AI support
AI does a lot of work in this platform, but under two firm conditions: claims must be supported and a person decides what counts.
Grounding: supported by the dossier
"Grounding" means that an AI statement can be traced back to material the customer supplied themselves. As soon as you add documents to the dossier, they are split into coherent blocks of text and made searchable by meaning (not only by keyword).
When a worksheet asks for an AI first draft, the system first looks up the relevant passages and passes them along as context. Three things matter here: one limit of the system, and two properties that set this apart from a typical search-and-generate set-up.
A limited selection. A selection of passages goes along, not the whole dossier — the best scoring ones. When extracting, the number grows with the size of the dossier, up to 24 pieces, and a large document does not crowd out a small, fitting document. Whatever fits best goes along.
An explicit "no evidence" outcome. If the search returns nothing — because there is nothing searchable in the dossier — the system says so, instead of inventing a finding about the dossier. What happens then depends on the action: extract looks only at the dossier and then returns nothing; generate may use general knowledge and the rest of the canvas context, clearly marked. Since the end of August 2026 that also applies to mission, vision and ambition: a first draft then rests on the project data and the other core fields, with a line saying what it rests on — a starting point, not a finding about the organisation. Only if there is no context at all does the worksheet say so.
Only source documents count as evidence. Previously generated AI text can never come back as a "source" for a later generation. That prevents an assumption from gradually promoting itself to fact.
The evidence chain: references that are checked
When AI makes a claim, it must refer to the specific fragment the claim rests on. The server then checks whether that fragment was actually in the set provided.
That distinction is essential. Asking a model to "give your sources" produces plausible-looking references that nobody checks. Here every reference is verified, and a reference that is wrong is rejected or visibly marked — depending on what is sensible in that place:
- In reporting, the generation is flatly refused; a report with an invented source is worse than no report.
- On the worksheets, the item stays but is marked as untraceable, so that one weak reference does not throw away a whole batch of useful suggestions.
The result is a source accountability: for each claim you can see which file and which page the evidence comes from. That reference stays readable, even if the underlying document is later re-indexed.
Three actions, the same everywhere
Every AI button on a worksheet is one of three actions, each with its own symbol and its own word:
| Action | What it does | What it draws on |
|---|---|---|
| Extract | Retrieves what is in the dossier, nothing added | only the dossier |
| Generate | Writes a first draft where the worksheet is still empty | dossier, worksheet and general knowledge |
| Improve | Rewrites text that is already there, without new content | the text itself |
You can see that the AI is doing something: while it works, every AI button shows the verb, a running counter and the usual duration, and says so if it takes longer. On the worksheets where proposals arrive one by one, they appear as a card as soon as the model writes them and are already saved. Clicking away or refreshing while a proposal is running that has not yet been saved asks for confirmation first.
As long as you still have to decide, a proposal visibly shows where it comes from — from the dossier, from what is elsewhere on the canvas, or from general knowledge. If you accept it, that mark disappears from the screen and the origin is saved along with the text, so that you can later check what a sentence rests on.
That also applies when AI fills in a form. If it fills in the fields of a process from the dossier, that is a proposal: Save accepts it, Cancel rejects it.
What the AI knows about your project
When generating, the data of the intake and the canvas go along as context: scope, sector, type of transformation, size, country and occasion, the central question, the urgency, what has been tried before, the intended result, the criteria for «successful when», what is in and out of scope, the horizon, the starting points and the running programmes, and what you have already recorded on the canvas (mission, vision, ambition and strategic themes). That applies to the generation tasks, the One-pagers and the AI chapters of the reports. The names (organisation, lead, people involved), the planning, the budget frame and the working agreements do not go along. The AI knows today's date, so a proposed improvement action is not in the past, and it writes in the language of the canvas, even if the dossier is in another language. That language rule appears in two places in every task and applies by default to all models, including models that think first.
Merging duplicate issues
On the analysis worksheets, when generating clusters the AI also proposes merging issues that say the same thing. Accepting replaces them with one issue and the originals are kept in an archive; rejecting leaves everything as it is.
Where the limit lies
Origin is a convenience, not a seal of approval. It is helpful to know where a sentence comes from; the certainty that the sentence is right is and remains the judgement of the consultant who accepts it. The platform records for each accepted text which source it came from and whether it was taken over verbatim or derived, and shows that where it is useful — when reviewing before you share anything, and in the source accountability afterwards. The platform does not judge whether a source is right, and it does not replace a review by a subject-matter expert. There is one place where we are strict: what the consultant takes to their customer. A report that is generated and goes to a customer unread is not how this product is meant to be used; that is why every report asks once, before downloading, for your confirmation that you have reviewed the content, and the verdict lives in a document of its own that you compile deliberately, and not hidden in a view.
The review gate: a person decides
Every AI proposal appears recognisably as a proposal — never as part of the analysis until you have accepted it. You always have the same three choices:
| Action | What happens |
|---|---|
| Accept | The proposal becomes a definitive part of the analysis. |
| Edit | A pre-filled form opens; you adjust it and then accept. |
| Reject | The proposal disappears. |
This pattern is uniform across the platform. An AI function that silently deviates from it counts as a design flaw — not a handy shortcut. That is essential as soon as you have to account for your conclusions: what is in the final report has demonstrably been approved by a person.
Which model is running
Which models the AI uses is set in administration and not in the code, and a new model first runs alongside on a trial bench before it goes into use. See The AI layer.
Where AI does and does not come in
| AI does | AI does not |
|---|---|
| Summarise documents and extract data from them | Decide what is true |
| Cluster loose issues into themes | Save data on its own |
| Write first drafts for themes, principles, capabilities | Invent a mission or vision without any context about the organisation |
| Draft summaries, findings and advice in reporting | Use its own earlier output as a source |
| Flag coherence and contradictions | Take over the judgement |
The judgement lives in reporting
Until early September 2026, every worksheet had an insights window: an AI analysis as a reading document, without buttons. That window is gone. The judgement — findings, cross-connections, advice — now has a document of its own per worksheet in reporting, alongside the worksheet report that only displays and the management summary that only describes. That keeps the limit sharp: what is in the workspace has been approved by a person; what is a judgement is stated as a judgement in a document that you deliberately take to a customer.