What building an AI-native finance function taught me
The finance industry's rush to bolt AI onto old workflows is running into a harder truth: precision, not creativity, is the job.
Written by OutOfToken AI
August 10, 2026 · 5 min read · Synthesized from reporting by OpenAI Blog · How this works
A blog post attributed to OpenAI CFO Sarah Friar frames a set of lessons drawn from rebuilding a finance function around AI rather than layering AI on top of it. The piece reportedly touches on automated forecasting, stronger controls, and how to measure AI's actual return on investment. Independent reporting hasn't corroborated the specific post or confirmed the five lessons in detail, but the themes align with a broader shift already visible across finance teams experimenting with large language models.
AI-first isn't AI-native
Across the finance-AI conversation right now, one distinction keeps surfacing: bolting a chatbot onto existing spreadsheets and approval chains is not the same as redesigning workflows so AI is the default interface. Practitioners building genuinely AI-native teams describe a slower, more deliberate process — feeding models business context, structuring data so it's machine-readable, and rethinking what a human actually needs to approve versus what a system can handle end to end. The gap between demoing an AI feature and scaling it into daily operations is where most of these efforts stall.
Why finance resists the free-form model
Large language models are prized for flexibility — they're good at ambiguity, brainstorming, and interpreting loosely defined requests. Finance needs almost the opposite. The same input has to produce the same output every time, and every number needs a traceable path back to its source, which is why auditability and determinism keep coming up as prerequisites rather than nice-to-haves in this space.
"Finance and accounting demand deterministic precision and mathematical certainty — the same input must always produce the same output, every single time."
The context trap
One recurring lesson from teams that have actually tried to get AI to analyze financial scenarios, rather than just describe a chart, is that more context isn't automatically better. Dumping full historical data, entity definitions, and hierarchies into a single prompt can overwhelm a model and degrade its output rather than sharpen it. The more effective pattern involves pre-computing what can be calculated deterministically outside the model, then handing the AI a narrower, cleaner task — separating summary generation from detailed analysis rather than asking for both in one shot.
Controls and ROI as the real test
The claimed emphasis on 'stronger controls' and 'AI ROI' in Friar's reported lessons tracks with what's emerging as the actual bottleneck in enterprise AI adoption: not whether a model can produce an answer, but whether an organization can trust it and justify its cost. Multi-modal models that read PDFs, spreadsheets, and images are increasingly treated as the baseline for finance deployments, but the harder work is building the guardrails — audit trails, deterministic checks, human sign-off points — that let a finance team actually rely on AI-generated numbers rather than just admire them.
Whether or not this exact article and its five lessons can be pinned to Sarah Friar, the direction of travel is consistent across the finance-AI landscape: automation is easy to demo and hard to trust. The functions that get this right will likely be the ones that treat determinism and auditability as design requirements from day one, not patches applied after the first embarrassing mistake.
Editorial Note
The research strongly corroborates the article's core claims about AI-native finance functions, determinism requirements, context limitations, and implementation challenges. Multiple independent sources (LinkedIn articles, Reddit discussions, company blogs) validate the key themes. The one unverified element is the specific attribution to OpenAI CFO Sarah Friar and her 'five lessons' — the research discusses these topics broadly but doesn't confirm the original blog post's existence or specific framing.
Claim Tracker
AI-assessed
Source 6 (Jennifer Morrison) states: 'AI-first vs. AI-native. The difference is subtle, but it can make or break your product. AI-first means building because the tech exists. AI-native means building where the tech actually fits.'
Source 5 (Building Luca) directly states: 'But finance and accounting demand the opposite: deterministic precision, mathematical certainty, and complete auditability. The same input must always produce the same output, every single time.'
Source 3 (Reddit discussion on financial scenario analysis) confirms: 'Dumping everything—full entity definitions, all historical data, component hierarchies—ballooned prompts to 15K+ tokens and _confused_ the model. More context isn't always better; it's often worse.'
Source 5 (Building Luca) states: 'This flexibility makes LLMs brilliant at tasks like writing, brainstorming, and interpreting ambiguous requests.'
Source 6 (Jennifer Morrison) notes: 'Everyone loves the demo. Scaling it is where teams bleed.'
Ask AI about this story
// discussion
sign in to join the discussion