Design the agentic workflow that grows your AI product
Capture the signal, turn it into action, keep the customer. The architecture that turns growth spikes into a system that compounds.
The playbook from the session: how a product signal becomes an action, who decides, and where the human gate goes.
Find the leak, then build the loop that closes it
The problem and the gap
Growth spikes, retention leaks, and nobody drew the architecture.
The foundations
Positioning, ICP, and message come before any agent.
What actually breaks
Six failures that are all really one: the data layer.
Automation, workflow, agent
What each one is, and which one your loop actually needs.
The core mechanic
Signal, interpret, action, gate. The loop everything is built from.
Where the leaks are
The signals you already have, and the ones worth acting on.
Your leak, named
Which stage of your funnel is losing the most, and the signal that proves it.
One loop, drawn
A signal, its read, the drafted action, and the gate tier that governs it.
The first agent to build
Not a tool recommendation. The one loop worth building first, and why.
You leave with an architecture you drew yourself, not a list of tools to evaluate.
Growth in spikes. Retention that leaks.
A launch lands, a post takes off, then it flattens, while the customers you already won drift away. A spike is one action. Compounding is a system.
An agent amplifies the decisions you already made
A team spends six weeks evaluating an AI research workflow. Then someone asks how many customer conversations they had that quarter, and the answer is none. The tools are genuinely good. They are not a substitute for the unglamorous work underneath them.
Bedrock, then system, then agent. Automate a vague position and all you scale is the vagueness.
The hard part is never the model. It is the data layer underneath it.
The data layer
No unified definition, no versions, permissions tied to people. Nobody can say which copy is true.
Design for it. Decide the source of truth before an agent reads it.
Rules in people's heads
The real logic lives in veterans, old docs, and legacy workflows, not in any document.
Design for it. Write the rules down as executable logic first.
No accountability boundary
A report missing a column can be re-run. A wrong quote or payment cannot.
Design for it. Map autonomy to reversibility, per action.
The delivery gap
Plenty of newcomers. Few who have carried a first agent to production.
Design for it. Ship one small loop end to end, then widen.
Layer mismatch
A general agent is a browser. Browsers never replaced the CRM.
Design for it. Treat the harness as the entry, not the system.
Buying a model
Models and harnesses get swapped every release. Domain constraints do not.
Design for it. Build governance and workflows that survive a model swap.
Before a business goes from digital to intelligent, it has to decide again what counts as true. Every ambiguity you leave in the data gets amplified the moment an agent acts on it.
Do you even need an agent, and who decides?
Cost, and the surface where it can fail, rise left to right. Use the simplest thing that hits the target.
Automation
A fixed rule with no model in it. Same input, same output, every time. You pick the path, in advance, and it never changes.
Example. Usage drops for 14 days, so the account is flagged and a task lands in the CRM.
Workflowmost growth loops
A fixed path you design, with a model step inside it doing the reading and the writing. You pick the steps, the model handles judgment within one step.
Example. That same flag, but the model reads the account history and drafts the check-in for approval.
Agent
A goal, a set of tools, and the freedom to choose its own path until the goal is met. The model picks the path, at run time. You set the goal and the limits.
Example. Research the account across the site, the CRM, and product usage, then choose which play fits.
Auto
Safe, reversible, read-only. Reads, scoring, enrichment. Let it run.
Notify
Impactful but recoverable. Draft an email, create a task. It acts, tells you, and you can undo it.
Block
Irreversible or high-stakes. Publish, spend, change a CRM stage. It waits for an explicit yes.
Gartner expects more than 40 percent of agentic AI projects to be canceled by 2027, mostly where a workflow would have done. An agent burns roughly four times the tokens of a chat, so cap the spend per run and match the model to the job.
A product signal becomes a growth action, in four moves
Underneath the tooling, every agentic growth workflow is the same four moves: signal, interpret, action, gate. You keep the judgment call. It is a worker, not an oracle.
Signalwhat happened
A fact from your own product or market, with a threshold attached so it fires once and not every day.
- Source: analytics, CRM, support, the web
- Threshold: 14 days idle, not just idle
- Freshness: hours, not weeks
Interpretwhat it means
The read on the fact. Rank it, explain it, and discard the ones that do not carry intent.
- Score the intent and the fit
- State the reasoning in one line
- Kill the false positives here
Actionwhat to do
The drafted move, written to the specific account and the specific behavior. Never a template blast.
- One channel, one clear ask
- Grounded in the actual event
- Ready to send, not sent
Gatewho decides
The reversibility check. Auto if it can be undone, notify if it can be recovered, block if it cannot.
- Auto: reversible, let it run
- Notify: recoverable, tell me after
- Block: irreversible, I decide
Then it closes. What the action produced becomes the next signal. Sent, opened, replied, ignored. The loop learns from its own output.
First, capture the signal. You cannot act on what you do not see.
A signal is only worth capturing if something watches for it and can act. Instrument the few signals that carry real intent, then wire each one to a drafted action and a gate. Timing beats volume.
Behavioral signals
A user stalls at a specific step. Onboarding drop-off, abandoned setup, a repeated failed action.
Usage-based signals
An account approaches a token, seat, or usage-limit threshold.
Adoption signals
A shipped feature shows flat or declining engagement over a defined window.
If you cannot name the exact product event that fires the loop, you do not have a trigger. You have a cron job wearing an agent's clothes.
Route on ambiguity, not on importance
A high-value account does not automatically warrant a frontier model. An ambiguous situation does.
Low ambiguity
A small model or rules engine deploys a standardized growth play instantly.
Low cost, low latency.
High ambiguity
Escalate to a frontier model to parse behavior history and synthesise a bespoke intervention.
Higher cost, reserved for signal that matters.
The action is where the loop earns or loses trust
It runs inside the guardrails without a human in the loop, and it always leaves a trace of what it did and why.
Restructure the UIInterface level
Dynamically surface an unadopted feature where the user already is.
Inject a grounded tipContent level
Hyper-personalized and grounded in live retrieval, not a static template.
Flag the accountEscalation level
Hand Customer Success a brief on the predicted churn vector. The agent decides who acts.
One intent signal, one named account
The signal, the intent, and the call
Three people from one domain compared Product A and Product B on the pricing page this week. The likely intent is a live evaluation with a budget holder involved. That is not a lead, it is an account. It fits tier one, so treat it as one account, one story, across every channel.
Detect
Watch the page and the domain, not the visitor. Same for a checkout or a demo form they never finish.
Resolve the account
Match the domain to the CRM record and the ICP tier. Enrich what is missing.
Read the intent
Which two products, which pages, how many people, how fast.
Gate
You approve the account and the angle before anything is built.
The coordinated build
A page for that account, a sequence off the drop-off, and paid aimed at the domain.
Close the loop
The CRM logs every touch against the account and tells you whether the intent converted. That answer is the next signal.
A mid-tier account crosses 85% of its allotment, nine days before renewal
Cost, privacy, and intelligence
Three dials you set for every agent, and the guardrail tiers behind each. This is the layer the generic content skips, and it is what decides whether your loop survives contact with reality.
Costwhat it can spend
A cheap model for the easy check, a capable one for the reasoning, and a hard cap per run. An agent uses about four times the tokens of a chat, so match the model to the job.
- Standardized play. Small or rules-based model, low fixed cost, stays in tier
- Bespoke intervention. Frontier model, capped per account per month, escalates on ambiguity
- Past the cap. Halt and route to a human. Zero further autonomous spend
Privacywhat it can see
Least privilege: scope it to the task, not the agent. Read-only unless it must write. The agent that touches customer data sees the minimum, and you can audit exactly what it touched.
- Account level. Usage, plan tier, adoption. No PII
- User level. Individual behavior, limited and task-scoped PII
- Restricted. Anonymised signal only, for regulated segments
Intelligencehow much it decides
Autonomy is a spectrum, not a switch. Most first agents should sit low, on a near-fixed path. Start deterministic, and let it earn the autonomy.
- Read-only retrieval. Autonomous
- In-product UI change. Autonomous, within guardrails
- Outbound or account write. Human review, always for account changes
If a row is blank, it is not a guardrail. It is a hope.
The three dials become eight rows you answer before the agent ships. Here is the worked example: a renewal-risk agent, every row filled in.
Per-action cost ceiling
$0.05 standard, $0.60 escalated, capped at $12 per account per month.
Model tier for standard tasks
A rules engine, with a small model for copy assembly only.
Escalation trigger to a costlier tier
A usage-pattern change inside the current billing cycle.
Data scope
Account level: the usage curve, the plan tier, the renewal date.
PII access
None. The Customer Success owner is resolved by account ID.
Actions allowed fully autonomously
An in-product upgrade prompt sized to the usage curve, and retrieval reads.
Actions requiring human review
Any outbound message, and any change to a plan or a limit.
Who reviews escalations
The named Customer Success owner on the account, same business day.
Pick one agent you already run. Write the trigger, the routing rule, and these eight rows. That is the whole exercise.
Research to retention, on one system
Every part of growth is the same signal-to-action loop. Research feeds content, content feeds outbound and paid, both feed the CRM, and retention compounds it. The hub holds it together: you approve every send, and the metrics tell you which loop is working.
One system, six stations, one gatekeeper. Each station is built from the same parts: a shared context file, guardrails, live tool connections, and a narrow sub-agent per job.
Acquisition fills the bucket. Retention is the bucket.
The same signal-to-action loop, pointed at the customers you already won. A kept customer is next month's revenue, next quarter's expansion, and the referral that lowers your cost to win the next one.
Decay shows up weeks before anyone cancels. Each leak has a signal you can instrument today.
First value
They signed and never reached the moment that proves it works. Signal: the setup step never completed.
Engagement
Used once a month is a cancellation with a delay on it. Signal: usage down for 14 days.
Adoption
One seat, one feature. A single active user means no switching cost. Signal: a single active user.
Growth in place
The account outgrew the plan and nobody noticed. Signal: hitting a limit.
Win back
They left for a reason you can now fix, and most people never go back. Signal: the churn reason resolved.
Point the retention agent at your customer list, not just your lead list. Same loop, different question: who is drifting, and what is the one thing that would bring them back?
Draw your first architecture
You pick one signal from your own product and fill five blocks. That sheet is the deliverable: your agentic workflow, drawn, with the first agent to build sitting in the last column.
The loop
Signal, interpret, action, gate. What fires it, what decides, what it drafts, and who approves.
Cost
Which model runs it, the budget, the per-run cap and stopping condition.
Privacy
Exactly what data and tools it can see, read-only unless it must write.
Intelligence
Where it sits on the autonomy line, and which guardrails apply.
Architecture
One orchestrator, a few subagents, and a verifier so nothing grades its own homework.
One signal. One agent. One gate. That is the whole assignment.
Who I am
Grace Man
Agentic AI growth advisor and practitioner
- 1
Creative Director
Hyundai
- 2
Digital Marketing Manager
Pfizer
- 3
Marketing for M365, Surface, and Xbox
Microsoft
- 4
AI Adoption Lead and Data Analyst
Canon
- 5
Founder, agentic AI for marketing, growth, and CRM
AI Strategy League
More free insights, every week
I break down the systems behind agentic growth in the newsletter: the signal loops, the guardrails, and the builds. Free, and straight to your inbox.