ATL276

ATL276: The Secrets Agents Keep (Guest: Alexis Kingsbury, author "Accrual Intentions")

Accounting Technology Lab — ATL276

ATL276 for September 25, 2026: The Secrets Agents Keep

The Secrets Agents Keep

• Series: Accounting Technology Lab
• Episode: ATL276
• Title: The Secrets Agents Keep
• Guest: Alexis Kingsbury, author of Accrual Intentions
• Hosts: Randy Johnston and Brian F. Tankersley, CPA.CITP, CGMA
• Release Date: September 25, 2026
• Approximate Runtime: 57 minutes
• Primary Topics: AI agents, agentic AI, AI governance, human review, workflow design, deterministic controls, delegation, token economics, vendor lock-in, MCP, organizational knowledge, AI risk

---

Episode Summary

In ATL276, Randy Johnston and Brian Tankersley sit down with Alexis Kingsbury, author of Accrual Intentions, to unpack what he learned from building a deliberately extreme experiment: an accountancy practice staffed by eleven AI agents with their own roles, personalities, and responsibilities. Kingsbury explains that the project began as part parody and part management-science experiment, but quickly became a practical test of delegation, workflow design, controls, and human judgment.

The central lesson is not that AI is either brilliant or useless. It is that AI can perform impressively on difficult tasks and still make simple, checkable mistakes. Kingsbury argues that firms should separate probabilistic AI work from deterministic processes, document workflows, insert stage gates, and decide explicitly where human review belongs. He warns leaders not to confuse delegating the work with delegating the thinking.

The discussion also covers token economics, local models, model portability, and the risk of locking organizational knowledge inside one AI vendor's project environment. Kingsbury recommends keeping core context, processes, and organizational knowledge in systems the firm controls, then connecting AI tools to that context. His closing advice: do not wait for perfect AI, and do not attempt a giant transformation. Pick a painful, valuable problem, solve it deeply, learn, and expand.

---

Key Takeaways

1. AI can be brilliant and stupid within the same workflow. Kingsbury describes Claude completing sophisticated work correctly and then altering an API-provided link enough to break it.
2. Delegate execution carefully; do not accidentally delegate judgment. If AI takes over production, humans need to make the thinking, objectives, assumptions, and review gates more explicit.
3. Deterministic controls matter. Where possible, use AI to create repeatable calculations, scripts, tests, templates, checklists, and processes.
4. Review capacity has to scale with AI production capacity. Eleven virtual workers can generate an enormous volume of output, and that output still requires testing, prioritization, and accountable review.
5. Your organizational knowledge should not belong to your AI vendor. Portability matters when prices, models, jurisdictions, or vendor strategies change.
6. Token efficiency can become an economic issue quickly. AI consumption may require active optimization as usage scales.
7. Waiting for “perfect AI” is not a strategy. Context, processes, guardrails, and review will remain necessary.
8. Avoid the giant AI transformation project. Start with one sufficiently painful or valuable business problem, solve it well, and expand.

---

Short Promotional Copy

One-Sentence Promo

What happens when you give eleven AI agents jobs, personalities, responsibilities—and enough autonomy to expose everything that can go brilliantly right and spectacularly wrong?

Three-Sentence Promo

Alexis Kingsbury built an accountancy practice staffed by eleven AI agents and turned the experiment into Accrual Intentions. In ATL276, he joins Randy Johnston and Brian Tankersley to discuss what the experiment revealed about AI errors, delegation, controls, token economics, organizational knowledge, and human judgment. The big lesson: AI can dramatically expand what a firm can do, but only if governance and review expand with it.

Promotional Paragraph

Your AI agent just completed the sophisticated analysis, updated the documentation, called the right tools—and then broke the link it was supposed to give you.

That kind of contradiction is at the center of ATL276: The Secrets Agents Keep. Alexis Kingsbury joins Randy Johnston and Brian Tankersley to explain what he learned building the AI-staffed accounting experiment behind Accrual Intentions. The conversation moves beyond “AI good” versus “AI bad” and gets into the management problem: how do you provide context, separate thinking from doing, design deterministic controls, scale review, manage token costs, and keep your intellectual property portable instead of trapping it inside one AI provider?

---

Timestamped Pull Quotes

Time	Speaker	Pull Quote	Promotional Angle
05:03	Alexis Kingsbury	“It'll just be easier if I do it myself.”	AI vs. delegation
12:40	Alexis Kingsbury	“If you want a really good decision made, you don't want two people who think the same.”	Diversity of perspective
26:08	Alexis Kingsbury	“There's value, but also risk… maximize value and mitigate the risk.”	Governance
34:28	Alexis Kingsbury	“If you can get rid of the things that we don't enjoy… you get more time on the things that do add value.”	Human value
38:33	Alexis Kingsbury	“If you are delegating the doing, you have to pull out the thinking and do the thinking up front.”	Management
47:03	Alexis Kingsbury	“I've been able to improve token efficiency by 4,000x.”	AI economics
50:48	Alexis Kingsbury	“One other big mistake I would suggest people avoid is giving the keys away to big AI firms.”	Vendor lock-in
53:15	Alexis Kingsbury	“The technology is already good enough. It's just that you need all the guardrails and the context.”	Implementation
53:15	Alexis Kingsbury	“Pick something in your business that is either incredibly frustrating and painful… or is a current missed opportunity for value.”	Where to start

Production note: Timestamps are based on the supplied transcript/SRT. Recheck them against the final edited episode before cutting promotional video.

---

Best Short-Video Pulls

The brilliant AI that breaks the easy thing — approximately 22:55–26:40

Discussion of deterministic versus probabilistic work, culminating in Alexis's story about AI completing difficult work and then altering a valid link.

Delegate the doing, not the thinking — approximately 38:30–45:00

Alexis explains why removing the manual production step means leaders must intentionally preserve the thinking step.

Tokens, local AI, and vendor economics — approximately 47:00–52:30

Discussion of AI subsidies, token efficiency, local models, and changing provider economics.

Don't give away the keys — approximately 50:45–53:10

Governance discussion about keeping organizational context outside proprietary AI workspaces and making it accessible through tools such as MCP.

How accounting firms should start — approximately 53:15–56:30

Don't wait for perfect AI, don't start with an enterprise-wide transformation, and don't waste time on trivial experiments. Pick a valuable problem.

---

Suggested Show Notes

• Why Alexis created an accountancy practice staffed by eleven AI agents.
• Why the experiment began as both parody and serious management research.
• Custom GPTs and the early evolution of the virtual accounting team.
• Giving agents different roles, personalities, backgrounds, and perspectives.
• Why multiple viewpoints can improve AI-assisted decision-making.
• Moving from Custom GPTs to Claude Code, Claude Cowork, Codex, and agentic workflows.
• Why LLMs sometimes excel at complicated work but fail at simple tasks.
• Probabilistic AI versus deterministic calculations and processes.
• Using AI to build conventional software, scripts, templates, and controls.
• Why “I checked it” is not the same thing as having a testing methodology.
• The danger of delegating thinking along with execution.
• How AI changes management and supervisory review.
• Jevons' paradox, Parkinson's law, and what happens when automation lowers the cost of work.
• Why human judgment, context, prioritization, and care may become more valuable.
• Token economics and how dramatically AI efficiency can vary.
• Local models and private/on-device workflows.
• Why firms should retain control of organizational knowledge.
• Model portability and vendor lock-in.
• MCP as a way to connect models to context the organization controls.
• Why “wait until AI gets better” is poor implementation strategy.
• Why a giant, waterfall-style AI transformation can fail.
• Starting with an important, painful business problem rather than a trivial AI experiment.

---

20 Suggested Social Media Posts

1

AI can ace the hard part—and then break the link at the end.

That may be the most important AI governance lesson in ATL276.

Alexis Kingsbury joins us to discuss what eleven AI employees taught him about automation, errors, controls, and human judgment.

ATL276: The Secrets Agents Keep

1. AccountingTechnology #AI #AIAgents

2

“If you are delegating the doing, you have to pull out the thinking and do the thinking up front.”

That is a management problem, not a prompt-engineering problem.

ATL276 with Accrual Intentions author Alexis Kingsbury.

1. CPA #Leadership #AI

3

What happens when your firm suddenly has 11 additional workers who never sleep, type incredibly fast, and occasionally do something completely inexplicable?

Congratulations. You manage AI agents.

ATL276 explores what changes next.

4

The dangerous AI error isn't always the one you catch.

It is the mistake in a subject where you aren't expert enough to realize the answer is wrong.

That's why AI review isn't just proofreading.

ATL276: The Secrets Agents Keep

5

AI governance in one sentence:

“There's value, but also risk.”

The firms that succeed will build systems that increase the value while controlling the risk.

1. AIGovernance #AccountingTech

6

“Claude said…” is not a management memo.

AI can research it. AI can draft it. AI can format it.

But somebody still needs to decide what matters.

ATL276 gets into the increasingly important difference between delegating doing and delegating thinking.

7

If one employee suddenly became eleven employees, would you give all eleven unrestricted access, no written procedures, no review process, and vague instructions?

Probably not.

So why do it with AI agents?

1. AgenticAI #InternalControls #CPA

8

Accounting firms understand controls.

Inputs. Processes. Approvals. Reconciliations. Exceptions. Review.

Those same disciplines may be exactly what turns generative AI from an interesting demo into reliable business infrastructure.

9

AI doesn't make documented processes obsolete.

It may make them considerably more important.

Alexis Kingsbury explains why AI agents need context, roles, guardrails, and review just like human teams do.

ATL276.

10

Here's an AI strategy question that sounds suspiciously like an accounting question:

Which parts of this workflow should be probabilistic—and which should be deterministic?

That distinction comes up repeatedly in ATL276.

1. Accounting #AI #Automation

11

Your AI provider's project folder is convenient.

But what happens if pricing changes, the preferred model changes, regulations change, or you simply want to switch vendors?

Organizational knowledge needs a portability strategy.

12

One of the strongest warnings in ATL276:

Don't give the keys away to the big AI firms.

Your policies, procedures, knowledge, context, and institutional memory are business assets.

Treat them accordingly.

1. DataGovernance #VendorRisk #AI

13

ATL275 was about token economics.

ATL276 asks what happens when somebody actually starts optimizing them.

Alexis Kingsbury says he has seen enormous efficiency differences depending on how AI workflows are designed.

AI cost management is going to get interesting.

14

The most expensive AI isn't necessarily the model with the highest token price.

It may be the workflow that creates mountains of output nobody can reliably review.

Production without review capacity isn't productivity.

15

Want to implement AI?

Don't start by documenting the entire organization.

Don't start with a two-year transformation plan.

Don't start with a useless “quick win.”

Start with an important problem painful enough to justify doing the hard work.

16

“Wait until AI gets good enough” sounds sensible.

But future AI still won't magically know your clients, policies, risk tolerance, procedures, priorities, or what “good” means inside your firm.

Context is part of the system.

17

Eleven AI employees.

Names. Jobs. Personalities. Different perspectives.

And one increasingly sleep-deprived human trying to figure out what they were capable of.

There's a reason Alexis Kingsbury's experiment became a book.

18

AI may not eliminate human work.

It may move the bottleneck.

When execution gets dramatically cheaper, judgment, prioritization, review, controls, and client understanding become more important.

That's a much more interesting conversation than “Will AI replace accountants?”

19

Twenty years ago, Alexis automated six hours of Saturday work with a batch script.

He didn't run out of work.

He found higher-value things to do.

That story may tell us quite a bit about what happens next with AI.

20

ATL276 — The Secrets Agents Keep

Guest Alexis Kingsbury built the AI accountancy experiment behind Accrual Intentions.

We talk about:

• AI agents
• Human judgment
• Review and testing
• Data ownership
• Token economics
• Model portability
• Process design
• Vendor risk

If you're moving from “using ChatGPT” to actually deploying AI inside business workflows, this one deserves a listen.

---

Suggested Tags and Hashtags

Tags:

Alexis Kingsbury, Accrual Intentions, Artificial Intelligence, Agentic AI, AI Agents, Accounting Technology, AI Governance, AI Risk, AI Controls, Human in the Loop, Model Context Protocol, MCP, Claude, Claude Code, Claude Cowork, Anthropic, ChatGPT, OpenAI, Codex, Microsoft Copilot, Google Gemini, Perplexity, Hugging Face, QuickBooks, Xero, AirManual, Token Economics, Local AI, Vendor Lock-In, Process Documentation, Workflow Automation

Suggested Hashtags:

1. AccountingTechnology #AccountingTech #AI #ArtificialIntelligence #AgenticAI #AIAgents #AIGovernance #AIControls #Automation #CPA #Accounting #FutureOfWork #MCP #HumanInTheLoop #VendorRisk #DataGovernance #GenerativeAI #AccountingInnovation

---

Newsletter Copy

Suggested Subject: ATL276: The Secrets Agents Keep — When Brilliant AI Still Gets Things Wrong

Alternate Subject: 11 AI Employees Walk Into an Accounting Firm…

What happens when you stop treating AI as a chatbot and start treating it like a team?

In ATL276 of the Accounting Technology Lab, Randy Johnston and Brian Tankersley talk with Alexis Kingsbury, author of Accrual Intentions, about his experiment building an accountancy practice staffed by eleven AI agents.

What he discovered is more complicated than either the AI enthusiasts or skeptics might like.

The agents could perform surprisingly sophisticated work—and then make astonishingly simple mistakes. That led Alexis toward a management model based on documented processes, deterministic controls, explicit review points, and a clear separation between delegating the work and delegating the thinking.

The conversation also covers token economics, local models, MCP, data ownership, model portability, and why firms should be cautious about allowing their accumulated organizational knowledge to become trapped inside one AI vendor's environment.

The practical takeaway is refreshingly non-magical: don't wait for perfect AI, but don't launch a massive AI transformation either. Find a significant business problem, build the context and controls needed to solve it reliably, learn what breaks, and expand from there.

---

Suggested Visual / Promotional Image Direction

Concept 1 — The AI Agent Dossier

A sophisticated spy-thriller visual built around an accounting office. Eleven AI employee dossiers or personnel folders are spread across a desk, each showing a different fictional AI employee. A central file stamped CONFIDENTIAL contains warnings such as:

• HALLUCINATIONS
• CONTEXT
• REVIEW
• TOKEN COST
• VENDOR LOCK-IN
• HUMAN JUDGMENT

Headline:

THE SECRETS AGENTS KEEP

Subhead:

What Eleven AI Employees Taught Alexis Kingsbury About Controls, Judgment & the Future of Accounting

Concept 2 — The AI Staff Meeting

A conference room containing eleven glowing digital chairs or holographic employees and one very human manager.

Some screens show excellent completed work; another displays a hilariously broken link or obvious exception.

Headline:

YOUR NEW EMPLOYEES NEVER SLEEP. THEY STILL NEED SUPERVISION.

Concept 3 — Don't Give Away the Keys

A physical key ring labeled:

CLIENT DATA — PROCESSES — KNOWLEDGE — POLICIES — IP

Several giant AI-provider doors surround the key holder.

Headline:

WHO OWNS YOUR AI CONTEXT?

---

Products, Services, and Companies Discussed

The table concentrates on material technology and business references rather than incidental historical references. “Not confirmed” means an official current account was not sufficiently verified rather than an assumption being made.

Company / Product	Role in Episode	X	Facebook	LinkedIn	Instagram
Alexis Kingsbury / Accrual Intentions	Guest and AI-accountancy experiment/book	@alexiskingsbury	Not confirmed	Alexis Kingsbury / Accrual Intentions	Official account linked; handle not independently confirmed
AirManual	Process-documentation/onboarding software	Not confirmed	Not confirmed	AirManual	Not confirmed
OpenAI / ChatGPT / Custom GPTs / Codex	Models and agent/development tools	@OpenAI	@OpenAI	OpenAI	@openai
Anthropic / Claude / Claude Code / Claude Cowork	Core AI tools used in later stages of the experiment	Not independently confirmed	Not confirmed	Anthropic	Not confirmed
Microsoft Copilot	Frontier/productivity AI example	@Microsoft	Microsoft	Microsoft	@microsoft
Google Gemini	Frontier AI example	@Google / @GeminiApp	Google	Google	@google
Perplexity	Frontier AI service example	@perplexity_ai	Not confirmed	Perplexity	Not confirmed
Hugging Face	AI/developer ecosystem reference	@huggingface	Not confirmed	Hugging Face	Not confirmed
Intuit QuickBooks	Example accounting system/data source	@QuickBooks	QuickBooks	Intuit / QuickBooks	@quickbooks
Xero	Example accounting system/data source	Not confirmed	Not confirmed	Xero	@xero
Apple / MacBook	Local-model/on-device AI example	@Apple	Apple	Apple	@apple
Model Context Protocol (MCP)	AI/tool and context connectivity	N/A	N/A	N/A	N/A
Python	Deterministic computation/code example	N/A	N/A	N/A	N/A

---

Calls to Action

• Identify one substantial, painful workflow where AI could create measurable value.
• Document what “correct” looks like before automating the workflow.
• Separate probabilistic AI decisions from calculations and rules that can be deterministic.
• Establish human review points based on risk rather than reviewing everything equally.
• Test whether organizational context can move to another AI provider.
• Ask where prompts, documents, policies, and project knowledge actually reside.
• Measure AI cost by workflow and successful outcome—not simply by monthly subscription.
• Make employees part of implementation so AI is something they help design.

---

Production and Fact-Checking Notes

1. Book title transcription error. The transcript repeatedly renders Accrual Intentions incorrectly. Use Accrual Intentions: What Happened When I Built the World's First 100% AI Accountancy Practice.
2. “ChatGPT six with Astra.” The wording around approximately 19:39 appears to be a transcription or verbal-name issue. Verify against the source audio before publication.
3. Claims about AI IQ scores. Treat the discussion of AI systems operating around IQ 190 as commentary rather than an independently established measure of general AI intelligence.
4. “95% of AI pilots fail.” Identify and review the underlying research before using this number as a general statistic.
5. OpenAI versus Anthropic developer spending / Hugging Face reference. These are rapidly changing market claims and should be separately sourced if included in published show notes.
6. 4,000x token-efficiency improvement. This is Alexis's description of his own experiment and should be attributed to him rather than presented as a general benchmark.
7. AI pricing economics. Discussion of subsidized AI usage and future pricing is analysis and opinion, not a certain future outcome.