Agent Builder 3: Your AI doesn't know who your client is

Two drafts of the same email. Same AI. Same prompt. Both asking a client for missing documents. Both generated in about two seconds.

The first:

Hi [Client Name], I hope this email finds you well. As we approach the end of the month, we wanted to reach out regarding some documents we're still missing for your monthly bookkeeping. Could you please provide the following at your earliest convenience? [list]. We appreciate your prompt attention to this matter. Best regards.

The second:

Mike — month-end's coming up and I'm still chasing a few things for Glendale Brewing. Can you send through the Stripe payouts for April (the summary is fine, I don't need every transaction) and the last two fuel receipts? I know you'd rather email than use the portal, so flagging it here. No rush — just before the 10th. — Peter

Same tool. Same task. The second one isn't smarter AI. It's the same AI with a client context file attached.

This is Part 3 of a five-part series on building AI agents for your CAS practice. Part 1 covered what an agent is. Part 2 covered where to place your first one. This piece shows you what the single most valuable artifact in your AI stack actually looks like — and how to build it.

What we've been calling "client context"

For months I've been writing about client context and context engineering like those terms mean something self-evident. They don't. Across this series and the articles before it, I've argued that what your AI knows about your client is the difference between generic output and output that sounds like it came from your firm. I haven't shown you what that knowledge looks like on paper. This is the article where I do.

The short version: a client context file is one document, kept alongside everything else you have for that client, that captures every non-obvious thing a new staff member would need to know before drafting an email, reviewing a statement, or jumping on a call. Industry, owner, financial shape, compliance footprint, communication preferences, history — the institutional knowledge that normally lives in the partner's head.

Once it exists, every agent you run for that client draws from it. And the file doesn't degrade, doesn't walk out the door, doesn't need to be re-taught to a new hire in their third week.

Last week's backlog needs this

If you did the Part 2 exercise, you ended with a Z1 list — work items your first agents should handle. Drafting follow-up emails. Generating meeting prep notes. Writing engagement letters. Producing recurring summaries. Building onboarding checklists.

Every one of those agents needs to know something about this client. The email agent needs the owner's communication style. The meeting prep agent needs to know what the client cares about. The summary agent needs to know what's unusual this period. Same client, different tasks, heavily overlapping knowledge.

You could write the context into every agent as you build it. Or you could write it once, in a shared file all of them pull from. Every refinement improves every agent at once. That's the first axis of compounding.

One file, every agent — and the flip side

Here's the compounding pattern most practitioners miss.

Same client, every agent. One context file powers every agent you build for this client. The document-chaser, the meeting prep generator, the advisory follow-up, the year-end planning brief — all of them draw from the same file. Write it once, every agent improves. Refine it, every agent improves again.

Same agent, every client. Once you've built an agent that works — say, a post-engagement summary generator — you can run it against a different client just by swapping the context file. The instruction stays fixed. The knowledge changes. One agent definition, fifty client context files, fifty working agents across your book.

Those two axes are why context engineering compounds and why the client context file is the highest-return artifact in your AI stack. Every time you refine an agent, it gets better across all your clients. Every time you refine a client file, every agent gets better for that client. Nothing else you can build for AI in your practice has that kind of double reuse.

The shortcut — let AI do the first pass

Before we get to the field list, here's a practical shortcut. If your AI tool has access to your email and your file storage, you don't have to fill in the blanks from memory.

If your LLM is connected to your email — whether that's Claude with a Gmail or Outlook connector, ChatGPT with its integrations, or Copilot inside Microsoft 365 — give it this kind of prompt:

"Search my emails from and to Mike Jensen at Glendale Brewing over the last twelve months. Based on the correspondence, answer these questions: How does Mike prefer to communicate? What topics does he raise repeatedly? What has he expressed frustration about? What has he said he cares about? What requests has he made about how we work together?"

If your LLM is connected to your file storage — Google Drive, SharePoint, Dropbox, or whatever you use — point it at the client's folder:

"Look in the Glendale Brewing folder. From the documentation in there — financial statements, tax returns, engagement letters, prior working papers — answer these questions: What does this business do? What industry are they in? Who are their main customers? What does their recent financial performance look like? What sales tax jurisdictions are they registered in and at what frequency? What are their compliance deadlines?"

Take whatever the AI produces, check it for accuracy, correct what's wrong, and add what's missing. In about ten minutes you'll have 70 to 80 percent of the file drafted. The remaining 20 to 30 percent is the institutional knowledge only in your head — which is exactly the part that's hardest to capture any other way.

What to put in the file

Ten fields. Fill in what's relevant, skip what isn't, add anything specific to the engagement.

1. What the business actually does. Industry, core products and services, major customers or customer segments, revenue model. Not the tax-code classification — the actual business.
2. Financial profile. Revenue size, growth trajectory, profitability history. Is the business consistently profitable? Growing? Declining? What are the numbers a stranger would need to have an intelligent conversation?
3. Owner and stakeholder profile. Salary versus dividend strategy. Personal income goals. Exit horizon. Whether this business is the owner's sole income or one of several. Family members on payroll or shareholders.
4. Compliance and regulatory profile. Sales tax jurisdictions and filing frequency. Payroll tax obligations. Income tax filing entities, fiscal year-end, and deadlines. Any industry-specific regulatory requirements — liquor licensing, trust accounting, healthcare billing rules, whatever applies.
5. Standard deliverables and cadence. What you produce for them, how often, and by when. Monthly close packages, quarterly reviews, annual tax filings, ad-hoc advisory.
6. How the owner communicates. Preferred channel. Tone. Level of detail they want. Response patterns. Anything they've told you explicitly about how they want to be contacted.
7. What they actually care about. The real drivers behind their decisions — cash flow, growth, tax minimization, clean books for sale, keeping the spouse on side. Not what they list on their website — what they talk about on calls.
8. What's unusual. Business-specific quirks, seasonal patterns, ongoing disputes, non-standard accounting treatments, unusual related-party arrangements.
9. What went wrong last year. Problems, corrections, late filings, hard conversations. What you don't want repeated.
10. Known preferences and pet peeves. "I hate PDFs." "Don't CC my husband." "Just give me the number, not the commentary." The stuff they've explicitly asked for or complained about.

Now pull up your Z1 list from Part 2. Scan each item. For any agent on that list, is there something specific it would need to know about this client that isn't in the ten fields? Add it at the bottom.

Try this now — 30 minutes, one client

Pick the same client you mapped in the Part 2 exercise. Open a blank document and title it "[Client Name] — Context File." File format doesn't matter — a plain text file, a Markdown file, a Word doc, a Google Doc, anything your LLM can open will work.

If your AI has email and file storage access, run the two prompts above and review what it produces. That gets you most of the way.

If it doesn't, or you'd rather start from memory, fill in the ten fields directly. Fast. Whatever comes to mind first.

Either way, you're done when the file fits on two to three pages and a competent stranger could pick it up, read it, and write that client a coherent email, draft their engagement letter, or brief you for a meeting — without calling you first.

Build it once. Every agent, every client.

That document is the most valuable thing you'll build in this series. Every agent you run for this client will be better because of it. Every time you refine it, every agent improves at once. Every new agent definition you develop can be ported to any other client just by swapping context files. And every new client you onboard gets a matching file built on intake, not retroactively six months in.

Next up in the series: the other two pieces of a working agent — the instruction (we started this in Part 1) and the examples that show it what good looks like. We'll assemble all three into a working agent you can run in 30 minutes, using the context file you just built.

If you want to start building these skills systematically, AI Essentials is the structured course that takes you from foundations through to a fluent practitioner.