JOnRamp 2026

OnRamp 2026 · San Francisco · September 2026

Automating a bad process just makes it fail faster

I went to a conference about AI in Finance. The most useful thing I learned had nothing to do with AI.

01The trigger

Consistently wrong, at scale.

Spend dataIllustrative

Transactions

Same label, two meanings
ItemField AField BAmount
LabelLabel
LabelLabel
LabelLabel
LabelLabel
LabelLabel
Report
Report
Report
J
F
M
A
M
J
J
A
S
O
N
D

  1. Same label. Two fields. Two meanings.

    Found midway through a session, in our own spend data.

  2. People fill the gap with context.

    An agent doesn’t.

  3. Every transaction,

  4. every report,

  5. every month.

  6. Automation amplifies whatever you point it at.

    A good process gets faster. A bad one gets faster too.

The first question is not: What can we automate? It's: Is this process good enough to automate?

02What comes before agents

  1. 01 · Process

    Written down.

    An agent can only follow a process that exists outside someone's head.

  2. 02 · Ownership

    Every number has a name.

    When an agent produces a result, someone has to be accountable for whether it's right.

  3. 03 · Data

    One definition per thing.

    Agents need a clean chart of accounts, consistent categories and one source of truth.

  4. 04 · Controls

    Separate who prepares from who checks.

    The principle we have always applied to people applies to agents too.

  5. 05 · Agents

    Last, not first.

03What other finance leaders taught me

I expected to hear about different problems. I heard the same ones.

The Golden Gate Bridge at sunset.
  1. 01

    Processes live in people's heads

  2. 02

    Costs don't map cleanly to the business

  3. 03

    Reports take too long to build

  4. 04

    Where should an agent stop and a person start?

04Where a person still belongs

One agent prepares, a different agent checks, and a person approves.

ApprovalsIllustrative
  1. Input
  2. Prepares
  3. Checks
  4. Approves
Agent preparesDrafting

Drafts the invoice from approved hours

Second agent checksException

Compares it with the contract and timesheets

Person approvesYour call

Exceptions only

ApproveRequest changes
Person approvesJudgment

Exceptions only

ApproveRequest changes

  1. One agent prepares.

  2. A different agent checks.

  3. A person approves.

  4. The person in this model is not a rubber stamp.

    They handle what doesn't fit the rules. That is judgment, and it doesn't come from the data alone.

05How the team changes

None of this is about a smaller team. It's about where the team's time goes.

Where Finance time goesIllustrative
PeopleAgents
  • Building invoicesReviewing exceptions
    —Building invoices
  • Assembling payrollApproving it
    —Assembling payroll
  • Booking recurring entriesSupervising agent output
    —Booking recurring entries
  • Reporting what happenedExplaining why, with the business
    —Reporting what happened
  • Fixing data by handDesigning how data is structured
    —Fixing data by hand

What stays human

The future of Finance isn't about removing people from the process. It's about removing the work that keeps them from adding judgment.

Stanford campus fountain and Hoover Tower.

What I'm testing now

I'll share what works, and what doesn't.

  1. 01

    Writing SOPs for recurring processes firsttime tracking, collections and month-end close.

  2. 02

    Mapping every cost to a business line with a named ownerbefore automating anything that touches spend.

  3. 03

    Moving reports from one-off scripts into a shared data layerso they are repeatable before they are automated.

Follow along on LinkedIn (opens in a new tab)

Juan Cruz Pastorini

OnRamp 2026 · San Francisco