AI Is Writing Our Future (But It Still Sucks at Long Division)

Every week, I interview founders and operators rewriting the rules of finance, tech, and leadership. This round, I sat down with Ray Sang, former Google controller, CPA, law student, and the visionary behind Inkwise.ai (policy-aware writing copilot) and Chipmunk Robotics (process automation that feels like shop class on steroids).

We spent an hour screen-sharing exactly how generative AI is already being used in audits, policy drafting, and deal work. Below is my deep dive into that conversation, paired with a month's worth of research on AI risks, opportunities, and field-tested guardrails.

If you love nerding out on the future of technical writing and want the full demo of Inkwise, catch the YT video [here] or Spotify podcast [here]. Let’s jump in.

TL;DR (1-Minute Version)

• Three live AI workflows dominating the marketplace: Copilot (assistant), Chatbot-Drafting, and Agent Orchestration.

• First-mover advantages: Less than 10% of firms use generative AI, yet funding for AI tools in tax & audit has surpassed $2 billion.

• Landmines for early adopters: AI hallucinations (especially in math), robotic writing tone, and million-dollar privacy fines.

• Actionable checklist: Sandbox a copilot with your team, anonymize sensitive data, and mandate citation verification on your close checklist.

If that’s all you have time for, no worries, I won’t be offended. For everyone else, grab a coffee. Let’s talk about keeping your workpapers safe from Skynet while shipping them faster than your audit partner hits "reply all."

1. The Three Camps of AI Workflows – Where Does Your Stack Sit?

Three clear AI workflows are emerging in professional writing, finance, accounting, legal, and beyond. Ray and I unpacked them thoroughly, and below is every insight, risk, and opportunity gleaned from our discussion and recent research.

Copilot / Assistant (Inkwise.ai lane)

• What it looks like: You write (imagine that!), the copilot suggests citations, flags inconsistencies, brainstorms competing arguments, and reminds you ASC 606 isn’t an energy drink.

• Sweet spot: Nuanced writing where precision is paramount, policy memos, SEC comment letters, complex contracts.

• Red flags: Potential over-reliance may dull critical thinking; ensure the team stays actively engaged.

Chatbot Fine-Tuned Models( TAbot lane from my prior episode - Harton Wong & Nina Zhao, CPA)

• What it looks like: Prompt a model, get a few pages of content, then edit out gems like “gross margin = 121%”.

• Sweet spots: Quick first drafts, outlines, preliminary research, and brainstorming. Recommend a human layer expert to review outputs.

• Red flags: High risk of hallucinations, compliance/privacy exposures, and instability from ever-changing LLM models.

Agent Orchestration

• What it looks like: A swarm of bots that scrape docs, draft sections, and alert humans to exceptions.

• Sweet spots: High-volume grunt work (“download 300 peer 10-Ks, tag footnote deltas”).

• Red flags: QA bottlenecks, bots often celebrate completed tasks even when botched.

Ray plants his flag in Camp 1: "Keep professionals in the driver's seat; AI should just be the assistant who never sleeps." After testing Inkwise, I get the appeal, it behaves like a hyper-contextual Track Changes pane: citing guidance, cross-checking memos, and suggesting alternatives only when asked.

No camp is inherently “right.” Choose the tools matching your risk tolerance, deadlines, and the patience of your audit committee.

2. Why Adoption Is Still Tiny

• Fewer than 1 in 10 accounting firms report using generative AI, even fewer within corporate controllerships.

• Among Fortune 500 finance execs, barely 1 in 3 have begun experimenting.

• Most early adopters are younger, tech-adjacent, or side-project enthusiasts, leaving an ocean of legacy workflows still stuck with Word templates and copy-paste macros.

If you’ve watched audit seniors lug binders of tick-marked workpapers years after Excel became mainstream, you know tool adoption moves glacially. Same story here, just magnified. For builders, slow adoption equals runway. Lean teams can redefine stale products as AI-native: faster, cheaper, and less headcount-intensive. Investors are already noticing.

That means the competitive moat isn’t tech, it’s courage. If you’re live-testing generative AI while your competitors debate it, every race starts at the 50-yard line.

3. The Big Risks We Can’t Ignore

3.1 Hallucinations – Math Edition: LLMs confidently deliver faulty arithmetic. Berkeley researchers got ten different incorrect answers from ChatGPT when solving the same algebra problem multiple times. I’ve personally witnessed:

• Invented discount rates in cash-flow models

• Nonsensical ratio spreads

• Phantom ASC references that don’t exist

Guardrails:

• Independently recalculate any AI-generated numbers.

• Require manual verification of every AI-sourced citation.

• Controllers should revisit SOX controls and review checklists accordingly.

3.2 Authenticity and Trust: Frequent ChatGPT users spot the tell-tale AI rhythm (heavy bullets, awkward contrasts, phrases like "In conclusion"). Readers and investors notice. Ray recommends: Write first, AI-polish later, keep your voice authentic.

3.3 Privacy and Confidential Data: CFOs hesitate to upload sensitive contracts to public LLMs - rightly so. Regulators frequently fine firms for mishandling user data. Cyberhaven found companies leaking sensitive data to ChatGPT hundreds of times weekly.

My recommendations:

• Obfuscate identifiers before uploading.

• Use structurally identical anonymized data, replacing real figures offline.

• Document this anonymization process within SOX controls. Mastery here is a competitive edge.

4. Inkwise in Action – A Deep Dive on Co-pilots

Ray designed Inkwise to feel like Microsoft Word with superpowers. The functionality is novel, so here's the full breakdown:

• Policy/Reference Vault: Drag accounting manuals or style guides into a private library.

• Context Snippets: Highlight the context and portions of references being relied on for the AI-generated content automatically.

• Live Predictive Writing: Iteratively auto-complete paragraphs with contextual knowledge of the document structure, current thought process, and references.

• Inline Brainstorming: Request alternate structures or devil’s-advocate feedback without leaving the document.

Hundreds of lawyers and accountants already use Inkwise or similar tools. They don’t just get faster, they become better professionals. Experimentation is highly recommended.

5. Where We Go From Here

Generative AI will revolutionize technical writing as Excel replaced paper ledgers, but only if your firm learns it. To recap, we have:

• Copilots augment experts without sacrificing authenticity.

• Chatbots quickly draft but need ruthless editing.

• Agent orchestration scales grunt work but requires robust QA.

Adoption remains single-digit, leaving ample space to leapfrog incumbents. Winners will master all three modes and shift gears fluidly.

6. Watch the Demo, Join the Dialogue

• Video: Full Inkwise demo with Ray, including policy uploads and consistency scans [here].

• Podcast: Listen on your commute [here].

• Ask Me Anything: Drop your questions on hallucinations, privacy, or agents. I'll queue them up or respond directly.

7. Up Next on Ambition Aligned

• Claire Tsukuda (Founder & CEO, Kipsi): Building finance & accounting products that survive customer scrutiny and startup culture insights.

• Kelsie Neibel (Top Bay-Area Sales Leader): Balancing discipline, family life, and sales tech overload without burnout.

• Nickolaus Violin (CEO & Founder, Konnekted): Startup realities, securing funding, and navigating traps while staying ethical.

Got a spicy take on hallucinations or agent QA? Hit reply or message me on LinkedIn. The best insights make the next episode.

Stay curious, edit your ratios, and please, don’t let a chatbot calculate deferred revenue alone.

- Devon