The short answer
Start with a workflow that is frequent, slow by hand, low in variation, spread across a few systems, cheap to get wrong and rarely off the happy path. Score candidates on those six factors, and pick the highest score that clears four gates.
Four gates first
A candidate that fails any gate waits, however well it scores.
The systems it touches won't be replaced in the next six months.
A named owner will review output during supervised runs.
Automated access is allowed by the software license and portal terms.
The six factors
Score each from 1 to 5. Higher always means a better first pick.
Score = 20 x the weighted sum, so the maximum is 100. 70 or more: first wave. 50 to 69: split or redesign first. Under 50: not yet.
Why these weights
Volume and handling time together are 40% because they set the hours at stake. Risk gets 20% on its own because the first project shapes everyone's opinion of agents, and an incident in week three can end the program.
Systems peak at two or three. One system often means the ERP can already do the job. Eight means a long chain with many places to fail.
Three examples
Payment settlement reconciliation (property management): 72, first wave
Daily, four clear paths, internal and reversible, few exceptions.
W-9 collection (construction): 63, split first
The agent requests, chases, reads and enters clean forms. Name and TIN mismatches go to a person.
Credit hold review (manufacturing): 49, not yet
Nine-plus paths, high judgment, and releasing a hold ships goods to someone who may not pay.
Split before you score
Credit hold review fails as a whole and passes as a half. Gathering the evidence for each held order (aging, payments in transit, open orders, credit limit, dispute notes) takes most of the time and changes nothing in the ERP. That half scores 72. The credit manager keeps the decision.
Almost every process has an evidence half and a decision half. Scoring the whole thing averages them and hides a strong first candidate inside a weak one.
Mistakes that skew the score
Scoring the loudest complaint instead of the biggest measured cost
Using catalog hour estimates as your own numbers (time 20 real items instead)
Forgetting cadence: a monthly process needs two cycles to validate
Treating more systems as pure difficulty (re-keying between them is where the hours are)
Ignoring who reviews
Seed your list
LaunchAI's Process Atlas catalogs 1,554 back-office processes across construction, manufacturing and property management, with typical systems, roles and planning-estimate hours. Use it to build your candidate list, then score with your own counts. Process Atlas →
- 458
- construction
- 501
- manufacturing
- 595
- property management
Start with the good half
In LaunchAI Studio every workflow is built from small atomic steps, so splitting is literal: evidence steps run on their own while the decision stays a review step with the packet on screen. Product → Then price it. ROI of AI Agents →
Questions
Should the first workflow be the biggest?
Only if it clears the gates and scores well on risk and exceptions. The first project's job is to run cleanly and prove the method. The biggest can go second.
How many workflows in a first pilot?
One, or two that share a system and a reviewer. A pilot is judged on whether its results are clear.
Keep learning
- Agent-Ready SOP Put one action on each line. Read the guide
- ROI of AI Agents ROI is the cost of the hours an agent gives back, minus the cost of the exceptions, reviews, upkeep and oversight it creates, minus software and model costs, set against one-time build and parallel-run costs. Read the guide
- Human in the Loop Put a person where a wrong action is expensive and hard to undo, and where the agent's evidence is weak. Read the guide
No prompts. No babysitting. No Integrations.