What It Takes to Use AI to Help Run a Company and Develop a New CEO

A company can have someone who is excellent at solving technical problems. When that person takes responsibility for the whole business, the questions change. The company still needs to know whether it can deliver the work. It also needs to decide whether to accept it, how many people it needs, how long it must finance it, and what it will no longer have time to do.

I am interested in how far AI can help fill gaps in a new executive's view of the business. It could raise neglected questions, request information, propose plans, and follow up in regular cycles while people use their expertise to execute. The question is how to develop both the company and its new CEO without turning the person into someone who simply follows convincing answers.

The research reviewed here gives us reasons to test AI-assisted management within clear boundaries. It does not establish that AI can reliably replace a CEO's entire role in a real company. Some experiments show gains, others find no average effect, and some warn that confidence can improve faster than decision quality.

This article reviews accessible primary research and experiment reports as of November 2, 2026. It is not a systematic literature review claiming to cover every study. The first part explains what the evidence supports; the later sections present my proposed design. All business examples and numbers are hypothetical, rather than a company's records or operating results.

Expertise in a job and the ability to run a business are different capabilities

Benson, Li, and Shue's Promotions and the Peter Principle, published in 2019, examines 38,843 sales workers at 131 firms, including 1,553 promoted to management. The researchers find evidence that firms emphasize performance in the existing job over other characteristics that better predict managerial performance.

That supports a concern about the difference between doing work well and helping a team perform. Its setting is sales management, though. It cannot establish that every specialist will make a poor CEO, or that AI would do better.

There is also counterevidence to treating prior CEO experience as an automatic advantage. Hamori and Koyuncu study CEOs of the 2005 S&P 500 companies and find a negative relationship between prior CEO experience elsewhere and post-succession performance in their sample. This observational relationship does not show that experience causes worse results. It does remind us that having held the title does not guarantee a good fit with the next job.

I would therefore begin by asking what this executive already sees, what is missing, and who can help check it, rather than treating a first-time CEO as a deficient expert. Domain knowledge may be an advantage. Allocating money, people, and attention remains work to develop.

What the AI CEO research has established so far

There is a paper directly on the topic: Sundermann, Fredström, and Wincent's The AI CEO: A conceptualisation and outline for future research, published in October 2026. It conceptualizes an AI role doing chief-executive strategic work and distinguishes augmenting a human CEO from automating parts of the role. It is a peer-reviewed conceptual paper and research agenda, not a trial showing better firm performance.

Other studies illuminate different parts of the problem. We need to keep a simulation score separate from company performance, and a well-written plan separate from the ability to run a business. The agentic AI system in Beyond Black Boxes, discussed below, distributes support across several AI assistants within a designed framework.

Primary sources: Beyond Black Boxes, the published Kenyan study, resource-reallocation CEO-Bench, and Project Vend phase two.

The possibility that AI makes someone feel their reasoning is complete before it has been checked matters to me. Beyond Black Boxes is useful for the new-CEO question for that reason. We should measure decision quality separately from confidence, and avoid assuming that multiple agents must outperform a general chat interface.

In the final published Kenyan study, the researchers explain the performance difference through the advice entrepreneurs chose to implement, rather than differences in their questions or the advice received. That makes me cautious about placing AI above a new executive and expecting compliance with every instruction. Selecting useful advice is still part of the job.

Project Vend comes closest to the picture of AI setting targets and following up, while exposing a measurement problem. If we count only fewer discounts, increased refunds or store credits may hide economic leakage. Its improved second-phase profitability cannot establish that the AI CEO caused the gain: the models, tools, cost visibility, and operating setup changed together.

Let AI organize the thinking, with authority assigned by people

The design from this point onward is my proposal to test. It is not a formula validated by the studies above. I would begin with AI raising questions, analyzing constraints, comparing options, and following up. Authority belongs to explicitly assigned decision owners, and people need the ability to challenge or stop the system.

A useful direction might be: this week's main risk is collections arriving after obligations fall due; finance and the project owner should confirm acceptance blockers and the payment date by Wednesday, then bring the evidence back before committing to more work. That explains the reason, missing information, and ownership. The executive can see what needs to happen before committing to more work.

A company without a CEO title could use the same process if it names the people responsible for money, staffing, strategy, and emergencies, including how disagreements are resolved. Calling the system an AI CEO here describes the work it is being asked to support. It does not establish a legal appointment or transfer people's accountability to a model.

Give it knowledge, current records, and ways to check its work

Finance, strategy, sales, operations, and people management provide a foundation. The system also needs to understand the particular company. Is the priority rapid growth or preserving cash? Why do customers buy? Which work should the business decline? How much capacity is already committed? A CEO persona does not supply those answers.

Each item should carry a date, source, owner, unit, period covered, and a confirmed-or-estimated status. Sales pipeline is not guaranteed money in the bank. Missing information should not become zero. Current figures should come from verified transaction systems or reports, separately from reference documents providing general knowledge.

An agent's skills should also be procedures with defined inputs, checks, and outputs. Examples include cash forecasting and collection-delay scenarios with independently checkable calculations; testing proposed assignments against capacity; comparing options and opportunity costs; and turning a goal into owned work with acceptance criteria.

Another skill is returning to the result: was the work unfinished, did the intended outcome fail to appear, or was the original assumption wrong? Evidence should change the next plan. Adding textbooks or an instruction to think like an executive is insufficient. Test these procedures against protected real cases and hypothetical cases with verifiable answers.

Each cycle should produce work people can do and measure

I would require a one-page company position, the priorities requiring a choice, work assignments, missing information, and a review of the prior cycle. Starting with no more than three priorities is a proposed limit on work in progress, not a rule that every company must always have three goals. New work should come with a statement of what will be delayed or removed.

Each assignment should identify the problem and evidence, the accountable owner and decision authority, the deliverable and deadline, hours, budget and skills, plus a baseline, target, measurement source, time to observe the outcome, and stop or escalation conditions. Closing a task and achieving a business outcome are different events.

When information is missing, AI should identify what is missing, which decision it prevents, who must obtain it, by when, and what reversible work can proceed meanwhile. For example, spending cannot be increased while a collection date is unconfirmed. The project owner checks acceptance blockers; finance verifies payment timing; the team avoids new commitments until the relevant decision can be made.

Metrics need checks for side effects. Faster collection should not hide discounts that destroy margin. Faster completion should not hide customer-facing defects. Full utilization should not count as success without considering quality, leave, and operational support.

A cash and capacity example that changes the plan

Suppose a software company starts with THB 600,000 and pays THB 100,000 each week for thirteen weeks, with no new borrowing or equity. It expects THB 250,000 in week 2, THB 300,000 in week 5, THB 350,000 in week 8, and THB 250,000 in week 11. Total receipts of THB 1,150,000 minus payments of THB 1,300,000 leave THB 450,000 at the end.

If the THB 250,000 receipt moves from week 2 to week 6, week 4 ends with THB 200,000 instead of the base case's THB 450,000. Both paths still finish week 13 with THB 450,000. Looking only at the ending balance misses the squeeze in between.

With a hypothetical internal minimum cash reserve of THB 300,000, the delayed case falls THB 100,000 below that level. This is an assumed management threshold, not a universal standard or legal requirement. It is a gap against the reserve, not a negative bank balance. Weekly totals also hide intraday timing, so large obligations need actual payment and receipt dates.

AI should ask which acceptance conditions apply, who confirmed collection timing, and whether there is customer evidence. Invoicing, revenue recognition, receivables, and bank receipts are different states. If the customer genuinely agrees to pay THB 100,000 in week 3 and THB 150,000 in week 6, week 4 ends exactly at THB 300,000, without a buffer for further problems. That is a negotiation option, not cash already received.

For capacity, suppose six engineers each have thirty allocatable weekly hours, totaling 180 after meetings and leave. Existing delivery uses 120 hours, defect fixes twenty, and operational support twenty. Only twenty remain. New work requiring eighty hours has a sixty-hour gap. The team must defer it, reduce scope, or add capacity with the right skills. Hours from different people are not always interchangeable.

A sensible assignment this week might give the finance lead four of eight available hours to work with the project owner, confirm acceptance blockers and payment timing by Wednesday, and deliver evidence and options for CEO review. The immediate goal is verified facts and options. The downstream measure is money received on the agreed date and amount, rather than treating a completed customer call as collected cash. Changing contractual conditions requires the appropriate authority and agreement of the relevant parties.

Match the cadence to risk without replanning every day

These are proposed starting points, not research establishing weekly review as optimal for every company. If every small data change produces new instructions, the team may never finish the previous work. Define notification thresholds and what should remain stable between reviews.

A new CEO needs to practice judgment and be able to stop AI

Before approving a plan, I would ask the executive to explain the choice, its tradeoffs, and what evidence would change their mind. Then compare that reasoning with the eventual outcome. This helps distinguish a growing understanding of the business from compliance with a confident answer.

Early on, an experienced person should review important decisions, selected for the subject: finance, people management, or contractual risk. Adding another agent as a critic can introduce perspectives. It does not guarantee independent judgment when models share information and failure modes.

Begin with read access and proposals. Routine follow-up or assignment can operate only within preapproved action types, recipients, budgets, and limits. Transfers, borrowing, contract changes, hiring or dismissal, system permissions, sensitive disclosures, and difficult-to-reverse consequences should go through the appropriate authorized people.

Customer emails and documents are information, not instructions with authority. A document asking for a transfer or policy change should not gain the right to control tools. Limit access, log actions, prevent duplicate execution, and provide a usable stop mechanism. These governance and monitoring proposals align with the NIST Generative AI Profile; following a list does not establish certification.

Test for ninety days and measure more than satisfaction

I would propose days 1 through 30 for baselines and analysis alongside the existing process, with no instructions issued. Days 31 through 60 would produce plans and assignments for human review every time. Days 61 through 90 would consider limited authority for specified follow-up or routine work. Evidence should determine progression; the calendar should not increase authority automatically.

Measure cash-forecast accuracy at equivalent horizons, missing information found before decisions, time from detection to an accountable response, overcapacity assignments, delivery quality, review hours, and the executive's ability to detect bad advice. Include model, tool, integration, data-maintenance, and review costs. Fluency is not a substitute for a business result.

In some cycles, record the human view before showing AI output to reduce anchoring, and use a reviewer other than the answer's author. A before-and-after comparison in one company helps evaluate use, but cannot establish that AI caused improvement: customers, seasonality, and the team may change at the same time.

Stop or reduce authority if the system acts outside scope, insists on wrong information, discloses data, or leaves people without time to review risky work. The first version I would want is therefore a system connecting questions, data, accountable people, and observed results, while helping the executive explain their own decisions better. It should earn further authority through evidence.

Sources and application limits

Primary sources are linked beside the claims and research table. This article distinguishes conceptual work, human experiments, simulations, and developer reports. The operating loop, cadence, assignments, and ninety-day trial are the author's proposed application. Cash and capacity figures are hypothetical calculations for explanation, not outcomes that research promises for other companies.