The Operations Audit

There is a striking finding in Microsoft’s 2026 Work Trend Index, a survey of twenty thousand knowledge workers across ten countries. Organizational factors account for 67% of AI’s measured impact. Individual mindset and behavior account for 32%.

Read that in practical terms. Two thirds of whether AI works for you is determined by how the work is designed and supported, not by which tool you bought or how enthusiastic your people are. Buying a subscription and handing it to your team is, statistically, the smaller half of the problem.

The rest of the 2026 research says the same thing from different angles. MIT’s analysis of enterprise deployments found returns were consistently stronger in operations and finance, while budgets overwhelmingly went to sales and marketing. The work that pays is the unglamorous, high-volume, rule-shaped work. Invoice reconciliation. Order entry. Status updates. Chasing paperwork.

Nobody has ever been excited to buy software for that. Which is precisely why the returns are still sitting there. So this worksheet does not ask what tool you want. It asks where your hours go, which of those hours are rule-shaped, and what they cost.

List up to eight recurring tasks. Not projects. Not decisions. Tasks: work that happens on a schedule or a trigger, gets done roughly the same way, and produces a predictable output.

Three prompts that surface most of the list in about five minutes:

• The Friday question. What does someone assemble every week that is mostly copying from one place into another?
• The interruption question. What arrives by email and has to be typed into a system?
• The chase question. What does someone spend time chasing because it did not arrive on time or in the right format?

What does not belong on this list: anything that requires a judgment call that changes case by case, anything done fewer than a dozen times a year, anything where the hard part is a relationship. Those are real work, they are just not this.

If you cannot fill four rows, this is not your bottleneck and no amount of automation will help. That is a genuinely useful answer.

Annual hours is the only unit that makes tasks comparable. A twenty minute daily job beats a four hour monthly one, and almost nobody guesses that correctly.

Hours per year = times per week × minutes each time × 52 ÷ 60

Anchor: one full-time role is 2,080 hours a year. Hold onto that. It becomes the most important number in step 6.

Use fully loaded cost, not salary. Payroll taxes, insurance, benefits, and overhead put the real number 1.25 to 1.4 times base salary, and 1.3 is the sensible middle.

Fully loaded hourly rate = salary ÷ 2,080 × 1.3

Annual cost of the task = hours per year × fully loaded hourly rate

That total is your gross number. It is real, and it is also the most optimistic version of the truth. Step 5 makes it honest.

Cost tells you what a task is worth. This tells you whether it can be handed off at all. Score each task 0 to 3.

• 3 · Fully rule-shaped. Same inputs, same steps, same output. A new hire could follow written instructions. Exceptions are rare and recognizable.
• 2 · Mostly rule-shaped. Standard path covers most cases, with a handful of known exceptions someone routes by hand.
• 1 · Judgment-heavy. The steps are consistent but the decisions are not. Requires context about this customer, this account, this situation.
• 0 · Not a task. It looked like a process. It is actually someone's expertise.

Most people score their own work too low. The test is not "could a computer do this," it is "could you write down what you do." If the answer is yes, it is at least a 2, whatever it feels like from the inside.

Also mark the blast radius. H if a mistake would reach a customer, a regulator, or the bank before anyone caught it. L if it gets caught downstream. This does not change whether you automate. It changes whether a person approves the output before it commits.

Here is the part most worksheets like this leave out, because leaving it out produces a bigger number. Saved hours are not recovered hours.

Two things eat the difference. The first is supervision: half of AI users in Microsoft's 2026 survey named quality control of AI output as a skill that is becoming more important, not less. Somebody checks, and that check is real work.

The second is exceptions. A task that is 90% routine still sends 10% of cases to a human, and those are the slow ones. The routine 90% might take four minutes; the exceptions take twenty.

Net annual value = annual cost × recovery factor

A worksheet that told you the gross number was your savings would be lying to you, and you would find out in month three. This one will be roughly right, which is more useful.

Two businesses can arrive at the same net number and face completely different decisions. What separates them is whether the hours sit in one place.

Concentrated. The hours cluster in one or two people, and together they approach or exceed roughly 1,250 hours a year, about 60% of a full-time role. At that density the hours are convertible. You can genuinely not make the next hire, or move that person onto work you have been deferring for a year. The saving is real money and it shows up in the P&L.

Scattered. The same total hours, spread thinly across six people at three hours a week each. Nobody's job changes. What you get is capacity: three hours a week back for six people, which is worth having and is worth roughly half what the arithmetic suggests, because reclaimed fragments are always worth less than reclaimed blocks.

Both are worth doing. They are just not the same purchase, and confusing them is how automation projects get killed by a CFO in month four. If you promised headcount savings and delivered scattered capacity, the project failed even though the automation worked perfectly.

Concentrated or scattered?The single most important line on this worksheet.

Plot each task. Hours per year up the side, rule-shaped score across the bottom.

A freight brokerage, eleven people, running on a TMS plus email plus a shared drive.

Gross: $47,476 a year. Applying recovery factors task by task brings the net to roughly $33,000.

Concentrated or scattered? Three of the five tasks belong to one operations coordinator, totaling 1,040 hours. That is half a role sitting in one person, and adding the report and reconciliation work pushes the cluster past 1,250. So: concentrated, and convertible.

What they should do: order entry is high hours, rule-shaped 3, low blast radius. Unambiguously first. Invoice reconciliation is equally clean but high blast radius, so it goes second with a human approving before anything posts. What they should not do: all five at once.

Gross annual cost

Net after recovery factors

The one task you would start with

That net number is what one more year of doing this by hand costs you. It is not a sales figure and we did not choose the inputs. You did.

It is probably conservative, because almost everybody underestimates task frequency and forgets the chase work entirely. And it excludes error cost, which for document-heavy operations is not small: industry estimates put the cost of resolving a single invoice error at up to $53.50, and manual data entry error rates typically run between 1% and 4%. If your list is full of typing, add a line for that.

In every case, before you build anything, check three things. Can you get credentials to the systems involved, and how fast? Is there one person who can make decisions about this without a meeting? Can you measure your net number today, before anything changes? That last one is where most projects quietly fail, because a deployment that worked but cannot be proven gets canceled anyway.

Those questions come from our companion worksheet, the AI Readiness Assessment. This one finds the task and prices it. That one tells you whether you are in a position to deploy it.

1. Microsoft Work Trend Index 2026, fielded February 18 to April 20, 2026 by Edelman Data x Intelligence. 20,000 full-time knowledge workers across ten countries. Source of the 67% organizational versus 32% individual split, the 50% citing quality control of AI output as increasingly important, and the 66% reporting more time on high-value work.
2. MIT NANDA, "The GenAI Divide: State of AI in Business," 2025. 52 executive interviews, 153 survey responses, 300 public deployments. Source of the operations-versus-marketing return finding and the focused-use-case finding. Directional rather than definitive given sample size.
3. Fully loaded cost multiplier, 1.25 to 1.4x base salary. Consistent across 2026 US compensation guidance, covering payroll taxes at roughly 7.65% FICA plus benefits and overhead. We use 1.3 as the midpoint.
4. Error costs. IOFM estimates up to $53.50 to identify, investigate, and resolve a single invoice error. Published manual data entry error rates cluster between 1% and 4% for structured fields, higher for free text.

On numbers we left out: several widely circulated figures about hours lost to repetitive work, ranging from 41% to 62% of the working week, trace back to sources we could not verify and disagree with each other by a factor of nearly two. We left them out. This worksheet uses your numbers rather than a benchmark for the same reason: a benchmark that is wrong for your business produces a confident answer in the wrong direction.