switched off entirely. The record works without it; the drafting surfaces simply require the work to be typed rather than reviewed.
See what the AI drafts, and what you approve.
A walkthrough takes forty minutes and uses a synthetic organisation, so nothing of yours is entered until you decide it should be.