14 articles
A working agent is a one-off; a practice runs on systems. Three things have to happen for a single agent to become practice infrastructure — it persists across time, ports across clients, and composes with other agents into chains. This piece names the operational disciplines, the platform differences, and the manual chain that lets most CAS practices start today.
You have two pieces of a working agent — an instruction from Part 1, a context file from Part 3. The third piece is the example, and the trick is you don't write it. Run the agent without one. Fix the instruction or the context — not the output — until what comes back is what you'd send. That output is your example.
On April 17, Anthropic launched Claude Design — free on every paid Claude plan, turning prompts and source data into branded slides, PDFs, and reports. For CAS practices, three patterns are worth your time this quarter — and one principle decides whether the tool actually pays off.
Sage and PwC's Beyond the Black Box research put a number on AI's hidden cost: finance professionals spend 12.9 hours a week reconstructing and explaining AI outputs, with 26% of AI time savings lost to verification. The 71% rejection rate the headlines led with is the future cost; the 12.9 hours is what your firm pays now — usually out of partner time. The fix is to stop shipping AI deliverables alone and start shipping them with a research memo and a cross-model verification report.
For months I've been writing about client context like its meaning is obvious. It isn't. This is the article where we stop talking about context engineering in the abstract and actually build one — a single document that makes every agent you run for a client better, and that lets you take the same agent to every other client just by swapping files.
For six weeks, the AI tool most CAS firms now run on was producing degraded output, and the vendor didn't know until customers complained. Three tests every CAS firm should build this month — drift detection, portability evaluation, and version trail accountability — to know when your AI vendor breaks something quietly.
During a webinar on AI-powered client coaching, a colleague shared a 25-question diagnostic survey in Excel. It was ugly, had scoring errors, and looked like what it was — a spreadsheet pretending to be a tool. Thirty seconds of prompting later, it was a polished interactive HTML page that captured, scored, and analysed responses. The quality ceiling on what you hand your clients just disappeared.
I built an AI skill for something completely unrelated to accounting — booking flights — and learned more about context engineering in two hours than months of client work had taught me. The first draft captured my obvious preferences. The corrections surfaced knowledge I didn't know I had. That's the encoding gap, and it's the reason most firms' AI stays mediocre.
The research was valuable. What happened next was more valuable — a 10-minute reflective conversation with the AI that turned a one-off project into a reusable methodology anyone in the firm can follow. That's context engineering. And it's how you stop losing what you learn.
Google Cloud says multi-agent orchestration is the architecture of 2026. Basis — now valued at over a billion dollars — is already running it for Top 25 accounting firms. The question isn't whether this model works. It's whether your practice will build it with your own specifications or let a vendor do it for you.
Xero's updated developer terms now prohibit external AI agents, ban API data training, and impose $99–$895/month access costs — while launching their own proprietary AI. The message is clear: platforms will monetize AI access and restrict yours. The highest-leverage response is to stop waiting for platform access and start building in the gaps where vendors have no gatekeeping power.
A software engineer tried three times to build an AI dashboard before realizing the answer was markdown files and structured knowledge. The same lesson applies to your CAS practice — the competitive moat isn't software, it's encoded practice knowledge that compounds.
