Follow-up to our conversation

The three agents you asked for, in the order that makes them compound.

Any model can spin up an agent that works. That stopped being the question. The question is which ones compound, and in what order. Each agent's real payoff isn't the task it does, it's the foundation it leaves behind that makes the next one cheaper and safer to build. So the order is the strategy.

An agent is a loop with a stop condition, not a thing you query. Each of these triggers, decides, acts, and stops.


Build them in this order

01

Onboarding agent

Triggers when someone is hired, builds their role-specific ramp plan, stages access for approval, schedules the right intros, tracks milestones, and escalates when something stalls. It runs the ramp, it doesn't answer questions about it.

new hire → plan, stage access, track, escalate → stop when ramped

As CEO I'd greenlight this first

It costs almost nothing to try and it pays back every hiring cycle. Faster ramp means productive sooner. Lower first-90-day attrition means I stop paying to re-hire the same role. It's the lowest-risk way to prove agents work in a regulated environment before I point them at anything sensitive.

Metrics: ramp time, 90-day retention, rehire cost avoided

Leaves behind The first map of the company. Who owns what, which system holds which truth. Every later agent needs that map.
Foundation 33%
02

Credit card and loyalty conversion agent

Triggers on customer and batch events and decides which prescreen offers to fire, suppress, or route, reasoning across the issuer, loyalty, and CRM data, then adapts targeting on the acceptance signal. Offer cleanup and currency reconciliation ride along.

customer event → decide fire / suppress / route → adapt on acceptance

Autonomous loop. Human approves every customer-facing send, within legal's bounds.

As CEO I'd greenlight this next

This moves margin on both sides, and margin is the thing I can't move by pricing in this category. On revenue, the card is one of my highest-margin levers, and right now advisors skip it because the value prop is too confusing to explain at the register. Clean that up and prescreen smarter and acceptance goes up. On cost, smarter real-time targeting cuts mailer spend and manual compliance cycles. I measurably moved acceptance and cut mailer spend on this exact problem, and I'm happy to walk through the specifics in person.

Metrics: card acceptance and activation, mailer cost, reconciliation accuracy

Leaves behind A reconciled credit-and-loyalty identity layer, built on the system map from step one, plus proof the governance pattern holds in a regulated, compliance-heavy setting. Both are what earn the right to point an agent at customers next.
Foundation 66%
03

Beauty-journey agent

Decides the next best action for each member across touchpoints, takes it, and adapts on the response. Not a recommendation waiting to be shown. An agent that moves the member forward, respects their permissions, and learns from what happens.

member signal → decide and act → adapt on response

Autonomous loop. Human holds the gate on anything that reaches a member.

As CEO I'd greenlight this last

I can't change what a category margin is, but I can change how much each member I already have buys, and how often. That's conversion, basket size, repeat rate, and how well I'm using Beauty Insider and the co-brand credit. It's top-line growth from members I already have, not paid acquisition. I put it last because it touches customers, and I only point an agent at customers after the first two have earned the identity and trust controls to do it safely.

Metrics: conversion, average order value, repeat rate, loyalty and credit utilization

Leaves behind The trust and identity layer, proven first on internal onboarding and then in the regulated card program, so customer-facing is the earned last step rather than the risky first one.
Foundation 92%

The fourth one you don't have to invent

Cross-functional intelligence agent

It doesn't wait for a question. Once those three connect, it watches across all of it and surfaces the thing before anyone knew to ask -- when a customer-behavior shift, a loyalty-and-credit signal, and an operations signal converge into something no single team can see, then it routes or acts.

This is the one I reached for in our call. It wasn't wrong, it was early. The foundation has to exist for it to watch.

cross-domain pattern → detect → route or act (no human query)

You already work this way

This isn't my method. It's yours. DealFlow did exactly this:

Phase 1

Started with the most painful bottleneck: getting sales notes into Salesforce. Ran it with a small group, proved it, expanded to hundreds.

Phase 2

Built on that foundation by adding call transcripts to surface buying signals and deal risk. Cheaper because Phase 1 already laid the ground.

Most companies invert it and start with the ambitious agent, then wonder why it dies in production. Painful thing first, prove it, build on the layer.


Where I'd take it next

The agent none of your teams can build alone

Same move as the fourth one above, lifted from Sephora to your business. Support, sales, and finance each leave their own foundation. Once they connect, there's an agent that watches across all three and acts on what no single team can see. No one team owns it. No one team can build it.

Still in progress. You understand a layer by building inside it, not by mapping it from outside. I sketched this from your public material. To actually understand it I'd want inside one real account, because the layer's real shape only shows once it's integrated. That's the part I'm working on now.

The same test on three agents I didn't pick to fit it

It builds what compounds and rejects what doesn't. The one it rejects is the point.

The model can build any of these. What it can't do is decide which ones earn the next. That judgment is the job.

Dispute and chargeback triage agent

build it, after the foundation

CEO view

Bottom line -- cost per dispute, resolution time. Routes each dispute to the right team on first contact instead of bouncing it, which cuts handling cost and speeds resolution.

Leaves behind

A categorized reason-code layer that a later fraud or collections agent can reuse.

Placement rationale

Sits on the system map an earlier agent lays down, then deposits structure the next agent inherits.

Proactive churn-save agent

build it, but last

CEO view

Top line -- retention, repeat rate. Reaches at-risk customers before they lapse, protecting recurring revenue from members you already have.

Leaves behind

Consumes the identity and trust layer rather than building new foundation.

Placement rationale

It acts toward customers, so it is only safe once identity, permissions, and trust are already proven. Powerful, but not first.

Standalone FAQ chatbot

don't build it

Rationale

It isn't really an agent. No loop that learns, no stopping condition tied to an outcome, nothing it leaves behind. Deflects a few questions but moves no real number you would run the company against. Buys fine, builds nothing, does not lower the cost of the next solve. The engine says no, and that no is what proves it discriminates.