In brief: An Owner can now get a decision pack in an afternoon that used to take an analyst three days — and the decision itself is no cheaper, because judgement, mandate and answerability cannot be delegated to an agent. This briefing is for business owners, managing directors and business-line leads whose evidence is arriving faster than ever, and for the leaders they report to. The practical test: an agent may prepare, compare and challenge, but if a recommendation arrives carrying authority, someone has handed over a mandate that was never theirs to give.
An Owner’s evidence is suddenly cheap; their judgement is not. This briefing follows one investment decision through an Owner’s hands — what agents can prepare, what only the Owner can decide, and the questions that keep the two apart.
Frauke Lehmann is the managing partner at Bruns & Lehmann, a family-owned Mittelstand manufacturer of dosing pumps in northern Germany — around 420 people, customers across Europe, an ERP that predates most of the vendors now pitching to replace parts of it. On Friday she takes an investment decision to the family’s advisory board: whether to fund the automation of supplier-order intake that operations has been asking for since spring.
The signals on her desk disagree. Operations has counted the hours lost to intake errors and wants to move now. Finance has read the vendor quote twice and does not believe the payback period. A key account has complained about confirmation delays for the third time this quarter. And IT has flagged that the integration touches the ERP in a place nobody has touched since 2019.
Months ago, Frauke built herself a small assistant for exactly this kind of week — a single constrained agent with a clear brief: pull the delivery reports, summarise the complaint log she is permitted to read, and draft the decision pack in the house format. Nobody assigned it to her. She started it on her own pain, the way the first agent in a firm usually starts, and IT later gave it a proper identity and locked down what it can reach.
The pack it produces is good. What it cannot contain is the decision — and this briefing is about keeping that line exactly where it is while the pack gets better and better.
What an Owner actually holds
The Compass groups the delivery roles of a transformation as Transformers — people, not the neural-network architecture; the difference between the two is exactly the accountability gap this briefing works. The Owner is the Transformer who ties an initiative to strategy and answers for the result: direction, value trade-offs, the mandate to commit resources, sponsorship when the initiative needs defending, and the consequential approval that makes everything downstream legitimate.
None of that appears in the decision pack. It appears in what happens after the pack. Someone with standing says yes. That person commits money and people. And if the yes was wrong, the advisory board, the works council or a key customer holds that same person responsible. An agent can hold the evidence. It cannot hold the standing.
What agents change for this role
For Owners, agents change one thing profoundly: evidence becomes cheap. The pack that took an analyst three days now takes an afternoon; the scenario comparison that never got built because nobody had time now arrives by default. Judgement does not get cheaper at all — and that asymmetry creates this role’s specific failure mode: the well-formatted recommendation that nobody actually owns. A pack can be so complete that it reads like a decision already taken; sponsors start approving the formatting. The Owner’s discipline is to treat a better pack as better raw material, never as a smaller decision.
In my work, this failure has recently grown a second face. Work that is visibly AI-inflated — too polished, the recognisable AI language, slides assembled from generated graphics rather than built by a person — is met in senior rooms with more distrust by default, not less. The questions arrive immediately: has this person actually verified all of it? Did this deck take thirty minutes? That is the contradiction leaders now live inside. Everyone is enabled, even pushed, to use AI — responsibly — and yet the people in the room still give their trust only when they can see your own work, your validation and your own opinion in the material. The polish is not the credential. The visible judgement is.
From one assistant to an Owner’s suite
Frauke’s single assistant is the first of the three shapes from the previous briefing. If the decision load grows, the honest next step is not a bigger assistant but separated purposes — a small suite around the role:
- A briefing and evidence agent — assembles the pack from declared sources, and only from declared sources, with every number traceable to where it came from.
- A scenario and trade-off agent — builds the two or three futures the committee should compare, with its assumptions printed, not implied.
- A challenge agent — briefed to argue against the emerging recommendation: the strongest case for not doing this, or doing it later, or smaller.
- A portfolio-signal monitor — watches the initiatives already running and flags drift early. It watches; it never acts.
What is deliberately not on this list is an owner agent — something that weighs the trade-offs and recommends with authority. The moment a recommendation carries authority, someone has delegated a mandate, and mandates are the one thing this role cannot hand over. If a vendor’s slide shows an “AI decision engine” for leadership, the two questions from the previous briefing apply unchanged, and the answer to the first one had better be no.
Cost note: every agent on this list is another brief to keep true, another thing to re-evaluate when a model or a policy changes, and another output crossing the Owner’s desk — the suite’s real price is paid in her attention before any invoice arrives.
In my work with PathPatron, these roles exist — but not always as separate agents. I call on them repeatedly across tasks, most of all the challenger, briefed as a devil’s advocate against our own articles, our evidence and my arguments. For new project ideas I run a practice I call “Call the board”: agent perspectives on finance, market research, product–market fit and technical feasibility, with a legal or ethics view added when the topic demands it. And the practical honesty is this: you can prompt a single agent into all of these roles, or store the brief as a reusable instruction set, without building a suite at all. The suite pays off when the roles are needed on a regular basis. For one-offs, in my view, a suite is overkill.
None of this requires system integrations on day one. Most of what an Owner’s suite reads — the delivery report, the vendor quote, last quarter’s minutes — already crosses her desk as documents, and handing them to an agent by hand is an honest first shape: lower fidelity, no permission battles, and enough to learn whether the suite earns real connections. Where access to systems does become the wall — an IT mandate against connecting AI tools, or simply nobody who knows how — the next briefing in this Path takes that blocker on properly, because it bites the Organizer hardest.
Owners are already building the first shape
The pattern in Frauke’s story — the role-holder starts on their own pain, engineering hardens it later — is not invented. Two German craft-trade owners documented doing exactly this in 2026. Jennifer Kirschbaum-Konsek, a sheet-metal master — Spenglermeisterin — about to take over her family’s roofing firm, built a ChatGPT-based agent that applies her trade’s calculation rules to roof and sheet-metal cut-offs and flags the classic errors; it was grown case by case from real jobs, not delivered as an IT project. And Jonas Winkler, a master joiner, built a quote agent on his own past quotes and invoices after investing more than a hundred hours in a project that never paid — his words on why enquiries went unanswered before: sometimes the effort was simply too big, so they stayed lying there (KI im Handwerk — IT-Matchmaker News, August 2026 — self-reported, single-agent, small scale, published with OpenAI’s involvement, and both owners are themselves well-known trade content creators, so read it as illustration rather than evaluation). What survives the vendor tint is the order of events: the person who knew the work decided what the agent was for, what it could read, and what it must never touch. The boundary came from the owner, not the tool.
The Owner’s data and evidence map
Frauke’s signals disagree: operations has counted the lost hours, finance doubts the payback. That disagreement is not noise for an agent to clean up. It is the most important thing in the pack, because it shows her where the real decision sits. An Owner’s evidence design has four parts, and all of them are decisions, not configurations.
Internal sources. Authoritative sources are declared before the pack exists. If the ERP’s order data and the ops team’s spreadsheet disagree, which one wins is not a question the agent may settle by picking the more convenient number: it surfaces the conflict, and the pack carries both figures with their sources.
External sources. The Owner names which outside sources count — the industry association’s statistics, the analyst service the company already pays for, a named publication. Two checks follow. Whether the agent may actually reach the paid sources, because licences often do not allow automated access — a contract question before it is a technical one. And what the agent does when it cannot: one that quietly substitutes open-web material for the paywalled source it could not reach fills the pack with the internet’s opinion instead of the firm’s evidence base. Anything not on the declared list is labelled unverified — visibly.
The exclusion list. Some material never enters the system at all: margin structures, personnel judgements, the family’s own capital discussions. The Owner decides this list first, the way EMES Kabelbaum Konfektion, the Saxon switchboard manufacturer from the opening briefing, decided its margins would not leave the firm before it built anything else. In practice, the workflow shows the seam. The exclusion is enforced by permissions, not politeness — the assistant’s identity simply cannot reach the costing tables. The pack carries the gap visibly, as a marked line: “margin impact — out of scope for the assistant, assessed by the Owner.” Frauke adds that layer herself, on one page the system never sees, before the pack goes to the advisory board. And the brief forbids the workaround that matters most: the agent may not estimate its way around a gap it was told not to fill.
Designed challenge. A challenge agent briefed to disagree is genuinely useful — it forces the recommendation to survive an argument. It is not independent challenge: it reads the same sources, runs on the same model, and inherits the same blind spots. Real independence needs separate purpose and separate evidence, sometimes a separate human — a later briefing in this Path takes challenge design properly. For now the Owner’s rule is modest: know which kind of disagreement you are getting.
The four questions that travel
Everything above compresses into four questions. Frauke asks them of her own pack before Friday; they work just as well on a vendor pitch, a colleague’s proposal, or a “decision engine” demo.
- What can this prepare that it must never decide? If the answer is unclear to anyone in the room, the boundary does not exist yet.
- Which sources are authoritative, and who declared them so? “The agent found it” is not a source.
- Who challenges the recommendation — and is the challenger just the same model in a different hat?
- What happens when the evidence conflicts? The design answer must be that conflicts surface to the Owner unresolved — an agent that averages a disagreement away has taken a decision nobody gave it.
If the Owner is on your team
Half the readers of this briefing are Frauke. The other half lead a Frauke — a business-line or product owner who now arrives with agent-prepared packs. Your job then is not to re-take their decisions; it is to check that the boundary this briefing describes actually exists. Three checks do most of the work.
- Ask who decided. The answer must be a name, given without hesitation.
- Ask to see the seams — the declared sources, the marked gaps where excluded material was assessed by hand, the challenge that was considered and rejected with reasons. An Owner who can show the seams is running the agents; one who cannot is being run by them.
- Watch the tempo. When packs arrive faster but the decisions in them stop changing, formatting has begun to substitute for judgement.
You are the senior room from earlier in this briefing — your instinct to distrust over-polished work is a management instrument, not a prejudice. These three checks are the short version. Leading people who work with agents — the checks, the warning signs, the review cadence, for all six roles — is covered in its own briefing later in this Path.
The boundary, kept
On Friday, Frauke’s pack goes to the advisory board with the operations case, finance’s counter-case, three scenarios with printed assumptions, and one page her challenge agent produced arguing for a six-month delay — followed by her written reasons for rejecting the delay. The reasons are the point. The pack is agent-made and says so. The decision has one name on it, and it is not the assistant’s.
That is the Owner’s version of the series’ claim: agents redistribute the work of deciding — the gathering, the drafting, the stress-testing — and leave the decision itself exactly where it was. An Owner who understands that gets more decision-ready weeks per year. An Owner who forgets it gets a very well-formatted way to stop being an Owner.
Related reading
- One Agent, a Suite, or an Agentic Workflow? — the three shapes and the two questions this briefing applies.
- Delegate, Automate, or Keep Human: The Agent Handoff Map — assigning authority level by level, once the Owner’s boundary is drawn.