The Startup Ideas Podcast

Don't apply AI.

The exact workflows behind
the $1M-a-year AI job.

Vas MozaCEO · Varick Agents

This is already happening.

Funds buy accounting firms, IT shops and law practices, put engineers inside, and re-engineer the work. One of many ways to get value from AI. The one with the most money behind it right now.

1Buy a firm that runs on people.An accounting firm, an IT shop, a law practice. Manual work, thin margins, no software of its own.

2Put engineers inside it.Find the real process. Rebuild it with agents in the systems the firm already uses. Thrive has 35 engineers across 70 firms.

3The numbers move.Tax returns 30% faster at 98% accuracy. Half of help desk tickets closed by agents. Gross margin at one call-centre firm: 60% and above.

4The firm is worth more.A services firm is valued on its profit. Double the margin at the same multiple and the firm is worth about twice as much. Keep it, or sell it.

Thrive Holdings · Josh Kushner70+ businesses. $2B raised at a $12B valuation. OpenAI took a stake in December 2025.Its accounting arm, Current, is 50 firms and 2,000 professionals.

General Catalyst · AI-enabled roll-upsBillions committed to buying service firms and rebuilding them.Legal, IT services, call centres, property management. 25 to 30% productivity gains in HOA management.

The people doing steps 2 and 3Forward deployed engineers.Same job whether the owner is a fund, a public company, or the founder. This video is how it is done.

Sources: TechCrunch, 14 August 2026 · The Next Web, 13 August 2026 · General Catalyst, The Future of Services. Step 4 is arithmetic, not a specific deal.

Your life is already complicated.

Seventeen services. Each one used a different way. None of it written down.

5inboxes

Gmailreceipts, family

Gmailclients, same day

Outlookold job, unread

iCloud MailApple receipts

Yahoonewsletters

4file stores

Google Driveclient docs

Notionnotes, unsorted

iCloud Drivephotos, scans

Desktopfinal_v3.pdf

3streaming

Netflixyour login

Prime Videopartner's login

Spotifyfamily plan

5messaging apps

iMessagefamily, two lines

WhatsAppfriends abroad

Slack2 workspaces

Google Calendarpersonal

Outlook Calendarwork

The hard part

Which app is easy. How you use each one is the work.

ifa client writes→Gmail work, reply today iffamily writes→iMessage, two lines ifit is a receipt→Gmail personal, file it ifit is a newsletter→Yahoo, never read ifit is a new document→Drive, house template ifit is Friday night→Netflix, your login

Now imagine a company.

Eight acquisitions in five regions. Each one kept its own systems and answers to its own laws.

1

2

3

4

5

6

7

8

companies acquired

5

regions, five sets of laws

23

systems of record

1

Chicagoacquired 2016 · US GAAP, state tax

2

Torontoacquired 2021 · CRA, provincial tax

3

São Pauloacquired 2019 · e-invoicing mandates

4

Londonacquired 2023 · UK GAAP, VAT

5

Frankfurtacquired 2018 · GDPR, IFRS

6

Bangaloreacquired 2020 · data residency, GST

7

Singaporeacquired 2022 · PDPA, GST

8

Sydneyacquired 2017 · Privacy Act, payroll rules

ERPCRMHR

Understand a company's process.

Interviews get tribal knowledge, and process mining agents get real data and documentation.

1 · Interviews

The human side.

Controller

AP clerk

Deal desk analyst

Sales rep

why a rule existswho really decideswhich step is theatrethe workaroundthe unwritten threshold

What they give youThe why behind every number in the log. Ten to twenty people, in person, showing you.

2a · Mining the systems of record

What the systems know.

throughputvolume per stepcycle and wait timeerror rateexception shareentity map and data types

What the agents give youEvery case, every timestamp, every hand-off. All of the volume, none of the memory.

2b · Mining everything else

What lives outside them.

SOPs and policiesthe spreadsheet on the deskapproval threadsemail chasesexceptions settled in chatthe exception list

What the agents give youThe rules nobody typed into a system, and where the documented process drifted.

Interviews give the why. Mining gives the what. Together they give the map: every step, its volume, its wait, its owner.

Example 1. A public software company.

Sales operations, the deal desk. The document says 7 steps. The CRM log says 20.

Drawn from a real engagement · names and figures changed

Revenue

~$5B

global B2B software

Products

~150

in 3 segments

Solution consultants

~175

plus the deal desk

Brought in by

CRO

the Chief Revenue Officer

What made it hard

Public

Every discount lands in a reported quarter.

What the process document says · 7 steps

1 Build quote

2 Submit

3 Deal desk

4 Approve

5 Send

6 Negotiate

7 Sign

What the CRM log shows · 20 steps · 7 loops back, the three largest drawn

Terms review sends it back to the rep · 61% of requests

Legal sends it back · 12%

A new quote repeats approval · 30%

1 Build quote

Ask RevOps

Rebuild

2 Submit

3 Deal desk

Terms review

Back to rep

Resubmit

Deal desk

Finance

Legal

Security

4 Approve

5 Send

6 Negotiate

New quote

Resubmit

Approve

Send

7 Sign

In the documentNot in the documentA documented step done againA loop back

documented → actual

7 → 20

Sort every step. Delete, code, agent, human.

Twenty logged steps, one bucket each. Eight are deleted, not automated. Agents prepare every review. A person approves, negotiates, and signs.

Delete

8 steps · rework, and a review run twice

Plain code

4 steps · rules, no judgement

5 agents

prepare each review, then stop

3 human decisions

approve, negotiate, sign

The agent works the CRM. A person decides.

A discount request arrives. The agent assembles policy, floor, and precedent, computes the band, and stops.

◆ ARDEN GROUP · PRODIllustrative reconstruction by Varick Agents · simulated data · no client systems connected · not a reproduction of a Salesforce application

🔍 Search...

+

?

⚙

🔔4

VA

SalesHomeOpportunities ▼Quotes ▼Accounts ▼Contacts ▼Reports ▼

📜

Quotes

Pending approval ▼

NewSubmit for ApprovalChange Owner▼

7 items · sorted by Quote Number · updated a few seconds ago

🔍 Search this list...

⚙ ▼⟳✎

	Quote Number ▼	Account Name ▼	Segment	Term	Grand Total	Discount	Approval Tier	Status
1	QTE‑4468	Northwind Foods	Mid‑market	12 mo	184,000	10%	Tier 1	Approved	▼
2	QTE‑4469	Contoso Health	Enterprise	36 mo	96,500	12%	Tier 1	Approved	▼
3	QTE‑4470	Fabrikam Logistics	Enterprise	24 mo	310,000	15%	Tier 1	Approved	▼
4	QTE‑4471	Acme Industrial	Enterprise	36 mo	412,000	22%	Tier 3	Pending	▼
5	QTE‑4472	Tailspin Media	Commercial	12 mo	58,000	8%	Tier 1	Approved	▼
6	QTE‑4473	Woodgrove Bank	Enterprise	36 mo	1,240,000	14%	Tier 2	In Review	▼
7	QTE‑4474	Litware Systems	Commercial	12 mo	77,200	5%	Tier 1	Approved	▼

Arden Group✎

💬 Threads

🕔 Activity

▼ Channels

# general

# sales-floor

# deal-desk1

# pricing

# legal-review

▼ Direct messages

Tom Beckett

Priya Raman

Varick Agent

# deal-desk☆Illustrative reconstruction · simulated data · not a reproduction of a Slack application👥 18

Today

TB

Tom Beckett8:52 AM

Quarter‑end is Friday. Anything still open by Thursday noon goes on the commit call.

◆

Varick AgentAPP9:08 AM

@Rachel Okafor, one request is outside policy and it is yours to call.

QuoteQTE‑4471 · Acme Industrial · Enterprise

The ask22% off list · 36 months · Net 60 · $412,000

Outside policy7 points above the Enterprise cap of 15% · policy v4.2

Everything else2nd quartile of 9 comparable approvals · median 18% at Net 45 · paid on time 11 of 12 invoices

📎 Approval case · QTE‑4471.pdf · 3 pages

◆Varick Agent: On it. Rebuilding at 18% on a 36‑month term with Net 45, and diffing the order form. Your conditions are on the quote.

RO

Rachel Okafor9:11 AM

Approve with conditions. 18% at Net 45, 36 months. Send it back to the rep.

Message #deal-desk

+Aa🙂@►

Cleared on policy: 12% before, 68% after · days a deal sat idle: 23 before, 6 after · the agent stops, a person grants it

Example 2. Nine companies, one private equity owner.

The sponsor brought us in and opened every door. Twenty-six portfolio companies. Two engagements: five on NetSuite, then four on Dynamics.

Drawn from a real engagement · names and figures changed

The sponsor

Mid-market fund · 26 portfolio companies · 7 sectors · operating partner

Co. 1

Software

“We’re different”

Co. 2

Software

“We’re different”

Co. 3

Software

“We’re different”

Co. 4

Software

“We’re different”

Co. 5

Software

“We’re different”

Co. 6

Services

“We’re different”

Co. 7

Services

“We’re different”

Co. 8

Services

“We’re different”

Co. 9

Services

“We’re different”

Co. 10

Services

“We’re different”

Co. 11

Services

“We’re different”

Co. 12

Health

“We’re different”

Co. 13

Health

“We’re different”

Co. 14

Health

“We’re different”

Co. 15

Health

“We’re different”

Co. 16

Industrial

“We’re different”

Co. 17

Industrial

“We’re different”

Co. 18

Industrial

“We’re different”

Co. 19

Industrial

“We’re different”

Co. 20

Consumer

“We’re different”

Co. 21

Consumer

“We’re different”

Co. 22

Consumer

“We’re different”

Co. 23

Logistics

“We’re different”

Co. 24

Logistics

“We’re different”

Co. 25

Insurance

“We’re different”

Co. 26

Insurance

“We’re different”

Grouped by sector · ERP on each tileEngagement 1 · 5 on NetSuiteEngagement 2 · 4 on Dynamics

Why these nine

One ERP per engagement.Five run NetSuite. Four run Dynamics. One integration each, built once, deployed to every company on it.

One problem.Accounts payable, receivables, and the month-end close. Excel files keyed in by hand, in all nine.

One set of politics.The sponsor introduces you to each CFO. Top-down buy-in from day one, nine times over.

One close. Five ways to run it.

Month-end close in the five NetSuite companies, mapped from their own systems. Same start, same end, five paths. Red steps are chasing or waiting.

Co. 1Excel, offshore team
1Chase statements
2Download PDFs
3Normalize workbook
4Pull ledger
5Match by hand
6Flag mismatches
7Email controller
8Wait for reply
9Chase reply
10Post adjustment
11Roll up region
12Close
12steps
Co. 2Inside the ERP
1Import statements
2Auto-match
3Review unmatched
4Email controller
5Wait for reply
6Post adjustment
7Intercompany recon
8Sign off
9Close
9steps
Co. 3ERP, sheets, a shared mailbox
1Check mailbox
2Save PDFs
3Key into sheet
4Pull ledger
5Match by hand
6Flag mismatches
7Email controller
8Wait for reply
9Chase reply
10Re-key adjustments
11Post
12Roll up
13Second review
14Sign off
15Close
15steps
Co. 4NetSuite plus the old ledger
1Chase statements
2Download PDFs
3Key into workbook
4Pull old ledger
5Pull NetSuite
6Map accounts
7Match by hand
8Flag mismatches
9Email controller
10Wait for reply
11Chase reply
12Post adjustment
13Intercompany recon
14Re-key adjustments
15Roll up
16Second review
17Sign off
18Close
18steps
Co. 5ERP and a shared mailbox
1Check mailbox
2Import statements
3Auto-match
4Review unmatched
5Key exceptions
6Email controller
7Wait for reply
8Post adjustment
9Intercompany recon
10Roll up
11Second review
12Sign off
13Close
13steps
All fiveOne definition of the close
1Statements arrive on a feed
2Rules match
3Agent assembles exceptions
4Person works the queue
5Post
6Close
6steps

17 steps became 7. 6 loops became 1.

Accounts payable, mapped from the transaction log. Same policy, same approvers, full audit trail.

Before · 17 steps · 6 owners · 6 exception loops

Owner ↓ · Phase →

A Receive

B Classify

C Record

D Approve

E Prepare payment

F Pay & close

1Suppliersends the invoice

2Invoice processing teamkeys every invoice

3Finance systemthe ledger

4Budget holdersapprove their own lines

5Finance leadershipreleases the payment run

6Payments desksends the money

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

After · 7 steps · 4 agents · 3 human decisions

Owner ↓ · Phase →

A Receive & check

B–C Classify & record

D Approve

E Prepare payment

F Pay & close

1Suppliersends the invoice

2Agentsfour of them, 24/7

3Finance systemthe ledger

4Approversdecisions only

5Payments deska person sends the money

◆Agent

Process steps

17→7

Cycle time

24 days→6 days

Exception loops

6→1

Straight-through

18%→87%

Cost per invoice

$31→$6

Example 3. A 60-person accounting firm.

Monthly books for 400 small clients. The engagement letter says 6 steps. The practice log says 14.

Composite · a typical firm · figures illustrative

Revenue

~$12M

4 partners

Clients

~400

monthly books

Systems

4

Xero, QuickBooks, Excel

Brought in by

Owner

the managing partner

What made it hard

400 ways

Every client sends its books a different way.

What the engagement letter says · 6 steps

1 Collect

2 Categorise

3 Reconcile

4 Partner review

5 Deliver

6 Bill

What the practice log shows · 14 steps · the two largest loops drawn

Client chased for documents · 70% of months

Partner sends it back · 35%

1 Collect

Chase

Re-ask

2 Categorise

Ask client

Redo

3 Reconcile

Flag

4 Review

Send back

Fix

Review again

5 Deliver

6 Bill

In the letterNot in the letterA step done againA loop back

documented → actual

6 → 14

Same four buckets. Smaller firm, same method.

Fourteen logged steps. Five are deleted. Rules do the books. Agents do the chasing. A partner still signs, and a person still talks to the client.

Delete

5 steps · rework and repeats

3Re-ask

6Redo

10Send back

11Fix

12Review again

Plain code

4 steps · bank feeds and rules

1Collect

4Categorise

7Reconcile

14Bill

3 agents

chase, ask, assemble the exception

2Chase

5Ask client

8Flag

2 human decisions

the partner signs, a person delivers

9Review

13Deliver

A $12M firm and a $5B company. Same method, same four buckets.

Three people in one.

That is why the job pays. All three examples needed all three skills, and very few people hold all three.

1 · The operatorUnderstands how the work actually runs.Finance, sales, operations, legal, support. Knows which step is theatre and who really decides. Consultancies have this person.

2 · The engineerShips production software into the systems of record.Salesforce, SAP, NetSuite, Dynamics, Workday. Integrations, permissions, rollback. Software firms have this person.

3 · The AI engineerBuilds evals and a model-agnostic harness.Knows what a model can be trusted with. AI labs have this person.

Almost nobody has all three in one person.

Do this next week.

One thing a day, on your own life. Before any AI is involved. Friday, do it to a business.

MondayList every app that holds your stuff.Five inboxes, four file stores, three logins for one restaurant app. Write which one wins when two disagree.

TuesdayPull 20 things you did last week.Paying a bill, booking a table, filing a receipt. Count how many went the same way twice. Almost none will.

WednesdayWrite one process down, step by step.How you pay a bill, from the email to the bank. Every click, every tab, every wait.

ThursdaySort every step.Delete it, give it to code, give it to a model, or keep it yourself. In that order. Most people start at the model.

FridayDo Monday again, on a business you know.A dentist, a law firm, the shop on the corner. Time the work against the waiting. Then build the smallest thing that removes one step.

So don't apply AI.

Six steps. They work on a dentist, a law firm, or a public company.

01

Find the real process.Interviews for the why. Agents on the systems of record and everything else for the what.

02

Measure elapsed time against touch time.The waiting is the process. The work is minutes.

03

Define one process with one owner.Somebody who can sign it across every department. Nothing about this step is AI.

04

Sort every step.Delete, plain code, an agent, a human. In that order.

05

Baseline before you build.Write the number on the wall. Same policy, same approvers, full audit trail.

06

Build once. Deploy everywhere. Measure again.Twelve months later, that is the only proof it worked.

Sitevarickagents.com Weekly write-upvarickagents.com/newsletter Thank you.