↓ Skip to main content

Let AI Do the Routine, Keep People for Judgement

C-level corner · Executive brief

AI can take on much of a team's repetitive work, but only if leaders can trust what it does unattended. The pattern that earns that trust: AI does the routine work, people decide how far it may go and approve the risky steps, every AI change is logged and can be undone, and one switch stops it all. The people doing the toil today become the people supervising it.

The business problem
#

Much knowledge work is routine: fixing the same kinds of rules, sorting alerts that turn out to be nothing, writing up what happened, formatting the deck. It crowds out the work that needs experience.

Security operations is an extreme case. Detection rules are slow to write, break silently when a log format changes, and noisy ones get switched off. Analysts drown in triage. AI could help, but the technical article names the catch: an AI that dismisses alerts can also learn to hide an attack. Hand over work without controls and you trade toil for a blind spot.

What changes
#

The inspiration is a personal prototype security operations centre in which AI writes and repairs detection rules, triages, investigates, writes the reports and requests the response. It runs on synthetic data: not a product, not production-ready and not an official or endorsed solution. Its operating pattern is the transferable part.

  • Autonomy is earned, per type of task. Each kind of change gets one step more freedom only once its approvals, test results and accuracy earn it, and drops back automatically after a mistake. Every week a sample of unattended changes goes to a person: would you have made this change?
  • Every AI change is logged and undoable. One ledger records what the AI changed, why and on what evidence, with an undo button on each.
  • Humans approve the risky steps. Low-risk actions run on their own. Anything that could hurt, such as suspending a user’s account, waits for one human click. Some limits no policy can lift: the AI never deletes data or overrides a rule a person has pinned.
  • One switch stops all automation. Anyone on shift can pause it. Only an administrator can resume.
  • People teach, the machine doesn’t teach itself. Only human decisions train the model that suppresses false alarms, dismissed items re-open when new evidence supports them, and people spot-check what was suppressed.
  • Figures come from the data, words from the AI. In reports, every number is computed from the data and the AI only writes the narrative around it. I’d recommend this for any AI reporting.

The idea applies well beyond security. In office work, someone can ask for “a Q1 sales presentation with five slides” or “a budget tracker” and an AI assistant builds it, so people spend their time on the content, not the formatting. As that grows, I’d apply the same rules: log what the AI changed, require approval before granting autonomy, and keep a person on anything that leaves the building.

What it’s worth
#

  • People’s time back. In the prototype, AI took on nearly all of the routine work around the clock, and the human role shifted from doing the work to supervising it: approvals, exceptions and what is genuinely new. Shown only on synthetic data, so treat it as a direction, not a forecast.
  • Lower risk from automation itself. Undo, human approval for risky steps and a single stop switch mean a mistake is contained and reversible rather than silent.
  • Evidence for auditors. A ledger of every AI action, and figures that come from data rather than the model, make AI-assisted work explainable.

Decisions for leadership
#

  1. Pick one team with a heavy routine load and agree which tasks AI may do, and which always need a person.
  2. Require a log of every AI change, with undo, before any automation goes live.
  3. Set the rule for granting autonomy: per type of task, based on a measured track record, and reversible.
  4. Give whoever is on duty a stop switch, and decide who may turn automation back on.

Questions to ask your team
#

  • If an AI tool changed something overnight, where would we see it, and could we undo it?
  • Who decides when an automated task gets more freedom, and on what evidence?
  • How would we know if an automated filter started hiding something important?
  • Can anyone on shift stop all automation in one step?

Technical background, for your architects

Figures come from the linked articles and their sources; check them against your own situation. Snowflake-related articles explore the art of the possible and are not official Snowflake solutions or endorsed by Snowflake.

Share this article