AI for COOs and Operators in 2026: KPI Reviews, Vendors and SOPs

AI can prepare a COO's operating rhythm: flag which KPIs moved before the review, turn the meeting into owners and dates, surface vendor terms and redraft SOPs from how work runs now. It can't tell you a metric's definition is right, hold a peer function to a date, or certify a procedure.

This is for whoever owns the rhythm, whatever the title says: COO, VP of operations, general manager. O*NET's occupation summary for general and operations managers names the recurring desk work plainly: reviewing performance data to measure goal achievement, monitoring suppliers against budget, and establishing procedures. So this page skips the company-wide AI program and starts with your next review.

Four loops, and where each one leaks

An operator's admin runs in four loops on four different clocks, and each loop leaks at a predictable point. AI is useful at the leak and mostly wasted elsewhere. Products named here are examples from vendor pages, in no particular order.

KPI reviews: automate the "what moved", distrust the "why"

BI tools with AI can now tell you what moved before the meeting, but any explanation of why depends on the fields in the model. Tableau Pulse, on Tableau Cloud, checks followed metrics every 24 hours, detects drivers, trends, contributors and outliers, and sends digests by email and Slack.

Copilot in Power BI summarizes reports and answers questions, but needs paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher); a Pro license alone isn't enough. Its email subscription summaries don't take your own prompts and don't cover scorecards or dashboards.

Microsoft's own documentation names the trap. Without a prepared semantic model, Copilot can misinterpret the data, and a question that isn't about the model gets answered from the language model's general knowledge. Say on-time delivery falls. A tool can show the drop sits in one distribution center and one carrier, if those fields exist. It can't know that someone redefined "promised date" last month or that the carrier's contract changed, and a general-purpose chatbot asked why may hand you a tidy, plausible cause anyway.

So split the work. Let AI write the first pass of the movement, and make the named metric owner responsible for the cause. This matters more if you run EOS: Ninety's default Level 10 agenda gives the scorecard review five minutes and sends off-track numbers to the Issues List, so the explaining has to happen before anyone sits down.

Metric: [name]. Definition: [numerator, denominator, source system, cutoff].
Attached: this period, last period, plan, broken out by [site / carrier / product].
1. State the change against last period and against plan, with the numbers.
2. Show which segments account for most of the change.
3. List causes the data can support, and label each one "in the data".
4. List questions only the metric owner can answer, and don't guess at them.
If the definition or a field looks inconsistent between periods, say so first.

Cross-functional follow-through: capture is the easy half

Note takers can list who agreed to what; the part that slips is the dependency between functions. Take a line from an ops review: "Sales gets ops the forecast by Friday so procurement can place the order." That one sentence holds three functions and two deadlines, and an action list can easily keep only the first half.

Ask for action lines that carry the dependency, in the same format every time, and keep them where the team already works:

Owner: [one name, never a team]
Deliverable: [what exists when this is done]
Due: [date]
Blocks: [who can't start until this lands, and their date]
Evidence of done: [link, report or number you can check]

Asana AI drafts updates and summarizes projects, with AI Studio Basic and AI Teammates available on all paid plans, though Teammates usage is charged per AI request. Read any such draft as a status built from what's in the tracker. If procurement hasn't touched its task since Tuesday, the draft can't know more than the task does, so a green update tells you less than a date and a link.

The chase itself should stay human above a certain level. A note to the head of sales saying the forecast is holding up the order is a conversation between peers, and it reads differently when software sends it. Let AI draft it and show you who is blocked; send it yourself or pick up the phone.

Vendors: track the notice date, not the renewal date

The date that costs you money is the last day to give non-renewal notice, which can fall months before the renewal itself. Ramp extracts contract details such as start and end dates and sends reminders 60 and 30 days before renewal. That's useful, and it also shows the gap: on a contract requiring 90 days' notice, both reminders arrive after the window has shut. Set your own alert from the clause.

Docusign Agreement Manager, included in select IAM plans, surfaces data from agreements and tracks key terms and deadlines. Whichever you use, open the source clause for your largest contracts and confirm the extracted date before it goes on a calendar. If you have no contract tool yet, a prompt run on a company-approved account gets you a first list, provided the contract's confidentiality clause allows it:

From this contract, list and quote the clause for each:
- Term, renewal type, renewal date
- Notice required to stop renewal, and the last calendar day to give it
- Price increase terms
- SLAs, service credits, and how a credit must be claimed
- Termination for convenience
- Audit, data handling and subcontractor terms
If a term isn't in the document, write "not found". Don't infer it.

Quarterly business reviews are the other half of the loop. Pull the SLA report, open tickets and any credits you're owed, then ask for a one-page agenda with the misses first. AI for vendor security review is more mature: Vanta says it flags high-stakes SLAs, DPAs and custom security terms in SOC 2 reports, and Whistic ties findings to source documents. Whistic is clear on the limit: its AI doesn't approve vendors, and human reviewers make the risk decision.

SOPs: drafts got cheap, ownership didn't

Capture tools have made the first version of a software procedure nearly free, while keeping it accurate remains the actual job. Scribe turns a captured workflow into a step-by-step guide, with a free Basic plan and Pro Team from $13 per seat monthly on annual billing, starting at 5 seats. Tango has a free plan limited to 5 shared workflows, and Pro from $15 per user monthly on annual billing at 3 or more users. Trainual drafts from a description and adds completion tracking and content verification reminders; it publishes no plan prices, only a $1,000 implementation fee.

Two gaps remain. Capture sees the screen, so the step on the dock or the shop floor never gets recorded. And a verification reminder only asks whether someone looked. The better trigger is the change itself: when a review approves a process change, ask which SOPs it touches, then walk the affected one with the person doing the work. A transcript of that walkthrough, compared line by line with the current SOP, gives you a redline for the owner to accept or reject.

Where AI should stop in operations

Three lines come up in this job more than most, beyond the sign-offs above.

• Allocating work and rating people. If you have staff in the EU, the AI Act's Annex III treats systems that allocate tasks based on individual behavior or traits, or that monitor and evaluate worker performance, as high-risk. Those obligations now apply from December 2, 2027. A follow-through report is fine as a list of open items, but think twice before it becomes a ranking of your managers, and ask counsel before a scheduling or scoring tool goes live.
• Recording vendor and cross-site calls. Participants may sit under different state laws. California, for one, requires all parties to consent before a confidential conversation is recorded. Announce the recording at the start of every QBR.
• Agents that commit money. Purchase orders, credit claims and contract notices create obligations. Keep a person on the approval step, and turn off any default that sends them unreviewed.

Where Merlin fits: your side of the follow-through

Merlin keeps track of what you personally owe and are owed, supplier by supplier and colleague by colleague. It's our desktop app for the Mac. If your work mail and calendar are on Microsoft 365, it won't connect to them for now, because Merlin's mail and calendar connections are for Google. For an operator, it covers four things:

• Reviews. Merlin records locally on every plan; getting the words transcribed takes a paid one. Enhanced notes follow a preset such as Standup or 1:1, or your own template, so you can build one with sections like "Metrics off track" and "Blocks". Once a recording is transcribed, Merlin may message you with your own to-dos and what people owe you, and you can ask in chat what a supplier or a function head committed to across your recordings.
• Walking in prepared. Meeting Brief lands some ten minutes before the call starts, covering who's attending, what you covered last time, open commitments and ideas for the agenda, which suits a QBR you last thought about three months ago.
• Who's waiting. The People panel flags people who emailed you directly in the last 2 to 30 days and still await your answer, so a vendor's account manager doesn't sit unanswered. Email is its only signal, so a call doesn't clear the flag.
• Repeat checks. Routines start from a timetable or an incoming Gmail message, for example telling you when a renewal notice lands. They work unattended, and if you allow Slack on one, its posts go out with no review. In chat, by contrast, a send waits for your confirmation.

Be clear about what it isn't. Merlin is not a BI tool, a contract repository, a procurement system or a shared action tracker. Recordings have no share option, so the team's log still belongs in your project tool. The To-do panel takes in Linear, Notion, Todoist and Granola, but not Asana or Trello yet. Its Subscriptions panel can show software renewals when a receipt or bank alert reaches a connected inbox, with estimated dates, so a bill that only reaches your AP team's inbox won't appear, and no notice clause gets read. It doesn't place phone calls.

Check the data path against your supplier NDAs too. A copy of the audio stays on the recording machine, and on paid plans the audio is also sent to a transcription service; transcripts and notes sit in your Merlin account, and the privacy policy says AI model providers follow their own data rules. For the wider question of delegating to software, see this guide to handing work to an assistant.

Test this on one review before you buy anything

1. Pick one recurring review and one metric that moved. Write its definition in a sentence, including the source system and cutoff. If the room can't agree on that sentence, fix it before adding any tool.
2. Run the variance prompt, then hand the result to the metric owner. Have them mark every wrong or unsupported line. Those marks tell you how far to trust the next draft.
3. Capture the next review, with everyone told. Compare the extracted actions with your own notes, and look for what happened to the dependencies between functions.
4. Take your ten largest supplier contracts. Find each last-day-to-give-notice date, check it against the clause, and put it on a calendar someone actually watches.
5. Choose one SOP that changed this year. Walk it with the person doing the work, get the redline, and name one owner for the document.