When I think about where AI can change professional work, I start by considering what happens outside the software today. This is not because the software is failing to do its job; it’s because the current software in place today was built around an assumed boundary of capability. That line is moving, fast.
A tax application can perform a calculation once the inputs and treatment are established. Getting a particular client’s records into that state is a different problem. Someone reconciles inconsistent files, investigates an exception, determines which rule applies, and builds the workpaper connecting those decisions to the result.
As a software builder, what interests me about agents that can write and run code is their ability to move that boundary: to take on more of the client-specific work between a standardized application and a real engagement. This is not about building yet another piece of software to do the existing work in a faster way. This is about software that does more of the work itself with human judgement applied where it’s needed today.
The economics of the exception
Traditional software earns its investment by solving repeatable problems. Developers anticipate a workflow, encode its rules, and make it useful across many customers. That remains enormously valuable. It also leaves a long tail of tasks whose details vary too much to justify a dedicated feature for each one.
Professionals have been the bridge for that gap. In tax, accounting, and audit, professionals have bridged that gap with spreadsheets because a spreadsheet is a useful way to express a calculation, expose assumptions, and adapt to facts that do not fit a standard screen.
Consider a client asset schedule assembled from several inconsistent files. The tax package may already handle depreciation perfectly well. The work around it includes reconciling records, identifying missing information, establishing classifications, and documenting why particular treatments were selected.
An agent that can generate and execute code offers a different way to assemble that work. It can help reconcile the files, run calculations under specified assumptions, and produce a spreadsheet for review. A revised assumption can lead to a revised calculation without rebuilding the workpaper by hand.
The change is not just that automation becomes cheaper. It is that software can take on a different kind of problem. Traditionally, a developer had to understand the task well enough to specify how the software should handle it. An agent can work from the task itself: interpret unfamiliar inputs, reason about what needs to happen, and write and run the code to carry it out. That moves the boundary beyond the workflows we could anticipate and encode in advance.
Running correctly is not being correct
There’s an important nuance though. Being able to execute the work is not the same as getting it right. A calculation can run flawlessly yet still be based on the wrong interpretation of the client’s facts or the rule that applies. A script can apply the wrong classification but still look impressive. It can use a rule from the wrong tax year or silently turn missing information into an assumption.
In the asset-schedule example, reconciling totals is useful evidence. It does not establish that the treatment of every asset is appropriate. The hard question is whether the chain from client facts to applicable authority to calculation holds together.
This is why I do not see professional AI as a contest to put the most capable model inside a familiar interface. Model capability is essential. But the system also has to interpret the client’s facts, apply the right authority, and carry that reasoning through to the work it produces. And most importantly, it has to make those connections visible, so the professional can steer the system and validate the work without having to do it all over again. That is the engineering problem we are solving.
Where domain expertise becomes engineering
That distinction shapes the work behind today’s launch of the next generation of CoCounsel Tax, Audit & Accounting. It connects research and reasoning with agents that write and run code for calculations and work across spreadsheets and documents. The aim is to carry more of the task through to work product, rather than leave the professional to reconstruct pieces of the answer across a disparate set of tools.
Thomson Reuters brings something important to that engineering problem. Checkpoint’s maintained tax content provides a basis for identifying relevant authority and interpretation. A well-written explanation of the wrong year’s rule is still the wrong basis for the work. The system needs the authority applicable to the client’s circumstances and the period being examined, not simply the latest document it can retrieve.
Content alone is not enough. Tax practitioners also help develop and test the system. Their expertise matters most when it becomes a test the technology has to pass, not simply advice supplied during product design.
The resulting work also has to preserve the connection between sources, assumptions, calculations, and conclusions. Otherwise, the professional saves time producing an artifact and spends it again reconstructing how it was made.
That is why I believe Thomson Reuters is well positioned to lead in this category. Access to frontier models is increasingly shared. The maintained body of professional knowledge, the expertise to test its application, and the connection to the workflows where that work is used take much longer to build.
Those assets are not a substitute for capable technology. They are how capable technology becomes useful in a profession. Our job is to connect them to execution and review, not simply put them beside a model and call it trusted.
Test the work, not the conversation
For a practice evaluating AI, I would start with a real engagement where the standard workflow ends and the workarounds begin. Ask the system to handle that gap. Then change a consequential fact. Can you see what changed in the work, why it changed, and what still needs your judgment?
That is a more revealing test than the fluency of an answer. It measures whether the technology reduces the cost of assembling client-specific work without hiding the decisions inside it.
The goal is not to remove professional judgment. It is to stop making professionals rebuild the machinery every time they need to exercise it.
Read more about the launch of the next generation of CoCounsel Tax, Audit & Accounting in the blog post by Chief Product Officer, David Wong.
