The Economics of Operational Excellence: Uptime, Not Headcount

Incident time is a payroll cost that never shows up as a budget line, and most of the value of fixing it comes from uptime, not headcount. A value model to fill in with your own numbers, and a weekly team scorecard that keeps it honest.

PerspectiveUpdated Nov 2, 2026

Cover Image for The Economics of Operational Excellence: Uptime, Not Headcount

The Economics of Operational Excellence: Uptime, Not Headcount

In the annual budget review, the CFO asks a simple question: what did incidents cost us last year? The operations report answers in minutes per incident. It has nothing in money, because the largest cost of an incident is engineers' time, and that sits in payroll, spread across salaries.

This part puts a number on it, and the larger number isn't the one most business cases lead with. Most of the value of faster investigation comes from uptime, not from engineer time.

It's the last part of the Agentic IT Operational Excellence series.

Incident time is a payroll line nobody budgets

The hidden cost of incident time covers the survey data: most teams report spending 40% or more of their time on incidents. Use that as the starting assumption and turn it into money. Take a team of ten SREs:

  • At international cost levels, about $150,000 per engineer a year, that's roughly $600,000 a year of payroll spent on incidents.
  • At Vietnamese cost levels, about $40,000 per engineer a year, it's roughly $160,000 a year.

Not all of it can be recovered. New failures still need people, and they should. A business case should rest on your own incident classification (how many incidents repeat a known failure, how many are noise, how many are new) instead of someone else's benchmark.

Most of the value is uptime, not headcount

When people build the case for agents in operations, they usually start with engineer time. It's the easiest number to find. It's also the smaller one.

Here's an illustrative model for the same ten-person team. The assumptions are deliberately conservative. Engineers spend 40% of their time on incidents, and agents recover only 20% of that. There are 12 customer-facing incidents a year, and half of them get cut from 99 minutes to 14 with Resolve doing the investigation, the illustrative timeline from the payments incident. A minute of degraded service costs $1,000 at international levels and $200 at Vietnamese levels.

Value line (illustrative)InternationalVietnam
Engineer time recovered$120,000$32,000
Uptime protected (99.77% to 99.87%)$510,000$102,000
Total value (before cost)$630,000$134,000

Uptime is about four fifths of the total in the international column and three quarters in the Vietnam column. Even where engineers cost less, it's the larger line, because cutting the degraded time of recurring incidents from about 99 minutes to about 14 protects revenue and SLA commitments.

That changes who should care. A case built on engineer time is an efficiency story, and it invites the question "so how many people can we cut?" A case built on uptime is a revenue protection story, and it belongs in front of the business.

Build the value model with your own numbers

The table above is a shape, not a forecast. Replace every assumption with your own data. The model has four lines.

01Incident payroll

Engineers × cost × share

Headcount, fully loaded cost and the share of time that goes to incidents.

02Recoverable time

× what agents can take

Only the part spent on repeats, noise and catchable changes. New failures still need people.

03Uptime

Minutes saved × cost per minute

Customer-facing incidents per year, minutes saved on each, cost of a degraded minute.

04Cost side

Your volume × your quote

Investigation volume times the price you are quoted, plus rollout effort.

A value model you fill in with your own numbers. Uptime is usually the biggest line.
  1. Incident payroll. Your operations headcount, their fully loaded cost and the share of their time that goes to incidents. If you don't know the share, measure it for a month. Most ITSM systems only record when an incident opens and closes, so you may need to ask engineers directly.
  2. Recoverable time. The part of that payroll spent on known, recurring failures, noisy alerts and risky changes that could have been caught before release. Start low. 20% is a conservative place to begin.
  3. Uptime. Your customer-facing incidents per year, the minutes an agent would realistically save on each, and what a degraded minute costs in revenue and SLA penalties. Your finance team usually has that last number, even if nobody in engineering has asked for it.
  4. Cost side. Your expected investigation volume times the price you're quoted, plus the effort of rolling out. Count the alerts you ignore today: once every actionable alert gets investigated, volume will be higher than your current incident count. Don't assume it falls over time either. Tokens get cheaper, but each agentic workflow tends to use more of them.

Then compare value with cost, using numbers you trust because they're yours.

You don't have to take the assumptions on faith. Run agents read-only for 30 days, alongside your team, on your own incidents. Compare their verdicts with your postmortems and their timelines with your real investigation time. By day 30 you'll know how close your recoverable share is to the 20% assumption. Control before trust covers the 90 days that follow.

The weekly team scorecard

A business case gets the program approved. A scorecard keeps it honest: every team, measured every week, on the same five pressures that years of transformation left behind.

01Cost

Cloud spend

Change in spend for the services the team owns. Optimize works this column.

02Reliability

MTTR

How long the team takes to recover service. Resolve works this column.

03Security

Open risks

Confirmed weaknesses still waiting for a fix. Cyber works this column.

04Velocity

Change failure rate

The share of changes that cause a failure. Review works this column.

05People

Toil per week

Hours on repetitive, manual operations work. Every agent should move it.

Five measures on every team's weekly scorecard, and the agent that works each one.

Here's what that looks like with five teams. Cloud spend is the change since the baseline, and the last column counts the measures that improved since last week. The numbers are illustrative.

TeamCloud spendMTTROpen risksChange failToil a weekImproved
Payments-18%12 min24%6 h5 of 5
Checkout-12%18 min36%9 h4 of 5
Platform-9%25 min58%12 h4 of 5
Data-4%41 min711%15 h3 of 5
Mobile+2%55 min914%19 h1 of 5

The value is in the last column: who's improving, and who's slipping. It turns operational excellence into a shared goal instead of a blame game, and it puts the five pressures on one page that engineering and the business both read. Teams that care about something the five columns miss can build an agent for it with Customize and add their own column.

Under the scorecard, keep the monthly operating set from why incident investigation is slow. Time to remember is the one that explains the others.

Where to start

  1. Take the baseline on the scorecard's five measures, with investigation time split out from the total.
  2. Set the bar for scaling before the pilot starts: the investigation-time improvement, the repeat rate and the minimum return that move you to a wider rollout. A bar decided afterwards always moves.
  3. Decide where recovered time goes. Time nobody assigns gets absorbed. I'd put it on the prevention items that keep slipping behind the roadmap, because those stop the next repeat.

Then fill in the uptime line from your own history:

"List last quarter's customer-facing incidents with the minutes of degraded service for each, and mark the ones whose root cause we'd seen before."

See how Resolve shortens investigation.

The payments incident that runs through this series is a composite. Every number in the value model and the scorecard is illustrative.

About the author

ST

Steve Tran

Founder & CTO, CloudThinker

Steve Tran is the founder and CTO of CloudThinker, which builds Frontier Investigation Agents for production cloud operations. He writes about incident investigation, cloud cost, and the practices teams use to safely run autonomous AI agents against production AWS, GCP, and Azure environments.

economicsroiuptimesreoperationscio