In April 2025, I wrote a post questioning how reliable AI agents could get.
My take at the time? 2025 wasn't the year of AI Agents, and I doubted 2026 would be either.
I was wrong.
As I write this in early 2026, and despite running Opzer, an AI automation consultancy, I underestimated how quickly AI would improve.
Here’s what changed my mind, how I define AI agents (in plain terms), and why now is the time to start using them.
The personal AI agent that changed my mind
I’ve experimented with browser use agents—AI tools that spin up a browser session to automate tasks—since ChatGPT’s native browser agent came out in early 2025.
It sucked.
Human-like browsing has been one of the most challenging tasks for AI to get right.
Last month, however, I asked an AI agent to log into my bank and report on my personal finances.
Despite having multiple currencies, three bank accounts, two investment accounts, a few credit cards, it logged into my bank portal, navigated menus, and downloaded six months of statements for all my financial instruments.
Once it had the data, it categorized six months of statements and built a beautiful dashboard in minutes.
I was shocked.
The final output of my personal finance agent, executed by browser use running Claude Opus 4.6 as its model.
Total time it would have taken me: realistically, never. I would have put this off forever.
Instead, I got a full picture of where my (personal) money is going for the first time in a long time.
The agent processed the files, made its own categories, categorized the transactions, and helped me visualize my expenses. Agents are getting us to actually do work that'd take an entire day (or days) in under an hour.
So now I'm a believer.
But what makes an AI agent different from traditional software, automation, or your regular AI chatbot?
What an AI agent actually is
An agent is a software system that acts, thinks, and adapts.
Embedded in that definition are 3 key features. The 3 As, as I like to call them: Action, Autonomy, Adaptation.
Let’s define each one with some examples.
The three essential features of an AI Agent. Anything less is traditional automation.
Hands (Action)
An AI agent does things outside the chat window, usually through connectors and integrations with other apps.
Examples might be:
Creating a vendor in QuickBooks
Posting a journal entry in Xero
Updating a work item status in Karbon
Downloading a bank statement
Rather than having to copy+paste the result or follow instructions, it just does the task for you.
Brain (Autonomy)
An AI agent makes its own decisions, usually by picking which tools to use and when, without being told.
Examples might be:
Deciding whether to use the QuickBooks connector, the API, or a browser to update a vendor
Choosing between searching contacts or organizations to find a client in Karbon
Breaking down “reconcile this account” into 10 sub-steps without being told
You don’t need to spell out every step. You give it a goal, and it figures out what to do to get there.
Loop (Adaptation)
An AI agent learns mid-task through memory. It has the ability to fail and try again without being told.
Examples might be:
Hitting a login error and trying a different page
Failing to get all the data it needed and retrying with a different approach
Remembering from a previous run that a particular client codes meals as ‘T&E’ instead of ‘Meals’
All three boxes have to be checked for me to call it an agent.
What about AI agents has gotten better recently?
A year ago, agents looked great in demos but fell apart in production. Four things changed in the last 12 months.
Better reasoning, more native connectors, reliable (but still needing supervision) browser use, and stronger security are making AI agents usable.
Reasoning feels more human than ever
The latest generation of models can take a fuzzy goal like “reconcile this account” and break it into the right sequence of steps without being hand-held.
Just as importantly, they catch their own mistakes mid-task instead of confidently barreling forward into a wrong answer.
More, better connectors
Native connectors—and in our experience, Claude’s connectors—have started performing very well.
More apps and software companies are shipping their own, which makes it dramatically easier to plug agents into the apps accountants actually use: QuickBooks, Xero, Karbon, Google Drive, Slack.
Browser use started actually working
A year ago, asking an agent to log into a portal and pull a document was too slow and clunky.
Humans are still faster, but agents can now navigate portals and weird legacy UIs reliably enough that the technical capability is no longer the blocker.
Unique risks remain—like prompt injection, where a rogue page can redirect an agent to click, type, or submit things you didn't intend—so human oversight is still highly advised. But this unlocks a set of use cases that connectors alone would never fulfill.
My personal finance agent above is an example.
Better guardrails
Security risks, although outside the scope of this article, are also being addressed. Human oversight is still warranted, but — for example:
Claude Cowork (the most user-friendly agent harness from Anthropic) made it out of research preview, which means it now ships with the kind of controls enterprise teams expect: role-based access, audit logs, group spend limits, and a sandboxed execution environment. Agents can only touch the folders, files, and apps you explicitly hand them.
OpenAI launched AgentKit with built-in guardrails. Released as an open-source, modular safety layer inside Agent Builder, it can mask Personal Identifiable Information (PII) and wrap tool calls with custom validation rules.
The theme is that every major model provider is now building guardrails to win enterprise deals, which also benefits your firm.
The combination is what matters. Better reasoning plus more connectors plus working browsers plus better guardrails is what’s turning agents from ‘cool demo’ into ‘I can put this to use.’
Now is the time
A year ago, I would have told you to wait. The tools weren't secure enough, the failure rate was too high, and the time investment to get something working wasn't worth it.
That's no longer true.
Agents are handling real work—complex, multi-step, judgment-required work. And if you’re ever unsure where to start, ask AI itself. Describe your workflow, explain the bottleneck, and ask what you'd automate first.
Try it. See what breaks. Iterate. It may not work flawlessly the first time, but today is the worst it’ll ever be.
As a process optimization expert, Isaac Perdomo helps accounting firms do more with their existing team and software. With a focus on reporting, automation, process optimization, they’ve helped 80+ firms streamline their operations. When he's not automating business processes, he's tinkering with new software apps or taking a long walk outdoors.
This is your crystal ball for the next decade in accounting, based on 14 predictions from 34 of the profession’s most forward-thinking leaders and innovators.
G2.com
:format(avif))
:format(jpeg))
:format(avif))
:format(avif))
:format(jpeg))
:format(avif))
:format(avif))
:format(avif))
:format(avif))
:format(avif))