At Sapphire 2026, SAP unveiled the Autonomous Enterprise: a unified Business AI Platform with 50+ domain-specific Joule Assistants orchestrating 200+ specialized agents across finance, supply chain, procurement, HR, and customer experience. The pitch is not AI that helps your team draft a quote, or summarize a meeting, or write a follow-up email. It is AI that runs the workflow end-to-end. SAP's Autonomous Close Assistant compresses the financial close from weeks to days — and the human's role moves from doing the close to auditing it.
You are not buying SAP. A 100-person digital services business never will. But the architectural shift behind the announcement travels — and most SMBs have not named it yet.
The shift: from "AI drafts, human sends" to "AI acts, human audits"
Almost every AI workflow shipped to SMBs in the last 18 months sits in draft mode. The AI prepares, the human approves, the work goes out. Sales follow-ups, support replies, meeting notes, content drafts — all draft. It works. It is safe. And it is a ceiling.
In act mode, the AI completes the workflow itself, in the customer-facing system, against real data, with a human-readable audit trail. The human's job moves from approver-of-each-event to auditor-of-the-stream — different cadence, different competency, different role description.
SAP's bet is that the act-mode ceiling, for the workflows that genuinely qualify, is an order of magnitude higher than the draft-mode ceiling. They are right. The question for SMBs is not whether to make the shift — it is which 2 or 3 workflows are ready for it, and which ones must stay in draft forever.
Steal #1 — Re-classify every workflow as "draft" or "act"
Most teams have not. The mode is implicit, set by whoever built the integration. Once you name it, you can budget for it.
Example A — Quote-to-invoice in a 100-person services business. Today: the rep drafts the quote in CRM, the manager reviews it on Slack, the customer signs the PDF, ops creates the invoice in QuickBooks. Every step is "draft" — the AI (if it touches anything) is summarizing the discovery call or auto-filling a template. Maybe it shaves 20 minutes per quote.
In act mode: the AI generates the quote directly from CRM data and the rep's discovery notes, sends it with the rep's e-signature on file for deals under the rep's authority threshold, and creates the QuickBooks invoice on the signature webhook. The rep audits the day's pipeline in 10 minutes instead of authoring it for three hours. The ops person stops being the integration glue and starts being the exception-handler.
Notice what is different. The quote is not better in act mode. The cycle time collapses, the headcount math changes, and the rep's job becomes catching the 5% the system gets wrong rather than producing the 95% it would have gotten right anyway.
Steal #2 — The audit log becomes the human's job, not the doer's
The instinct when teams hear "AI acts" is to add more approval gates. That recreates draft mode with extra steps. The right move is to redesign the human's job around the audit, not the action.
Example B — Support ticket triage and first response. In draft mode, the AI suggests a reply; the support rep edits and sends. Quality stays high because the human is in the loop on every ticket. So does the cost — you cannot scale past your support headcount.
In act mode, the AI sends acknowledgement plus first-line answers for resolvable categories (password reset, refund status, shipment tracking, return policy lookup) and routes the rest. The support lead does not review every ticket — they audit 20 random tickets each Friday, plus 100% of any ticket the customer replied "this didn't help" to. The role description for "support lead" just changed: from production-line approver to sampling-based auditor with escalation authority.
The audit cadence is the design choice that makes or breaks the shift. Too rare and bad patterns ship for weeks before anyone notices; too frequent and you have reinvented draft mode. Weekly sampling plus 100% review of negative-signal events is a good starting point for most categories. Higher-stakes categories — anything with a money or compliance dimension — move to daily or per-event.
When NOT to make the shift
Plenty of workflows must stay in draft mode forever. Anything where the cost of a wrong answer is irreversible — contract terms changed, sensitive customer email sent, regulatory filing logged — needs a hard human gate, full stop. Anything where the work depends on judgment that is not documented anywhere stays in draft for the obvious reason: the AI cannot act on a rule that does not exist.
The shelf list — workflows you decided not to promote to act mode — is more valuable than the list of ones you did. Write it down with the reason. Six months from now the reason may have expired (better model, better tooling, new audit pattern) and the workflow can be revisited.
Ignore the breadth
SAP wins with 50 shallow assistants and 200 agents because it owns the suite. Every Joule Assistant lives inside data SAP already manages, with permissions and audit trails SAP already wires up. The breadth is the moat.
A 100-person business has no such moat. You win the act-mode shift with 2 or 3 deep agents that own a workflow end-to-end inside the systems you already use — and a shelf list of 10 you decided not to build yet, with reasons. The 50-agent strategy is not available to you, and it is not what you need.
The rule
If a human still has to read the output before it lands, it is not an agent. It is a faster intern.
Move two workflows out of draft mode this quarter. Document the audit cadence. Update the role descriptions. That is the shift, scaled to your business — not theirs.