AI Agentic Workflows
September 22, 2026 · 6 min read · by Ananda Narasimhan
A recent Deloitte study of over a thousand B2B suppliers and buyers found that 45% of sales teams say they use AI. Only 24% have deployed anything genuinely agentic — software that plans a sequence of actions and executes it without a human approving each step. The gap between those two numbers is where most of the risk lives, because the teams in the 45% assume they're further along than they are.
"AI in sales" usually means a chatbot drafting emails or a copilot summarizing a call. Agentic means the system reads a record, decides what to do next, and writes back to your CRM on its own. That last part — the write — is where governance actually starts mattering, and it's the part most rollouts skip straight past.
Most teams either lock an agent out of the CRM entirely, which makes it a glorified autocomplete, or give it a service account with the same permissions as a full admin, which makes it a liability the first time it misreads a field. Neither is a decision — both are defaults nobody examined. The fix is a tiered model, the same one you'd use for a new hire: read access on day one, propose-and-approve once it's shown it understands the data, autonomous write only for the narrow set of actions where a mistake is cheap to catch and cheap to reverse.
We build this into every agentic workflow we ship, because the alternative — shipping an agent and hoping — is how you end up explaining to a VP of Sales why forty opportunities got auto-closed on a Tuesday.
An agent's first job is answering questions from CRM data — account history, deal stage, last activity — without touching anything. This tier looks boring, which is exactly why teams skip it. It shouldn't be skipped, because it's the only tier where you can watch the agent's reasoning against records you already trust and catch a bad assumption before it costs you anything. If an agent misreads your lifecycle stage definitions at read-only, that's a fifteen-minute fix. If it misreads them with write access, that's a data cleanup project — the kind we cover in our CRM hygiene audit.
This is where most durable agent workflows should live for the first several months, and where a lot of them should stay indefinitely. The agent drafts the next action — a task, a note, a field update, an email — and a human approves or edits it before it goes live. You get the speed of automation with a checkpoint that catches the failure modes agents actually have: hallucinated context, stale data, or a correct answer applied to the wrong record.
The approval step isn't overhead you remove once trust builds. It's a measurement tool. Track the approval rate by action type, and you get a real answer to "is this agent ready for more autonomy" instead of a gut feeling. An 95%+ approval rate on a specific action, sustained over a few hundred instances, is the actual bar — not a demo that went well once.
Full autonomy belongs to actions that are high-volume, low-stakes individually, and easy to audit in bulk: logging an activity, updating a lead score, enriching a contact record, routing a new lead to the right owner. Not: closing an opportunity, changing an owner on an existing deal, deleting anything, or sending an email to a customer without a review step. The line isn't about how smart the agent is — it's about how expensive a mistake is and how fast you'd notice it.
This is the same logic behind where AI actually earns its keep in sales workflows: the highest-leverage automation isn't the flashiest one, it's the one operating on the highest volume of low-risk, repetitive decisions that were eating a rep's day.
Before promoting any action from proposed to autonomous, we check four things: the action is reversible or cheap to reverse; the failure mode is visible fast, not something that surfaces in next quarter's forecast; the data it depends on is clean enough that a wrong output means the agent erred, not that the input was garbage; and there's an owner who actually reviews the audit log weekly, not a dashboard nobody opens. Skip any one of the four and you haven't built autonomy — you've built an unmonitored process that happens to run on AI.
None of this requires an elaborate platform. It requires deciding, action by action, which tier it's in — and writing that decision down somewhere your team can see it, inside your revenue operations function, before an agent starts touching records nobody's watching.