The AI C-Suite Is Already Here. It Just Doesn't Look Like a C-Suite.

HOVR is a Canadian mobility company founded by Harrison Amit. It publicly names four AI executives: Mia for operations, Tia for technology, Sia for strategy and marketing, and Ria for regulatory and administrative work. Harrison has also described roughly 250 agents that can be created for a workload and dismantled when they drift.

That sounds like a company replacing its executive team with software. The evidence points to something more practical and more consequential.

HOVR also names human executives. Harrison remains founder and CEO. The company's December 2024 offering document described a conventional management team and treated HOVR GPT as a future milestone. Public accounts show substantial AI work across functions, while corporate authority and accountability remain human.

The interesting shift is management becoming software

Most AI adoption begins at the task level. A person asks for a summary, analysis, draft, or piece of code. The person carries the context, decides what happens next, and restarts the process tomorrow. The four-level adoption map places this in the early stages: useful work is happening, but coordination still lives with the person.

An executive layer changes the unit of automation. The system holds a standing role. It remembers the company's policies and prior decisions. It wakes on a schedule, receives work from a shared queue, calls tools, coordinates specialists, checks results, and routes exceptions to a person. The useful output is no longer a single answer. It is continuity across a recurring management loop.

This is why the best examples feel more operational than a chatbot and less autonomous than an executive. They turn parts of coordination into infrastructure. People still set direction, handle ambiguity, approve consequential action, and answer for the result.

What HOVR appears to have built

The strongest current account comes from Harrison's May 2026 ChannelNEXT presentation. An independent event report says HOVR automated dispatch, onboarding, marketing, and data-structuring workflows. Harrison described Mia as an assistant that evolved into an AI COO able to query data, surface insights, and support decisions in real time. He associated Tia with technology, Sia with strategy and marketing, and Ria with regulatory work. He also said agents are retired when they drift or produce misinformation. Read the ChannelNEXT case report.

HOVR's own public material has advertised four named AI executive roles and a lean human team augmented across every function. It claims a 75 percent lower operating cost. The same material names Harrison as CEO, Gurjant Singh as CTO, and Chris Alicpala as CMO. That overlap matters. It supports a shadow-role interpretation: an AI COO can organize operational work while human leaders retain accountable authority. Review HOVR's public site.

PureBrain appears to supply at least part of the underlying system. Its public product description emphasizes orchestration, persistent memory, tool access, and a catalogue of more than 100 agents. In a vendor-hosted testimonial, Harrison said Mia works across the organization and produces useful results about eight times out of ten during beta. That figure is unusually candid. It also describes a system that requires verification, not unattended authority. See the PureBrain platform and testimonial.

The historical record sharpens the interpretation. HOVR's December 2024 offering document listed about 35 employees, a human CEO, COO, CFO, CTO, marketing leader, and security leader. It planned more than 80 hires over two to three years and listed a HOVR GPT for operations and logistics as an uncompleted Q3 2025 milestone. HOVR may have built a large agent layer after that filing. The filing prevents us from turning the later AI-native story into a claim that software held corporate authority from day one.

The evidence-calibrated conclusion is straightforward: HOVR appears to run a substantive, human-governed shadow C-suite. Confidence is medium. Public evidence supports recurring AI work across executive functions. It does not establish replacement of corporate authority, fiduciary responsibility, or final human judgment.

A shadow C-suite separates work from authority

A useful operating model has three layers. Human leaders set policy, priorities, risk limits, and approval rules. Executive or functional agents maintain a persistent lens on operations, technology, marketing, finance, or compliance. Specialist agents and workflows handle narrower tasks beneath them.

The return path matters as much as delegation. Data and work products move upward with evidence. The system surfaces exceptions. Verification catches false completion and drift. Requests for consequential action stop at a visible human boundary.

The boundary should be based on consequence, not seniority theatre. External communication, money movement, legal commitments, employment decisions, policy changes, safety, and irreversible actions need a person who understands the context. That person must also bear the consequences. Routine research, reconciliation, drafting, testing, and status work can sit deeper inside the synthetic layer.

This division also matches the NIST AI Risk Management Framework. It calls for organizations to distinguish human and AI responsibilities, define human oversight, and keep executive leadership responsible for AI risk. The framework does not prescribe a shadow C-suite. It does reinforce the boundary that makes one governable.

This design also makes failure easier to diagnose. If a marketing agent drafts a weak campaign, the system can identify the role, prompt, inputs, model, reviewer, and outcome. If an imaginary AI CEO simply announces that it has made a strategic decision, the title hides the mechanism that needs inspection.

“AI executive” describes five different operating modes

Public examples become confusing when a branded humanoid, a non-voting board observer, a functional role-agent, and a compressed service team all share the same label. A clearer taxonomy asks what work the system performs, what evidence it leaves, and where authority stops.

At IHC, Aiden Insight is formally integrated as an AI board observer. IHC's initial disclosure explicitly calls the role non-voting. Aiden analyzes financial and operating data, provides predictive insight, and recommends actions. Later company reporting says it generates reports, manages workflows, and executes strategic tasks. The board and C-suite still choose which recommendations become company decisions.

NetDragon gave Tang Yu the ambitious title of virtual rotating CEO. Its company account describes an AI analyst that supports organizational affairs and strategic execution. Formal reporting is more specific: the 2023 annual report describes business-process automation and structuring, while the 2024 annual report describes AI pre-execution, pre-evaluation, and pre-decision support. The title is stronger than the disclosed authority.

Dictador's humanoid Mika sits near the theatrical end. A company-supplied announcement assigns data insight, strategic provocation, and community liaison to the role. Public evidence for ordinary operating control is thin. Mika is useful because the case shows how easily a title can outrun an operating system.

An AI team becomes real through nine unglamorous mechanisms

Named roles help people understand a system, but names do not create continuity or control. The credible small-company cases converge on the same operational anatomy. They use durable instructions, company memory, a coordinator, schedules, shared work records, bounded tools, independent checks, approval gates, and a way to retire a broken agent. The guide to why AI systems get worse over time explains what happens when memory, instructions, and maintenance are left unmanaged.

SmartQix offers one of the clearest first-person descriptions. Its founder says a 19-agent hierarchy includes a COO, CTO, developers, QA, operations, finance, marketing, sales, support, and an inspector. Persistent sessions, task IDs, model routing, and a chain of command reduce duplicate work. The candid detail is verification: agents can claim work is done when it is not. QA and inspection exist to catch that failure. The human CEO approves anything customer-facing. Read SmartQix's operating account.

Jackson Yew describes twelve agents organized into executive roles and direct reports. Each role has principles, a heartbeat or daily checklist, persistent memory, shared tickets, budgets, and logs. The useful change is that work continues without Jackson relaying every piece of context. The limit is equally clear: he reviews decisions and still performs substantive delivery. His own summary is coordination compression, not team replacement. See Jackson's case and the open-source Paperclip project.

Aiprosol is younger and less proven, but unusually open about its failures. Ten role-agents run on a schedule and send customer-facing work to a Slack approval queue. The founder reports hallucinated facts during an auto-send test, tone drift, and brittle agent-to-agent chains. One pricing recommendation sounded reasonable but was wrong for the business. The human chairman remains the decision bottleneck and expects to hire people as the company grows. Read the 30-day field report.

These cases are vendor or founder accounts, not independent audits. Their value comes from inspectable mechanics and disclosed failure modes. They show what an operating team needs before they prove how well that team performs over years.

The strongest evidence appears where the work is bounded

Broad claims about AI running a company are hard to test. A bounded delivery team or service function produces clearer measures: planned staffing, elapsed time, review acceptance, response time, cost, and quality.

A 2026 Itaú Unibanco case-study preprint reports that one staff engineer worked with four specialized AI roles covering discovery, specification, core development, and non-core development. The team delivered a brownfield initiative scoped for four people in half the planned sprints. The paper reports 90 percent first-review acceptance of AI-generated code and full integration-test passage. This is one proprietary case, not a universal productivity law. The experienced engineer still validated the work, held institutional knowledge, and owned integration.

Klarna shows both the upside and the correction at function scale. In May 2024 the company said its AI assistant had engaged with more than four million customers and would produce US$40 million in annualized savings. It also said 90 percent of employees were using AI in daily work. See Klarna's results release.

Later reporting added the quality trade-off. Klarna's CEO said the company had over-indexed on cost reduction and was course-correcting. Human hiring resumed. The often repeated phrase that AI “replaced 700 jobs” compresses several facts into a misleading causal claim. Klarna compared chatbot service volume with the work of 700 staff while headcount, vendor choices, demand, and service design were changing together. Read the Reuters account of the course correction.

The lesson is not that narrow functions are the only valuable target. They are the easiest place to separate impressive output from a business outcome.

Agent count is an architecture detail, not a headcount equivalent

HOVR's roughly 250 agents, SmartQix's 19, Jackson Yew's 12, and Aiprosol's 10 are not directly comparable. One system may count a durable executive role. Another may count every specialist prompt, temporary process, or tool-specific worker. A third may keep a small set of roles that call many workflows without naming each one.

Start with responsibilities. What recurring work is completed? What state survives between runs? Which tools and data can the system reach? How often does a person revise or reject the output? What exceptions escape the normal path? How much review time moves from production to judgment? What happens to quality, cycle time, cost, risk, and revenue?

These measures also prevent agent theatre. A role called CFO that writes generic financial advice is less operational than a modest reconciliation workflow. The latter reads bounded data, produces a traceable exception list, and stops for approval before changing a record.

Nine cases, one recurring boundary

The cases differ in scale, evidence strength, and operating mode. They agree on one point: the credible system prepares, coordinates, executes, or recommends work while a named person or human body retains accountable authority.

Evidence strength also needs two dimensions. Existence can be well established while impact remains uncertain. IHC's official disclosures make Aiden's role clear, but the company reports the benefits. NetDragon's filings establish formal deployment, while the virtual CEO title says little about decision rights. Founder accounts reveal implementation mechanics while leaving financial outcomes unaudited.

A plausible org chart still does not produce executive judgment

The hardest work in leadership is rarely generating options. It is deciding under ambiguity, integrating conflicting perspectives, remembering how earlier choices changed the situation, and accepting responsibility for a path that may fail.

A 2026 preprint called CEO-Bench tested models in multi-round resource-allocation scenarios with conflicting advice from simulated CFO, CTO, COO, and CMO roles. The systems produced structurally valid plans but diverged on strategic calibration. The authors found adviser capture, conservative defaults under ambiguity, weak historical memory, and a trade-off between integrating perspectives and acting decisively.

CEO-Bench is a preprint and a simulation, not proof that deployed executive systems fail the same way. It is a useful counterweight to the idea that a coherent set of role names and well-formed recommendations establishes long-horizon judgment.

Public case selection creates another blind spot. Companies publish successes, not quiet abandonment. Founder accounts may omit cleanup. Human reviewers may repair weak work without recording the effort. Systems that look efficient at week four may accumulate stale assumptions by week fifty-two. No case in this set proves that software independently bears board, officer, employment, regulatory, contractual, or reputational accountability.

Leaders should design the boundary before the org chart

A company does not need an AI CEO to benefit from this operating model. It needs one recurring loop where coordination consumes meaningful time, the inputs are available, the outputs can be checked, and mistakes can be contained.

Begin with a functional role rather than a personality. Define the responsibility, source data, current process, output, review standard, permitted tools, prohibited actions, and named owner. Add memory only for information that should persist. Use a work queue so status and exceptions remain visible. Capture evidence at each handoff. Require approval before external, financial, legal, employment, policy, safety, or irreversible action. The three-zone decision framework is a practical test for where that boundary belongs.

Then measure the review burden. If the system produces more material but forces a leader to recheck everything from first principles, it has moved work rather than removed it. If it reliably prepares the right evidence, catches routine exceptions, and reduces the time a person spends relaying context, the management layer is beginning to function as software.

The strategic question is no longer whether an AI can wear an executive title. It is which coordination loops should become software, which decisions must remain visibly human, and what evidence will let the organization tell the difference.

If you want to inspect the underlying mechanisms, the Sidekick operating-system guide shows how roles, memory, orchestration, evidence, and approval gates fit together. The AI audit logs guide explains how to make the resulting work reconstructable.

Sources and limits

Sources were checked on July 30, 2026. Company and founder claims are attributed. Formal disclosures and research papers receive more weight than titles, testimonials, or promotional descriptions. Some pages can change after review.

• HOVR and Harrison AmitChannelNEXT report, HOVR site, PureBrain, and December 2024 offering document.
• IHC / Aiden InsightInitial non-voting board-observer disclosure, Q3 recommendations, and workflow expansion.
• NetDragon / Tang YuCompany role description, 2023 annual report, and 2024 annual report.
• Coordinated small-company teamsSmartQix, Jackson Yew, Paperclip, and Aiprosol.
• Measured compression and capability limitsItaú one-person-squad preprint, Klarna results, Reuters on Klarna's correction, and CEO-Bench preprint.
• Theatrical contrastReuters image record and Dictador-supplied announcement.
• Governance boundaryNIST AI Risk Management Framework Core.

Harrison Amit's LinkedIn wording was not used. The profile was inaccessible during the underlying research, and this review did not use an alternate route to bypass that restriction. HOVR's internal prompts, logs, permissions, override rate, and audited productivity or cost data remain unavailable.