Most technology your firm uses runs on explicit logic: rules, conditions, and defined outputs. If a client owes above a threshold, flag it. If a form field is missing, stop the process. The logic is traceable and the outputs are predictable.
Generative AI works differently. Tools like Claude and ChatGPT are built on neural networks that learn patterns from enormous volumes of text rather than following instructions someone programmed. That’s why they can write, reason, summarize, and explain in natural language. It’s also why they require a different set of expectations than the software you’re used to.
How a large language model actually works
A large language model reads your input, breaks it into fragments called tokens, and predicts the most useful response based on patterns it learned during training. Think of it as sophisticated autocomplete shaped by an extraordinary breadth of text. It isn’t just retrieving facts from a database (though modern models have tools to do this). It’s generating a response that fits the context you’ve given it.
Two things follow from this.
First, specific prompts produce better results than vague ones. “Summarize this client situation and identify any missing information before I respond” will consistently outperform “what should I say.”
Second, these models can sound confident while being wrong. Review their output the way you’d review work from a capable but new staff member.
Chat versus Cowork
Claude operates in two modes built for different types of work (Claude Code is third mode but can be thought of as its own product outside of Claude).
Chat is the familiar back-and-forth. It’s well-suited for research, writing, summarization, and working through a problem step by step. One underused technique: explicitly ask the model to challenge your assumptions or push back on weak reasoning. Left to its defaults, it tends to agree with however you’ve framed a question. Counteract that directly, and get in the habit of prompting with less leading. Open-ended questions surface more than narrow ones. Instead of asking “Isn’t this client situation a candidate for the safe-harbor election?”, try “Here’s the client situation — what options should I be weighing, and what would change your answer?” The first prompt steers the model toward the answer you already suspect; the second lets it tell you what you’re missing.
Cowork is more structured. It builds a plan, asks clarifying questions, and executes across multiple steps using connected tools and data sources. This is where the time savings become meaningful: automating routine workflows, aggregating information from multiple systems, and preparing outputs that would otherwise require several manual steps. Skills in Cowork handle the repetitive analyses your team runs the same way every time. Instead of re-explaining the steps in each new chat, you write them down once — “pull the trial balance, compare to prior period, flag accounts with variance over 10%, draft commentary in our standard format” — and invoke the skill by name. Good candidates: monthly close variance commentary, new-client intake summaries, standard responses to common IRS notices, and quarterly review prep packets. The first version usually needs a few rounds of refinement; once it’s right, the output is consistent every time anyone on the team runs it.
Connected systems are where the value compounds
AI tools do generic work without context. They do meaningful work when connected to the systems your work actually lives in.
Connecting Microsoft 365, for example, gives the tool access to your calendar, email, documents, and team communications. A practical result: a daily briefing workflow that pulls together the morning’s calendar, flags urgent messages, and surfaces links to relevant documents on a schedule. The individual pieces aren’t technically complex. The value is in eliminating the manual aggregation that would otherwise start every day.
Other examples from the session: reorganizing files, updating slide decks from a source folder, converting data from one format to another, and researching options before a purchasing decision. None of these require coding knowledge. They require clear instructions and the right connections.
What the tools won’t do on their own
The tools won’t spontaneously invent workflows you haven’t defined or take consequential actions you haven’t authorized. Cowork plans and executes within the tools, connections, and permissions you’ve set up — it can use your files, connected apps, installed skills, and a browser when you’ve given it access, but the scope is yours to set. If you want a repeatable analysis run consistently, define it as a skill: a saved set of instructions that encodes exactly what you want done. That keeps outputs consistent and the work under your control.
If anyone on your team uses personal AI accounts for client-related work, that’s worth addressing. Enterprise accounts under a service-level agreement don’t use your data for model training. Personal accounts may have different defaults.
Where to start
Install the desktop application and connect Microsoft 365 or Google Workspace first. That connection covers email, calendar, documents, and Teams (for Microsoft users), and it gives the tool the context it needs to do useful work. Set your memory preferences. Then identify one workflow that costs you time every week and build something that handles it.
The firms making real progress aren’t running large change initiatives. They’re solving one specific problem well and moving to the next.
