Refactoring the Office of the CFO

PriceWaterhouseCoopers and Anthropic have announced a partnership which effectively significantly disrupts how the office of the CFO will work going forward, Anthropic and PwC Expand Alliance, Driving Impact Across Client Work and the Firm.

The partnership accelerates a shift from “accountants doing tasks” to “accountants supervising AI agents that perform tasks inside governed workflows.”   This is already happening inside PWC’s own finance operations and in client deployments.

Let the refactoring begin!

Per the article, PwC and Claude are selling governance, not just agent speed, the value of the PWC and Claude combination is auditability, risk controls, and regulated workflow design; and not simply faster agent output.

According to Digital Applied; 3 of the Big 4 now run on Anthropic' s Claude: PWC, Deloitte, and KPMG.  EY is using Microsoft.

Anthropic’s foundation models and agentic tools are fully proprietary.  These include Claude, Claude Code, and Claude Cowork.

PWC’s deployment is proprietary at the model and workflow layer, but standards-based at the integration and governance layer it appears. PWC is building Claude-native workflows that are not open standards, they are PWC’s internal IP and client-facing delivery assets. Model Context Protocol (MCP) which is a standard is being used for integration between LLM applications and external data sources and tools. This approach lets PWC build defensible IP while still satisfying audit, compliance, and regulated‑industry requirements.

MCP, which was initially created by Anthropic was open sourced in 2024.  MCP is often described as the "USB-C of the AI world", providing a universal, standardized connection between AI systems and external tools, making AI applications more capable, interoperable, and scalable.

It seems to me that the power of MCP as a USB-C connector is important but not explained well.  MCP is not capable of magic. I say this because there are logical and physical problems that MCP simply cannot resolve.  First you have the problem of the document (as contrast to graph first knowledge) that I explain in Mind the Gap. Second, you still have the problem of semantic fragmentation that I go over in Fragmentation and Defensible Compliance. Third, you still have the whole problem related to Meaning.  If you don't understand this, please read Digital Information Organism and Professional Oriented Knowledge Frameworks.  And then, of course, you have the problem of "messy data" in general which is clearly explained by The AI Ladder.  And finally, you have the problem of the electronic spreadsheet where PWC admits the best they have ever been able to do is get up to somewhere between 74% and 87% accuracy; however, that was with OpenAI.  The work around is humans fixing problems.  You would be hard pressed to create what you would call a scalable industrial process on top of this.  These issues will need to be resolved for MCP to work effectively.

Some things worth noting about PWC's and Anthropic's announcement.  Per this LinkedIn post and per this analysis, the business model created is effectively a "land-and-expand" consulting play in the banking, insurance, and healthcare sectors. The AI product is the "door" and the transformation engagement of the business.

PWC has set up a center of excellence and will train and certify 30,000 US based consultants initially on Claude and eventually on Claude Code and Claude Cowork.

Deloitte, KPMG, and EY will very likely take similar approaches. This article, PwC Claude CFO vs Deloitte AI vs McKinsey QuantumBlack, compares and contrasts the offerings from PWC,  Deloitte, and McKinsey.

What I don't understand is the true capabilities being offered here.  As I have pointed out in this blog post, Human Task Performance, there are different "flavors" of artificial intelligence. Not sure if what is being offered is the full spectrum of flavors or only one flavor.

Regardless, changing the paradigm of how the office of the CFO operates will take years. What PWC, Deloitte, KPMG, EY, and McKinsey is still "pre-science"; the "new normal" has not quite emerged but it is beginning to take shape.  Also, while the Big 4 are focused on the large enterprise, what is  going on is going to affect every enterprise; large, medium, small; public; private; not-for-profit; government.

The full transformation is not just "building out the enterprise knowledge graph" so that artificial intelligence can be leveraged.  I predict that enterprises will shift from a document-based architecture to a model-driven architecture. This will take time, but it is inevitable because you need the model-driven architecture to build industrial processes.  I predict that accounting will be refactored to begin at the beginning with the business event as contrast to starting later in the chain.

The office of the CFO is only the beginning of the refactoring of the enterprise enable to take full advantage of the opportunity artificial intelligence brings to the table.

I cannot tell for sure, but it seems like what PWC and Anthropic have is perhaps not full neuro‑symbolic artificial intelligence with human teaming.  But it does seem to be a hybrid system where you have neural models constructed by the machine, symbolic workflows constructed by humans, governed connectors to information sources, enterprise schemas defined by MCP, and audit trails.  All these seem to work together in a way that is perhaps functionally neuro‑symbolic, even if not formally integrated. It is, perhaps, a practical version of neuro‑symbolic artificial intelligence.

BOTTOM LINE: Buyer beware.  Artificial intelligence absolutely is a real thing and it will absolutely enable a refactoring of the office of the CFO and the entire enterprise.  Right now, I don't know exactly were we are on the Gartner Hype Cycle, but I would speculate it would be at the "Peak of Inflated Expectations".  We still have the "Trough of Disillusionment" to go.  As The AI Ladder points out, 81% of business leaders don't understand the data and infrastructure required for AI.  Watch out for the snake oil salesmen trying to separate you from your hard earned money.  The "Plateau of Productivity" will be reached...but we are not there currently.


All that said, the office of the CFO and the enterprise will be refactored. Methodologies, best practices, frameworks, structures, and processes that are leverageable are key.  I began my path to the Seattle Method in about 2001. If you want to have a look at what I created, here is My Garden.  I am personally embracing the Accounting & Audit by Design (A&AD) Framework.  More to come.

* * *

The following is a list of commonly used "synthetic reasoning engines" (i.e. LLMs and transformer; I think the right word is) which focus on natural language understanding and generation:
The fundamental capability of  all these is that they excel at "semantic mapping" meaning that they convert linear, natural language into a dense web of conceptual meanings, allowing them to reason across the relationships between those concepts in context rather than just the words themselves. These tools do not possess a mental model of the physical world, memory, an internal monologue, or a true moral or factual compass. They possess a world-class, multi-dimensional map of human language. They are phenomenal at restructuring, synthesizing, and translating that language, but they cannot be blindly trusted to validate the truth of their output.

To adjust for these issues, things are being added to these tools.  For example, Claude includes the following capabilities which have been added:
  • Skills: Skills are  modular capabilities which extend functionality.  These seem to be equivalent to my notion of a "Theory".  
  • MCP Connector: An MCP connector is a global standards based connection to a resource that provides data like a database, an application, a Github repository.  These seem to be equivalent to my notion of an XBRL instance, an XBRL taxonomy schema, or an XBRL linkbase. 
  • Projects: A project is background information tied to a permanent workspace, it is like memory.  This seems to be equivalent to a reporting framework.  
  • Artifacts: Artifacts are outputs.  This seems to be equivalent to an XBRL instance.
Google Gemini has the notion of the NotebookLM.  Others are adding other types of capabilities.

Think of it this way. In "phase 1", those tools where "brilliant chat bots" which provided an interface, you could type in test, and the chat bot would provide you answers, and a lot of times the answers were correct or at least useful; but you really never could tell what was right and what was wrong.

Now in 'phase 2", those tools are aspiring to be "an operating system for work". By default, an LLM searches it entire "brain" to answer questions, but that brain might be out of date and it certainly has a lot of information that is not relevant to your context.  Claude's "Projects" and Google Gemini's "NotebookLM" completely flips that around.  Projects and NotebookLM force the AI to act on data that you provided it specifically, in the project or files in that notebook.  That means the software is better grounded and there is less chance to stay off.  Note that I am saying "less chance", it does not mean "no chance".  Also, Projects and NotebookLM provide a way to have a "persistent memory".

Basically, these AI companies are trying to move users away from simply "search and chat" more toward "read information, reason on that information, and build something or perform some type of work".

A big problem that not only LLM (e.g. these synthetic reasoning engines) but pretty much every software application has is interpreting what it is that it's human user means. Make no mistake: computers are dumb beasts.  Really; dumb, dumb, dumb beasts. Full stop.

I mean, think about it.  It is incredibly hard to get two humans to communicate effectively.  It seems pretty darn foolish to me to believe you can magically get a human and a machine to communicate.  Computers do not, and never will, "understand".  Ever.  The very best computers can do is interpret.  There are exactly two ways to interpret.  (1) humans tell the computer exactly what they want which is a "rules-based" or "symbolic AI" approach; (2) humans give the computer a bunch of "stuff" and then let the computer guess as best as it can which is "probability-based" or "machine-learning".

These synthetic reasoning engines have a really hard time understanding documents which contain structured meaning.  For example, electronic spreadsheets. Will these synthetic reasoning engines ever be able to figure out all the different ways different humans arbitrarily create electronic spreadsheets? Sure, to a degree.  But, there is ZERO probability that these synthetic reasoning engines will ever be reliable enough to create industrial processes. Will never happen. We need to separate the "document" costume from the information in those documents.  There are massive benefits from taking this approach. Particularly for accountants, auditors, and analysts. Simply read, Essence of Accounting.

Put a tool in the hands of someone with no skill, experience, or knowledge; that someone will very likely produce an output that is clumsy, ugly, and perhaps barely functional.    But put that same tool in the hands of a skilled, experienced craftsmen with the know-how to use that tool; they can produce works of elegance, utility, beauty, and durability.

If the "stuff" you give a computer is messy, the answer the computer will give you will be messy.  

If the "stuff" you give a computer is logical, well reasoned, well organized, and otherwise clearly communicated and within a set of well articulated boundaries; what the computer returns to you will be well reasoned, well organized, clearly communicated and  therefore the human user will be able to rely on it.

Don't be an "AI nob". Don't fall for the hype.  Be able to differentiate a snake oil salesmen from someone that can truly help you figure out AI.  When you are playing the "AI game"; understand how to "cut the cards" effectively so you don't get bamboozled.

Additional Information:


Comments

Post a Comment

Popular posts from this blog

Internationalized Resource Identifier (IRI)

Image
An Internationalized Resource Identifier (IRI) is a standard proposed by the IETF ( RFC 3987 ).  An IRI is a unique sequence of characters that identifies a resource.   IRIs are a generalization of URIs ( RFC3986 ) that permits a wider range of Unicode characters. A  Uniform Resource Identifier (URI) is a unique sequence of characters that identifies an abstract or physical resource.  The difference between an IRI and a URI is that IRIs use the more modern  Universal Character Set (a.k.a. UNICODE)  and URIs use the older ASCII character set. The newer IRI format is compatible with the older URI format. The most  current version of RDF  uses IRIs to identify resources. Don't confuse IRIs with URLs and URNs.  A URL is a web address. A URN is a Uniform Resource Name. This Venn diagram shows the relationship between these different tools: Why are IRIs and URIs important?  If you are thinking in terms of a single database, a silo ...
Read more

Digital Proficiency

Image
A paradigm shift is occurring.  The territory is changing.  A new map is necessary. Trying to use your old map to understand the new territory will be unsatisfying. This shift is caused by the difference between how "realspace" (the real world, analog) and "cyberspace" (the internet, digital) operate. While many things will stay the same; it is also the case that entirely new business models and products are possible. Probably the biggest mistake one can make is misunderstanding the notion of intelligence and grouping "human intelligence" and "machine intelligence" together. I would contend that it is a fool's errand to try and compare machine-intelligence to human-intelligence.  They are not even in the same ballpark and they won't be for a long time, if ever.  The notion of "superintelligence" in computers is a fantasy. Computers are nowhere close to achieving superintelligence which I would consider as being human-level inte...
Read more

Example Financial Statement Holon

Image
As I have mentioned in other posts on my blog, a financial statement is a formal semantic structure , in fact a financial statement is a special type of formal semantic structure referred to as a holon . As I mentioned in my blog posts, Kurt Cagle and Chloe Shannon are providing a lot of very good information about holons on their substack. ( Here is Kurt Cagle's and Chloe Shannon's answers to my questions. Here is an example Kurt Cagle created of a financial statement as a DataBook.) I have provided a bunch of good examples of financial reporting frameworks and reports   In this blog post I want to focus on one specific financial statement and see if I can explain that a financial statement is a "container" of  blocks of information and the "information blocks" that the financial statement contains. Cagle and Shannon point out that a holon is made up of four distinct knowledge graphs that maximize coherence and operationalize the holon which are: The Sc...
Read more