Suhrab Khan recently shared something that caught my attention: "You don’t need more AI. You need less chaos." He followed it with a line that is both funny and painfully accurate: "Most 'automation strategies' are just vibes with a Zapier login."

I have seen the exact scenario he described. A team asks for an AI agent to qualify leads, write outreach, book meetings, and update the CRM. The request sounds modern and ambitious, but the foundation is often a patchwork of tabs, handoffs, and unwritten rules.

Suhrab’s point is not anti-AI. It is pro-clarity. If you automate a messy process, you do not get a clean outcome faster. You get mess at scale, delivered with confidence.

The real failure mode: hidden decisions

Suhrab Khan explained that teams do not "fail at AI" so much as they fail at making decisions explicit. That framing matters.

A workflow is not just steps. It is judgments:

What counts as a qualified lead?
When should a rep follow up?
Which industries are in scope?
When do we route to AE vs keep with SDR?
What exceptions override the general rule?

Humans handle ambiguity by improvising. They ask a colleague. They remember what worked last quarter. They rely on institutional memory. AI agents do not magically resolve ambiguity. They will either (a) make a guess, or (b) ask for more structure. If you skip the structuring work, the agent will "freestyle" as Suhrab put it.

"If you can’t explain the workflow to a new hire, an AI agent will freestyle."

That is the heart of it. New hires and AI agents both expose what the team has not agreed on.

What "automation chaos" looks like in the wild

In Suhrab’s story, the revenue team had no clear process to hand over. When they opened the tabs, they found multiple tools, multiple handoffs, and multiple definitions of success:

Handoffs across Slack, HubSpot, Notion, and spreadsheets
A doomed document called "FINAL_final_v3"
"Qualified lead" defined four different ways
Notes fields full of unstructured, inconsistent writing
Two reps carrying critical exceptions in their heads since 2022

None of that is rare. Most organizations accumulate workflow debt the same way they accumulate tech debt: one workaround at a time.

Here are a few signals you have workflow debt (and AI will amplify it):

1) Your system depends on memory

If a step works only because a specific person remembers the rule, you do not have a process. You have a hero.

2) Your definitions drift by channel

If Slack says one thing, the CRM says another, and the spreadsheet has a third version, your data will not line up and automation will not know which truth to follow.

3) You have "soft" handoffs

A handoff like "I DM you when it is ready" is not a handoff. It is a risk.

4) Your exceptions are the real workflow

If 30 percent of cases require special handling, your "main" process is just a marketing diagram.

This is why AI theater happens, as Suhrab described: demos look great, production bleeds, and leadership blames the model. The model is often doing exactly what you asked, on top of a process you never clarified.

The playbook Suhrab outlined (and why it works)

Suhrab Khan shared three simple build steps. They sound basic, but they are exactly the work most teams try to skip.

1) Map one loop end to end (on one page)

Suhrab’s instruction was: "Trigger → inputs → decision → output → owner."

This is more powerful than a long process document because it forces completeness.

Trigger: What event starts the loop (new inbound lead, form submit, webinar attendee, list upload)?
Inputs: What data is required to decide (company size, role, intent signals, region)?
Decision: The explicit rule or rubric (if-then, scoring, routing thresholds).
Output: What the system produces (status change, task, email sequence, meeting link).
Owner: Who is responsible for the outcome (and who is accountable for exceptions).

If you cannot fit a workflow on one page, it is usually a sign that you are trying to automate a bundle of workflows at once.

2) Define edge cases in plain language until it gets boring

This is the unsexy part that makes everything else work.

Write the rules like you are teaching someone on day one:

If the lead is a student, do Y.
If the company is a competitor, do Z.
If the title includes "consultant," route to a different sequence.
If the region is unsupported, disqualify with a polite email.

Do it until you run out of surprises. The boredom is the point: it means your team is converting tribal knowledge into shared knowledge.

3) Instrument the automation: logs, fallbacks, human review, kill switch

Suhrab’s third step is what separates responsible automation from reckless automation.

At minimum, instrumenting means:

Logs: What decision was made and why (inputs, rule fired, output generated).
Fallbacks: What happens when required data is missing or conflicting.
Human review: A queue for uncertain cases, plus sampling for quality control.
Kill switch: One click to pause the system if something goes wrong.

This is not paranoia. It is operational hygiene. If an agent starts misrouting leads or spamming prospects, you need immediate containment, not a retroactive postmortem.

A concrete example: "Agentifying" SDR qualification the right way

Let’s take the exact request Suhrab received: qualify leads, write outreach, book meetings, update the CRM.

A clean, agent-ready version might look like this:

Trigger

New inbound lead created in HubSpot.

Inputs

Email domain, company name, website, role, country, form answers, lead source, prior activity.

Decision (explicit)

If country not in supported regions, mark "Disqualified - region" and send template A.
If role is student or job seeker, mark "Disqualified - not ICP" and send template B.
If company size is 50-500 and role contains "Director" or above, mark "Qualified - high" and create task to book meeting.
If missing company size, enrich first. If enrichment fails, route to human review.

Output

CRM updated with status, reason code, and next step
Outreach email drafted with approved claims and links
Meeting scheduling link sent only when qualified criteria met

Owner

SDR manager owns routing rules; RevOps owns data inputs and enrichment reliability.

Now an AI agent can help in the right places: summarizing research, drafting compliant outreach, proposing next steps, and handling enrichment. But the team is not asking the agent to invent policy.

A simple checklist for your next AI automation

Suhrab Khan ended with a challenge: pick one workflow you want to AI automate this quarter, then find where it depends on memory, a DM, or a spreadsheet nobody admits owning.

Here is a quick way to do that:

List every handoff and tool touch in the workflow.
Highlight every step that contains a judgment call.
For each judgment, write the rule in plain English.
Collect edge cases from the last 30 days.
Add logging and a kill switch before you add more autonomy.

If you do only one thing, do this: write down your definition of "done" for the workflow and make sure everyone agrees. AI cannot rescue disagreement.

Closing thought

Suhrab’s post is a reminder that AI is a multiplier. It multiplies whatever you already have: clarity or chaos. If you want AI agents that actually help your revenue team, start by reducing ambiguity, documenting decisions, and building instrumentation that keeps you safe.

This blog post expands on a viral LinkedIn post by Suhrab Khan. View the original LinkedIn post →