AI agents: Operations and quality

What it does: Every morning — ready-to-approve actions.

What a person decides: Approval.

For whom: Owner.

How to measure: Actions approved and completed.

Example result: “Action: approve 3 steps for today. 1) Request #214 ‘Fix the AC in the hall’ is a day overdue with no one assigned — assign a technician before lunch. 2) The alert ‘warehouse fridge: temperature above normal’ has been open for 2 hours — check it now. 3) 5 tasks for opening the second location are due Friday — move the deadline or add people. Impact: we tackle what would hurt customers and cash the most. Risk: step 3 needs the manager's decision, or Friday falls through. Fallback: any step can be left unapproved — nothing changes.”

What it does: Checklists, photo reports, deviations.

What a person decides: The review.

For whom: Staff member, manager.

How to measure: Deviations from the standard.

Example result: “Score: 9 of 10. Why: 9 of 10 items on the opening checklist are done, 8 with photos; no photo of the display (item 4), which the standard requires. Last Monday had the same issue. What to do: send a display photo by 10:00 and remind the shift about item 4.”

What it does: Assigns jobs, sets visit windows and notifies the customer.

What a person decides: The assignment.

For whom: Dispatcher.

How to measure: Visits within the window; travel time.

Example result: “Action: give the Samal-2 job (‘boiler not heating,’ window 2–4 pm) to Arman: per the schedule he finishes in Samal-1 at 1:30 pm, 10 minutes away, and he carries boiler parts. Customer gets a message with the 2:00–4:00 pm window after your confirmation. Impact: the technician arrives without rushing, and the customer gets an exact window. Risk: if the Samal-1 job runs long, Arman will be late — ask him to check in at 1:30. Fallback: if he doesn't check in, give the job to Yerlan, who's free from 3 pm.”

What it does: Tracks deadlines and warns in advance.

What a person decides: The final decision on the outcome.

For whom: Manager.

How to measure: Overdue tasks.

What it does: Writes and updates SOPs based on actual practice.

What a person decides: Approval.

For whom: Manager.

How to measure: Staff questions about processes.

Example result: “SOP ‘Product returns’ — 1 page. 1. When allowed: time limit and product condition — per store policy. 2. Steps: accept the product and receipt → check condition → process the return at the register → refund via the original payment method. 3. Timing: decision the same day, money per the bank's timeline. 4. Who decides: the cashier within policy, the shift lead for disputed cases. 5. Exceptions: no receipt, signs of use. To decide: the amount a cashier can refund on their own.”

What it does: Causes, takeaways and measures for each incident.

What a person decides: Measures.

For whom: Manager.

How to measure: Repeat mistakes.

Example result: “Situation: on October 3 a patient came for an appointment, but the doctor wasn't there — the booking was moved in the system, but the customer wasn't told (incident #41). Key point: this is the third such case this month — rescheduled bookings don't reach the customer. Options: 1) a rule ‘reschedule only together with a message to the customer’ — simple, but relies on people; 2) confirm every booking a day ahead — catches all such errors, but adds front-desk work. Recommendation: 2 — it covers this and similar errors. Today: assign someone to own confirmations.”

What it does: Scores a sample of conversations against the standard.

What a person decides: Conclusions.

For whom: Manager.

How to measure: Conversation quality score.

Example result: “Situation: 30 conversations from the week checked: 6 with no greeting, 4 answered after more than an hour, 9 with no next step — the customer wasn't told what to do next. Key point: ‘no next step’ is the most common and most costly — those customers book less often. Options: 1) add a closing next-step line to the saved replies — fast; 2) a short training with 5 examples — slower but more reliable. Recommendation: 1 now, 2 at the team meeting. Today: update the 10 most frequent replies.”

What it does: Deadlines, fines and fuel.

What a person decides: Expenses.

For whom: Manager.

How to measure: Vehicle fines and overdue items.

What it does: Repair planning and downtime.

What a person decides: Repairs.

For whom: Manager.

How to measure: Unplanned downtime.

Example result: “Action: schedule compressor maintenance for Monday, 9:00–11:00: the last service was in March, 7 months ago, and the manual says every 6 months; Monday morning has the fewest bookings. Impact: lowers the risk of a breakdown during the Friday rush, when downtime costs the most. Risk: the service company may not come Monday — confirm the request today. Fallback: if they're booked, move it to Tuesday morning.”

What it does: Deadlines for safety briefings and medical checkups.

What a person decides: Conducting them.

For whom: Manager.

How to measure: Workplace safety fines.

Example result: “Situation: per the logs, 4 employees have been overdue for a refresher briefing since October 1, and 2 have medical checkups expiring in November; the new warehouse employee has no record of an induction briefing. Key point: an overdue briefing is the first thing inspectors check. Options: 1) brief all four as a group on Wednesday — fast; 2) one at a time over the week — no work stoppage, but slower. Recommendation: 1. Today: set a time on Wednesday and book two people for checkups.”

What it does: Tracks the “who signs next” route, sends reminders and shows where things are stuck.

What a person decides: Signing.

For whom: Staff member, manager.

How to measure: Days in approval.

Example result: “Situation: 5 documents in approval; the Arman LLP contract has been with the lawyer for 4 days with no comments, and the director is next on the route; the reconciliation statement with Medtech has been waiting 2 days for the chief accountant. Key point: the Arman contract is needed by Friday — the start of work depends on it. Options: 1) remind the lawyer and ask for it back today — simple; 2) send it to the director in parallel marked ‘without legal sign-off’ — faster but risky. Recommendation: 1, and give the director a heads-up. Today: a reminder to the lawyer.”

What it does: Lays out a process step by step and finds the bottlenecks — from the description in the file.

What a person decides: Process changes.

For whom: Manager.

How to measure: End-to-end process time.

Example result: “Situation: from request to technician dispatch takes 6 steps and 2 days on average. Key point: requests sit with the dispatcher until evening — he handles them in one batch at 17:00. Options: 1) process requests every 2 hours — simplest; 2) send requests straight to technicians — faster, but distracts them. Recommendation: 1. Today: agree on the new routine with the dispatcher.”

What it does: Reviews a mystery shopper visit against the checklist: what went well, what didn’t, and what to fix.

What a person decides: Conclusions about staff.

For whom: Manager.

How to measure: Service standard compliance.

Example result: “Situation: October 14 visit, Abay Ave. location: 8 of 11 items met. Key point: the customer wasn’t offered a next appointment — the item repeat sales depend on. Options: 1) add a line to the front desk script; 2) a reminder in the POS system. Recommendation: 1. Today: discuss at the team meeting.”