Artificial intelligence is having another one of those moments.
There are warnings about autonomous systems. Questions about how fast the technology is moving. Arguments over guardrails, regulation and whether AI is becoming too capable too quickly.
Some of those concerns deserve to be taken seriously.
But I also think we are making a mistake by lumping every form of AI into the same conversation.
A frontier AI system operating with broad autonomy and an accounting firm using AI to prepare for tomorrow morning’s client meeting may both be called “AI agents.”
They are not remotely the same thing.
I’ve spent a lot of time thinking about this because we are not watching the agent debate from the sidelines. We are actually building these systems for tax and accounting firms.
And the more we build, the more convinced I become that the most important question is not:
Should AI be autonomous?
It is:
What should AI be allowed to do autonomously, and where should the professional take over?
That is where guardrails begin.
Start with Monday morning
Imagine you have 300 clients.
It is 7:45 on a Monday morning. Over the weekend, documents were uploaded, messages came in, a prospect responded to a proposal, someone partially completed an intake form, a tax notice arrived and several tasks moved closer to overdue.
You also have four client meetings today.
In most firms, the morning starts with reconstructing context.
You open messages. Check the calendar. Search for documents. Look at open tasks. Try to remember the last conversation and whether someone followed up.
Before the professional gets to the part of the job where their expertise really matters, they may spend 20 or 30 minutes simply figuring out what is going on.
Now imagine starting differently.
Before your first meeting, AI has already assembled the relevant context.
It notices that your 10 a.m. client uploaded a document Friday. It reads the document and finds an explicit response date. It connects that document with an open task and a recent client message. It prepares the questions you may want to ask.
Nothing has been filed.
No tax position has been chosen.
No client has received professional advice from a machine.
You simply walked into the meeting dramatically better prepared.
That is the part of the AI agent conversation I think gets missed.
AI agents are not all the same
When people hear the words autonomous AI agent, I think they often picture software being turned loose with unlimited access to their firm.
Give it a goal. Walk away. Hope it behaves.
I wouldn’t be comfortable with that either.
But that is not how I believe AI should be deployed inside a judgment-based profession.
An agent is simply an AI system that can do more than answer a question. It can work toward an objective by gathering information, using approved tools and taking a series of steps.
The important questions are much more practical:
What can it see?
What tools can it use?
What can it change?
What requires approval?
Where does the machine stop and professional judgment begin?
That is what guardrails actually mean.
What does that look like in the real world?
Take something as common as a tax notice.
AI can locate the document.
It can read it.
It can identify the agency, tax period, amounts and explicit deadlines contained in the notice.
It can explain the document in plain English.
It can extract the requested actions.
It can prepare tasks.
It can draft a message explaining what information the firm may need from the client.
But there should be limits.
The AI should not decide on its own that the tax agency is wrong.
It should not invent a statutory deadline that is not actually supported by the information available.
It should not choose the tax strategy.
And depending on the action, it should not create tasks, change records or communicate with the client without the appropriate approval.
That is human in the loop.
It isn’t putting a human somewhere vaguely near the technology. It is deciding intentionally which parts of the workflow belong to the machine and which parts belong to the professional.
The difference is bounded autonomy
This isn’t theoretical for us. We have had to answer these questions in the way we are building MAX Intelligence.
Today, MAX has a library of 77 specialized skills.
We did not build one giant agent and tell it:
“Go run the accounting firm.”
We built skills with specific jobs.
One can prepare you for an appointment. Another can give you a complete operational view of a client. Another can triage messages. Another can read a tax notice. Another can identify workflow bottlenecks or surface deadlines contained in documents and records.
The number itself is not really the important part.
What matters is that those skills can work together without each one having unlimited authority.
A document-reading skill does not suddenly get to decide tax strategy.
A messaging skill does not get to communicate professional advice simply because it can draft a response.
Some skills are read-only.
Some can recommend actions.
Some can prepare an action but require the professional to approve it before anything happens.
That is the difference between autonomy and unbounded autonomy.
AI changes where the professional begins
For years, the technology pitch to tax and accounting professionals has been about efficiency.
Save ten minutes here.
Automate data entry there.
Produce a return a little faster.
Those improvements matter.
But I think AI creates a much bigger opportunity.
It can handle an amount of preparation that humans simply cannot.
That is not because humans are incapable.
Humans have limits.
If you have 300 clients, you cannot wake up every morning having reread every email, every uploaded document, every previous conversation, every open task, every intake response and every recent change in every client’s situation.
A machine can continuously work through that information.
So instead of the professional beginning with:
“What is going on with this client?”
the professional can begin with:
“Given everything we know, what should we do?”
That is a very different starting point.
And I think it is a much better use of the professional’s time.
Human in the loop has to mean something
“Human in the loop” has become one of those phrases everybody in AI uses.
But it only matters if you define where the human actually enters the process.
AI can find it.
AI can read it.
AI can connect it.
AI can organize it.
AI can prepare it.
AI can recommend what should happen next.
And when the decision requires professional judgment, the professional makes it.
That is how we have approached our own notice and document-related skills. They can extract facts, identify supported deadlines, explain documents and prepare next steps, while deliberately avoiding unsupported tax conclusions or decisions that belong to the professional.
This is also where I think properly designed AI can actually make a firm more disciplined.
Things fall through the cracks today not because people do not care.
They fall through because there are too many clients, too many emails, too many documents, too many deadlines and too many open loops for any human being to continuously monitor all of them.
AI can keep looking.
The professional can keep judging.
Those are very different jobs.
This is different from a blank chat window
I use tools like ChatGPT constantly.
They can be incredibly useful.
But opening a general-purpose AI chatbot and asking it a question is different from operating inside an AI environment designed around a professional workflow.
Inside a controlled environment, you can define what information the AI can access.
You can define the tools it is allowed to use.
You can limit what it can change.
You can determine which actions require approval.
You can establish rules around sources, confidence and professional review.
The goal should not be to give AI unlimited freedom simply because the technology is capable of doing more.
The goal is to give it enough autonomy to remove unnecessary human work without giving away the judgment that makes the professional valuable.
The skills should work together
A good accounting firm does not have one person who does everything.
People have roles.
Someone gathers information.
Someone prepares the work.
Someone reviews it.
A professional makes the judgment.
Someone communicates with the client.
Someone makes sure the follow-up actually happens.
AI can work much the same way.
One skill gathers client context. Another understands the document. Another identifies action items. Another checks the work queue. Another prepares the client communication. Another looks across the firm for unresolved work that may be falling through the cracks.
That is where I think agents become much more interesting.
Our newer firm-level skills are already moving in this direction, including a Deadline Risk Radar, a Tax Season Command Center and a “Nothing Falls Through the Cracks” briefing that looks across tasks, messages, documents, proposals, payments, appointments and workflows.
None of that requires the AI to become the accountant.
It requires AI to do what computers are extraordinarily good at doing: processing enormous amounts of information, connecting the pieces and bringing the right things to the right person’s attention.
The real opportunity is not replacing judgment
This is where I think the conversation becomes much more important for firm owners.
The competitive advantage is not simply that AI saves time.
It is that a firm can know more, prepare faster and show up for clients with far more context than was ever practical before.
Think about the client experience.
The old model is often:
“Call me if something changes.”
The better model is:
“We already looked at what changed. Here are the two things I think we should talk about.”
That is a completely different experience.
And it does not require AI to replace the professional.
It requires AI to make the professional better prepared.
There are legitimate questions about increasingly powerful autonomous systems and what safeguards they should have.
Those debates should continue.
But tax and accounting professionals should not mistake that debate for the decision sitting in front of their own firms today.
We do not have to choose between ignoring AI and handing our firms over to machines.
There is a very large middle ground.
We can restrict access.
We can limit tools.
We can require approval.
We can preserve professional judgment.
And at the same time, we can use AI to perform preparation at a scale that no human team could realistically duplicate.
That is where I believe the real opportunity is.
Because the goal is not autonomous judgment.
The goal is autonomous preparation, with a better-prepared human making the decisions that matter.
