A few months ago I watched a colleague ask an AI agent to “clean up” a folder of client files. She meant organize them. The agent deleted duplicates. In 30 seconds it made decisions that took 20 minutes to sort out. The agent wasn’t wrong — she just hadn’t told it what she actually meant.
That’s the world we’re in now. Claude Cowork launched in January 2026. Microsoft Copilot Cowork followed shortly after. Intuit announced a partnership to bring QuickBooks and TurboTax intelligence directly into Claude. The tools are here, they’re being adopted quickly, and most of us haven’t thought through how to actually talk to them.
Context engineering is the answer. It’s not prompting. It’s not a cheat code. It’s a professional skill — one that maps well onto things accountants already know how to do.
Why This Is Different From the AI You’ve Been Using
Think about how AI tools have evolved over the past few years. ChatGPT came out and it answered questions. You asked something, it gave you text, you decided what to do with it. A bad answer was annoying but harmless — you just ignored it.
Then AI got embedded in tools. Microsoft Copilot started showing up in Excel, Outlook, and Teams. It surfaced information and drafted content inside workflows you were already in. Still mostly read-only.
Now we’re in the agentic era. These tools have read/write access. Claude Cowork can open your files, create new ones, move things around, run code, and connect to tools like Salesforce, Google Drive, and Slack — all in one workflow without you clicking each step. The Intuit integration means agents will soon be able to touch your actual client accounting data.
The stakes are different now. A bad prompt used to give you a bad answer you could trash. A bad input to an agent can overwrite a source file, misroute a document, or kick off a downstream process you didn’t intend. The agent doesn’t know it made a mistake — it did exactly what you asked.
What Context Engineering Actually Is
When AI first came out everyone was talking about prompt engineering. That era is mostly over. The models have gotten better and the tricks matter less. What matters now is the broader picture of information you give the tool before it starts working.
Context engineering is the discipline of deliberately shaping what an AI agent knows before it acts. The goal is consistent, repeatable outputs — not getting lucky on a single run. If you’re building a workflow that runs in the background regularly, you can’t babysit it every time. You need it to get it right a thousand times, not just once.
The Four Dimensions of Context
1. Task Context
What exactly are you asking the agent to do? Vague instructions get vague or dangerous results. “Reconcile this spreadsheet” is not a task. It’s a hope. The more specific version: “Compare column B in Sheet 1 against column A in Sheet 2. Flag mismatches over $500. Write results to a new tab called Exceptions. Do not modify the original data.”
2. Role and Expertise Context
General agents don’t know GAAP. They don’t understand three-way matching logic, ASC 842, intercompany netting, or what SOX documentation actually requires. They apply general pattern recognition — not accounting judgment. You have to supply that expertise explicitly, including attaching relevant memos and policies when they matter.
3. Environmental Context
What data and systems does the agent actually have access to? This is the most underrated question. My current approach: read-only access to live systems, with a dedicated sandboxed working folder for anything the agent creates or modifies. When I’m satisfied with the outputs I push the changes myself.
4. Constraint Context
What must the agent not do? Written instructions help — never delete files without confirmation, always show me your plan first — but they are not a hard stop. These are LLMs. If you want a true constraint, enforce it at the tool access level. Restrict what the agent can touch, not just what you’ve asked it to avoid.
The Accounting Issues This Creates
The Audit Trail Problem
Every action an AI agent takes has to be explainable and auditable. Most tools log their reasoning — but that means you have to read through that log before you sign off. Context engineering is how you ensure the agent’s reasoning is tied to real documentation, not inferred from ambiguous inputs.
Human-in-the-Loop Is Not Optional
AI is non-deterministic. Run the same task twice and you might get a different result. In accounting, where close enough is never an option, you need human verification checkpoints in any high-risk workflow. Agent gets context, shows plan, human approves, agent executes, human spot-checks. It sounds like more work. In practice it’s faster than fixing errors after the fact.
Two Mental Models That Actually Help
Brief your agent the way you’d brief a new staff member on a deadline. Give them the deliverable, what’s already been done, the constraints, and who to check with if something looks wrong. That’s the same information structure a well-engineered agent context needs.
The second is the workpaper mindset. Auditors already think about what a reviewer needs to understand their work. Context engineering applies the same question to AI. If you couldn’t document it in a workpaper, don’t assume the agent knows it.
Key Takeaways
- Agentic AI tools now execute multi-step tasks inside your files and workflows — a bad input can have real consequences.
- Context engineering covers four dimensions: task, role, environment, and constraints.
- General-purpose agents don’t know GAAP or SOX — supply that professional context explicitly.
- Enforce real constraints at the tool access level, not just in your prompt instructions.
- Build human review checkpoints into high-risk accounting workflows.
- The goal is consistency at scale — workflows that get it right a thousand times, not just once.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply