Steering with AI without handing it the decision.

Three concrete uses

1. 01An annotated dashboard for the executive committee

   AI Act level · Minimal risk

   1. 01Committee questionsWhat management must decide
   ↓→
   2. 02Defined indicatorsOne definition, one owner
   ↓→
   3. 03Sourced commentaryVariances explained, sources cited
   ↓→
   4. CheckApproval by the indicator owner
   ↓→
   5. 04Meeting papersSent after approval

   The problem

   Every month, each department presents its own figures and the meeting is spent reconciling the gaps. Nobody knows which version is authoritative.

   What the AI does

   An agent queries the indicators defined in the semantic layer, drafts a commentary on the variances and cites each source. It corrects no figure and settles no discrepancy: it flags them to their owner.

   Data needed

   The committee’s indicators, their written definitions and extracts from finance, sales and HR.

   Who signs off

   Each indicator owner approves the commentary before the meeting papers are sent.

   Minimal under the AI Act (EU): the agent summarises aggregated data and makes no decision about people.

   • Semantic layer
   • AI agent
   • Human oversight

2. 02A sourced decision memo for every trade-off

   AI Act level · Minimal risk

   1. 01Question framedThe decision, in one sentence
   ↓→
   2. 02Documents gatheredInternal sources, restricted access
   ↓→
   3. 03Options comparedCosts, risks, sources cited
   ↓→
   4. CheckReview by the file’s sponsor
   ↓→
   5. 04Reasoned decisionRecorded in the minutes

   The problem

   An investment, an acquisition or a reorganisation is decided on hastily drafted memos whose assumptions are scattered across dozens of documents.

   What the AI does

   AI gathers the internal documents, sets out the options, their costs and their risks, and links every statement to its source. It does not recommend in your place: the choice and its rationale remain the committee’s.

   Data needed

   Business plans, contracts, minutes and internal studies, with access restricted to committee members.

   Who signs off

   The file’s sponsor reviews the memo and corrects the assumptions before it reaches the committee.

   Minimal: internal drafting support. The main risk is leaking confidential information, which hosting and access rights must address.

   • RAG
   • Hallucination
   • AI hosting

3. 03A private assistant instead of consumer tools

   AI Act level · Minimal risk

   1. 01Usage inventoryWhat teams already do
   ↓→
   2. 02Usage policyAuthorised and prohibited data
   ↓→
   3. CheckPolicy approved by executive management
   ↓→
   4. 03Private assistantHosted, logged, partitioned
   ↓→
   5. 04Quarterly reviewUses, incidents, adjustments

   The problem

   Your staff already use consumer AI tools, sometimes with contracts, customer data or unpublished figures. You know neither what, nor where, nor on what terms.

   What the AI does

   An assistant hosted under your control, with guardrails and logging, covers everyday uses: drafting, summarising, translating. It does not approve what gets shared: each user reviews the output, and the policy states what must never be entered.

   Data needed

   No data to migrate at the outset; a list of authorised and prohibited categories of information.

   Who signs off

   Executive management approves the usage policy and the list of authorised uses; each user remains responsible for what they share.

   Minimal while the assistant is used internally and its output is reviewed. It would fall under the transparency obligations (EU · AI Act, Article 50) if it talked to your customers.

   • Guardrails
   • Logging
   • AI hosting