Uber pushed past engineering automation and into finance, legal, ops, and support. What worked was not a bigger model or a bigger budget. It was a two-week pairing ritual: agentic pods.
The backstory is what makes it land. Uber is the original tokenmaxxer; its operating chief spent the spring talking about token bills that made the market blink. They bought everything and let the experiments compete.
These pods are what survived.
The constraint was not compute. It was proximity to the real work.
What a pod is
One AI-fluent engineer, one domain expert, ten working days.
Days 1–2: shadow the expert. No code.
Day 3: pick one workflow, scored on scale, repetition, impact, and data availability.
Days 4–5: build with the operator, never for them.
Days 6–9: validate with several others who do the same job.
Day 10: ship.
Sixteen pods, sixteen functions, two months.
Finance won twice in Uber's own highlight reel
Capital allocation across 150 cities: 15 hours → 30 minutes
Financial pacing reports: 2 days → 10 minutes
Finance work is pod-shaped: calendar-repetitive, exception-heavy, spread across more systems than any other function, and rich in data that already exists.
Shadowing comes first because the process doc is a fiction
The real workflow lives in exceptions, workarounds, and side spreadsheets nobody wrote down. Automate the documentation and you automate the fiction.
Uber's own phrase: the wins were "hiding in plain sight." They were only invisible from the outside, where every consultant and steering committee stands.
The workflow is the unit of automation
Task automation saves hours, and hours leak unless you instrument them (our mid-year scorecard).
Workflow redesign removes line items:
Handoffs eliminated
Unnecessary approvals removed (Uber's word, and the right one: process sign-offs that never added control; the accounting controls stay)
Legacy tools retired
Vendor spend cut
Charter pods against workflows, and score them on what they remove.
To be clear about the ledger: Uber's finance wins are analysis and reporting workflows. Nothing in their post touches the books without a human, and neither does anything we build.
Two controls worth stealing
The two-week clock: a pod that cannot ship in two weeks picked the wrong workflow. The clock forces the pick before the build.
Day six: an agent that works for one person is a personal script. An agent that survives validation across several operators is a control-worthy tool.
Your version is one pair and a calendar
You do not need thirty spare engineers. You need: one fluent builder, your controller, and ten days.
Ship behind the usual bar: deterministic math, a checking layer, a named approver. Then charter the next pod against the next workflow.
Uber's real move was making the pod a repeating unit of how the company changes. Copy that part first.
THE NUMBER
THIS WEEK IN AI-NATIVE FINANCE
OpenAI launches ChatGPT Work, an autonomous agent that builds finished spreadsheets, decks, and docs, on new GPT-5.6 models. The Codex-powered agent gathers context across your connected apps and files and works on its own, with scheduled tasks and browser tools. GPT-5.6 ships in three token-priced tiers (Sol $5/$30, Terra $2.50/$15, Luna $1/$6). It aims squarely at the spreadsheet-and-deck work finance teams do daily, and the per-token pricing means you model this as variable, consumption spend, not seats.
Meta opens its first paid model API, pricing Muse Spark 1.1 at $1.25/$4.25 per million tokens to undercut OpenAI and Anthropic. A multimodal reasoning model built for agentic tool use, with a 1M-token context window and drop-in compatibility with both the OpenAI and Anthropic API formats. A credible third frontier vendor at roughly half the price of rivals widens your negotiating leverage and lowers the cost of re-pricing an existing AI stack.
Apollo and Blackstone assemble a record $35 billion private-credit package to finance Broadcom and Anthropic's AI chips. A special-purpose vehicle buys the custom chips and leases them to Anthropic, drawn over roughly a year, with Broadcom backstopping the senior tranches. This is the template for moving AI capex off the corporate balance sheet into off-balance-sheet SPVs financed by private credit. Expect leasing and vendor-financing structures of this shape in your own capex and covenant conversations.
Axos Financial agrees to buy AI-native fintech Arc Technologies to add agentic cash management. An incumbent bank buying an AI-native treasury and capital-markets platform, expected to close later in July. A clean signal that incumbents are acquiring rather than building agentic finance capability, and a reminder to underwrite the platform risk of the venture-backed fintechs holding your cash workflows.
FROM THE FIELD
In a finance R&D session this week we mapped what the AI accounting tools on the market actually automate, and where they quit. Nearly all of them solve the same three things: connecting the data sources, categorizing transactions, and reconciling accounts. Almost none automate the domain-specific work that eats a controller's month in a services business: collecting the billing and retainer data, making the accrual judgments, and posting the AR and AP journal entries. The crowded part of the market is the easy part.
If you are evaluating an "AI does your bookkeeping" tool, watch exactly where the demo ends. Most stop at a clean, categorized, reconciled trial balance, which is real work but not the hard part. The judgment-heavy middle, deciding what to accrue, splitting revenue and cost the way your business actually recognizes them, and posting the entries, is where the hours and the risk live, and it is largely still yours. "Connected and categorized" is not the same as "closed," and the distance between the two is the whole game.
WHAT WE’RE READING
71% of finance leaders would reject a 99%-accurate AI tool that cannot explain itself, and teams lose about 13 hours a week to a "verification tax" reconstructing and defending AI outputs. It reframes finance-AI ROI around explainability and hidden review cost, not raw accuracy.
Agents resend context and retry, so total spend climbs even as per-token prices fall. Goldman Sachs projects a 24-fold jump in token consumption by 2030. A usable unit-economics framework for 2026 AI budgeting, and a warning against the seductive assumption that cheaper models mean cheaper bills.
AI now does the manual Excel modeling that trained him for a decade, so tomorrow's finance talent needs data literacy (SQL, warehouses, business logic) alongside financial acumen. Essential reading for anyone rethinking hiring and apprenticeship.
The differentiator is not what an agent produces but whether it knows what it cannot discern and hands that back for judgment. A clean practitioner framing of where to trust autonomous execution versus require human sign-off.
Christian Sanford and Kenny Jen are the co-founders and managing partners of QuantFi, where they help PE- and VC-backed companies build investor-grade finance functions powered by AI-native infrastructure.
Christian came up through investment banking at Barclays ($20B+ in M&A and capital markets), buy-side investing at a hedge fund, and fractional CFO work for investor-backed companies. BBA and MSA in Accounting from Texas Tech.
Kenny led financial strategy at Pilot CFO, Pure Beauty, and Emil Capital Partners after starting his career at Credit Suisse. He has overseen finance for consumer, SaaS, and manufacturing businesses, and works directly with founders on pricing, working capital, and capital allocation. BSBA in Finance from Georgetown.