Designs the agent organisation, and names what must stay human
"Fifty-five AI agents to run a company" is a diagram, and it fails in contact with a real business for three predictable reasons.
The count is the wrong unit. Fifty-five agents is fifty-five things to configure, monitor, pay for and fix, and most will touch work that happens twice a month. What matters is which processes are frequent enough, rule-bound enough and evidenced enough to hand over — usually a much smaller number than the diagram.
The handoffs are where it breaks. An org chart draws boxes and the work lives in the arrows: what each agent produces, who checks it, what happens when it is wrong, and which named human owns the outcome. A diagram without arrows is a picture of a company nobody is accountable for.
And nothing in the genre says what should not be automated. Every company has work that is frequent, rule-bound, and still must not be handed over — because getting it wrong ends an employment, a licence, a treatment or a contract. That list is more valuable than the other one, and nobody publishes it.
This agent designs the operating model instead. It scores each process on whether it can genuinely be handed over, defines the handoff and the human owner for each, sequences deployment by dependency and payback rather than by enthusiasm, and produces the refusal list with the reason for each entry.
It deploys nothing and changes no organisation. Every output is a plan for people to decide on.
Six processes that genuinely work beats fifty-five that need configuring and watching.
And why, in writing, before someone tries it anyway.
Sequenced by what unblocks what, with the prerequisites named.
Frequency, rule-boundedness, evidence available, and the cost of being wrong. Most of a company fails at least one, and the diagram never checks.
What each agent produces, who checks it, what happens when it is wrong, and the named human who owns the outcome. The work lives in the arrows.
Work that is frequent and rule-bound and still must not be handed over. The most valuable list in the report, and the one nobody publishes.
Not by department and not by enthusiasm. Some of it cannot start until data exists that does not exist yet, and it says so.
Most of these plans fail on a prerequisite — a record nobody keeps, a system with no export, an owner nobody named.
It changes no organisation, buys nothing and configures nothing. Every output is a plan for people to decide on.
Why Operations in particular. The agent count is a marketing unit. What matters is the handoffs — who checks the output, what happens when it is wrong, who owns the outcome — and the refusal list, which no version of that diagram ever publishes.
Runs unattended
Started by you or by an event, and it finishes on its own. Nothing waits for someone to be at a desk.
The same standard every time
The two-hundredth item is held to the bar the first one was. Consistency is the part people cannot sustain.
It cannot act on its own
AI Operating Model Designer has no path to sending, spending or committing. That limit is why its output is safe to act on.
This agent runs server-side through the PROMIVO runtime. Each run is logged step by step and every tool call is permission-checked before it executes.
Demo dataIllustrative sample output, abridged.
{
"company": "A dealership group",
"currency": "USD",
"industry": "Automotive retail and service",
"headcount": 240,
"departments": []
}{
"escalate": true,
"sequence": [
{
"phase": "First — read-only, evidence already exists",
"unblocks": "Nothing depends on these, which is why they go first: they return value while the harder prerequisites are being cleared.",
"processes": [
"Warranty claim documentation check",
"Duplicate payment review"
],
"prerequisitesToClearFirst": [
"Load the manufacturer's warranty standard into a knowledge base."
]
},
{
"phase": "Second — needs data that does not exist yet",
"unblocks": "Capacity planning across all four sites.",
"processes": [
"Technician utilisation analysis"
],
"prerequisitesToClearFirst": [
"Job times are recorded on paper at two sites. Until that changes, no analysis of this is possible and no agent fixes it."
]
}
],
"disclaimer": "A design, not a deployment. Nothing has been configured, purchased or enabled; no headcount reduction is estimated and no recommendation is framed as replacing a role; savings are estimates from the figures you supplied and are stated net of the effort to supervise them; and no vendor or product is named.",
"assumptions": [
"Hours removed are calculated from the run frequency and durations you supplied, and are estimates.",
"Supervision effort assumes one reviewer per recommendation."
],
"recommended": [
{
"process": "Warranty claim documentation check before submission",
"readOnly": true,
"department": "Service",
"buildEffort": "Days",
"whoChecksIt": "The warranty administrator, who already reviews the claim run.",
"prerequisite": "The manufacturer's documentation standard has to be in a knowledge base. It currently exists as a PDF on one person's laptop.",
"agentProduces": "A list of claims that would fail, with the specific element missing from each.",
"agentReceives": "Repair orders with narrative, operation codes, parts and technician certifications.",
"whenItIsWrong": "A claim is submitted that would have failed anyway, or a good claim is queried. Both cost minutes, not money.",
"accountableRole": "Service Manager",
"capabilityNeeded": "Read each repair order against the manufacturer's documentation standard and report what would be reversed at audit.",
"netOfSupervision": 44,
"runEffortPerMonth": "About 2 hours of review",
"hoursRemovedPerMonth": 46
}
],
"countHonesty": "Six recommendations from 41 processes reviewed. A diagram of this company would have shown one agent per role and produced around forty. Thirty-five of those touch work that runs under twice a month, has no exportable evidence, or must stay human — and each would still have been something to configure, pay for and watch.",
"mustStayHuman": [
{
"reason": "Frequent enough and rule-bound enough to look automatable. It decides an employment outcome, and a wrong call ends someone's job on a machine's say-so.",
"process": "Deciding a technician's disciplinary outcome after a workshop incident",
"department": "People",
"couldAnAgentHelpAtAll": "It can assemble the evidence pack — records, timelines, policy clauses — so the person deciding has everything. It must not weigh it."
},
{
"reason": "A licensed technician's signature is the product. Nothing here can carry it.",
"process": "Signing off a vehicle as roadworthy after repair",
"department": "Service",
"couldAnAgentHelpAtAll": "It can check the job record is complete before it reaches them, which is a paperwork step and not a judgement."
}
],
"escalationReason": "Two processes the brief asked to automate decide employment and licensed sign-off outcomes, and one recommendation depends on job times that are currently recorded on paper.",
"notAnAgentProblem": [
"Chasing parts orders takes 30 hours a month because the supplier portal has no export. Fixing the export removes the work entirely; an agent would automate a workaround and make it permanent.",
"Two teams re-key the same customer record. That is one integration, not an agent."
]
}No integrations required.
What the company should actually hand over, and in what order.
Whether the process was ever a candidate.
A model with prerequisites and refusals, not a slide of boxes.
The map of what is automatable in a business they just inherited.
$599/month
Billed monthly through your PROMIVO subscription. Cancel at any time.
Runs consume your plan allowance for agent executions and tokens. See plan limits.
Almost certainly not, and that is the point. The count is a marketing unit. Most companies have a handful of processes frequent and rule-bound enough to hand over well, and a long list that would cost more to run as agents than to leave alone.
No. It designs the model. Deploying an agent into a live process is a decision with an owner, and it belongs to that owner.
It describes the capability each step needs — what it must read, what it must produce, what it must never do — so the description survives whichever tool you choose. Where a need has no product that meets it, it says so rather than naming a near-miss.
Then it says so with the reason, and that entry is treated as a finding rather than a gap. Work whose failure ends an employment, a licence or a treatment belongs to a person even when it is perfectly rule-bound.
No reviews yet. Reviews open once customers have run this agent.
Tell us what to change and our team will scope a customised version for your business.
Customize this agentRead-only by design. This agent has no path to sending, spending, publishing or committing anything. Where that limit is the product, removing it would remove the reason to trust the output.