Ask the Expert: Context and AI
David Yue, Co-Founder & CEO, Accordance, discusses why good results depend not only on prompts but also on “context.”
Accordance was founded by AI researchers-turned students-of-practitioners who dedicated their first years to working alongside tax leaders, professors, and firm partners. From Stanford AI Lab to CEO, David Yue continues with that same philosophy: learning from the practitioners and firms that shape Accordance.
Q Why does AI still feel hit-or-miss for me?
A Folks tend to focus on the prompt, but truly great results depend on the “context” you can provide.
Tax questions without facts are just hypothetical.
A tax question with the client’s partnership agreement, the prior year’s return, and the relevant Treasury regulation becomes a research problem with nuanced answers. Even the same model with the same prompt produces completely different output when one has rich context to reason from, and the other doesn’t.
The emerging skill in AI is structuring the context for your model. Models have gotten remarkably good at reasoning, but they need comprehensive context to really shine.
Q What actually counts as ‘context’?
A Think of it as layers.
First, the facts of the engagement: a client’s prior returns, the entity structure, the operating agreements, the depreciation schedules — whatever documents the position actually depends on. Be sure to anonymize or use models with robust data controls.
Second, the controlling authority: the actual Code section, the regs, the relevant rulings, the case law in your jurisdiction. AI trained on the open internet has seen plenty of tax content. Very little of it is primary.
Third, your firm’s prior thinking: the memo written two years ago on a similar fact pattern, or the reasoning your senior partner walked through when this came up last busy season. Most tools have no way to reach this material.
Each layer compounds on the others. Skip one and the answer gets thinner.
Q Doesn’t this mean more prep work
for every research question?
A Yep, but organizing your context upfront makes a big difference. Structure and store client documents to be accessible by a model with appropriate data controls. Support your research with primary authority and a rich citation engine. Curate your firm’s sample documents, templates, and operating procedures.
This context will compound over time. Last quarter’s engagement becomes the context for next quarter’s. A memo written once becomes part of the firm’s reasoning for similar circumstances. The accumulated thinking that used to live in a partner’s head, or in a folder no one opened, can start informing the model.
That’s the environment we’re focused on building within Accordance. We carefully curate client files, fact patterns, primary authority, and prior work products into a rich contextual foundation. That substrate supports an array of agents that research, fact-check, cite, reason, audit, and question each scenario you need to explore.
Sponsored by:
