Free calculator · no email required
What is the manual work costing you?
A task library from real ecommerce operations, priced at a fully-loaded hourly rate — and marked honestly for which work an agent can take today, which needs a human in the loop, and which should not be automated at all.
How to cost an hour of internal time
Fully-loaded hourly cost is salary plus employer overhead — tax, benefits, tooling, space — divided by the hours actually available, not by 2,080. Allowing for holiday leaves roughly 1,725 productive hours a year, so a $52,000 salary at 25% overhead is near $38 an hour rather than the $25 a naive division suggests.
Your week, in hours.
Prefilled at the midpoint of what we observe — replace each with your own. Set anything you do not do to zero.
Manual operations, a year
$89,267
51.5 hours a week at $37.68 fully loaded — 137% of a full-time role.
- An agent could ownfull-automation tasks only
- 29.0 h/wk
- Worththose hours, a year
- $50,267
- Net saving after running cost
- $48,107
- Payback on the build
- 3.0 months
16.0 hours a week are partly automatable — an agent drafts, a person approves — and 6.5 should stay human. Neither is counted in the saving above.
Where automation does not belong
Most of this category sells the opposite. Two tasks in the library are marked human, and the reason is the same in both cases.
- 1
Where a wrong answer costs more than a human minute
A refund decided badly costs the order, the customer and the review. The minute you saved is not worth the asymmetry, so an agent assembles the case and a person decides.
- 2
Where a false positive is invisible
Cancel a legitimate order as fraud and the customer does not complain — they leave. You never see the cost, which is exactly why the decision needs a human on anything ambiguous.
- 3
Where the data is not reliable yet
Automation built on a catalogue nobody trusts industrialises the error. Fix the data first; the automation is easy afterwards and impossible before.
- 4
And what that leaves
The rest — WISMO, subscription changes, feed fixes, reconciliation, reporting — is high-volume, deterministic and unloved. That is where the hours are, and it is what the ops automation build covers.
The task library
12 recurring jobs, the weekly hours we observe, and what an agent can actually do with each.
| Task | Hours/wk | Verdict | What an agent does |
|---|---|---|---|
| Where-is-my-order replies Support | 4–12 | Agent can own it | The tracking number exists in the system before the customer asks. An agent reads the order, the carrier status and the delivery window and answers without a human seeing it. |
| Order edits and address fixes Orders | 2–6 | Agent drafts, human approves | An agent can validate the address and apply the change before fulfilment picks it up. Past the pick, it becomes a warehouse conversation and a human should own it. |
| Refunds and goodwill decisions Support | 2–6 | Keep it human | Deliberately not automated. A wrong refund costs more than the human minute it saves, and a customer who feels processed by a machine at the worst moment of the relationship does not come back. An agent can prepare the case; a person decides. |
| Subscription skips, swaps and pauses Orders | 2–8 | Agent can own it | Every one of these is a rule against known data. It is also the highest-value thing to automate, because the alternative to a fast skip is usually a cancellation. |
| Product feed errors and disapprovals Catalogue | 1–5 | Agent can own it | Merchant Center disapprovals are a fixed set of causes with a fixed set of fixes. An agent reads the diagnostic, patches the field and reports what it changed. |
| New product setup and metafields Catalogue | 2–8 | Agent drafts, human approves | Drafting descriptions, specs and metafields from a supplier sheet is fast and safe. Publishing without a human read is how a wrong price goes live. |
| Purchase order chasing Inventory | 1–4 | Agent drafts, human approves | An agent can track what is late and draft the chase. Supplier relationships are not improved by a robot escalating on your behalf. |
| Stock reconciliation between systems Inventory | 2–6 | Agent can own it | Comparing two systems and flagging the delta is exactly what software is for. This one is usually done in a spreadsheet on a Friday afternoon. |
| Payout and channel reconciliation Finance | 2–8 | Agent can own it | Matching payouts to orders across Shopify, the marketplaces and the 3PL is deterministic work nobody enjoys and everybody does late. |
| Weekly reporting assembly Finance | 2–6 | Agent can own it | Pulling the same numbers into the same sheet every Monday. The judgement about what they mean stays with a person; the assembly does not need one. |
| Returns approvals and RMA admin Orders | 2–7 | Agent drafts, human approves | Policy-compliant returns approve themselves against the rules. Exceptions — outside the window, damaged, high value — go to a person, which is most of the value of the split. |
| Manual fraud review Finance | 1–4 | Keep it human | Not automated beyond scoring. A false positive cancels a real customer's order and they do not come back to tell you; the asymmetry is why a person makes the call on anything ambiguous. |
The hour ranges are our observation, not a benchmark. They are what we see doing this work inside $3M–$30M Shopify operations, offered as a starting point so the form is not empty. They are not measured across a sample we could publish, and the figure that matters is the one you replace them with. metric to confirm. Reviewed 2026-09-09.
What the saving figure leaves out
Your time specifying it. Somebody internal has to describe the rules, review the edge cases and sign off the behaviour. That is real time and it is not in the build cost.
Running both in parallel. Nothing goes from manual to automated in one step safely. Budget a month of doing both while you compare outputs.
Supervision. Automation moves the work from doing to supervising. Cheaper, but not zero — somebody still owns it when it breaks at 4am on a Saturday.
The hours you get back are not cash. Freeing twelve hours a week does not reduce payroll unless somebody leaves. It buys capacity — which is the point, but it is a different argument from a cost saving and worth making honestly to whoever signs off.
Definitions
- Fully-loaded cost
- Salary plus employer overhead, divided by the hours genuinely available.
- Agent
- Software that reads context, decides within rules and acts — as distinct from a script that follows a fixed path.
- Human in the loop
- The pattern where software prepares and a person approves. Where most operations work belongs.
- Deterministic work
- Tasks with one correct answer given the inputs. The safest and highest-value automation targets.
- Payback
- Months for the saving to cover the build. Counted here on fully automatable hours only.
Questions about ops cost
How do I calculate the cost of an employee's time?
Take the salary, add employer overhead — tax, benefits, tooling, space, typically 20–30% — and divide by the hours actually available. Not 2,080: allow for holiday and public holidays and you get roughly 1,725 productive hours a year. Dividing by the bigger number is how a task that costs $40 an hour gets budgeted at $25.
Which of these tasks should we automate first?
The ones with the most hours in the “agent can own it” column, not the ones that annoy people most. Reconciliation and WISMO usually top the list because they are high-volume, deterministic and unloved. The tasks marked human should be left alone regardless of how tedious they are.
Why do you say refunds should not be automated?
Because the cost of a wrong answer is asymmetric. A refund decision made badly costs more than the human minute it saves, and a customer who feels processed by a machine at the worst moment of the relationship does not come back to tell you. An agent can assemble the case — order history, policy, previous contacts — and a person decides. Anywhere a wrong answer is expensive, that is the right shape.
Are the prefilled hours a benchmark?
No. They are the midpoint of what we observe doing this work inside Shopify operations at this size, offered as a starting point so the page is not an empty form. They are not measured across a sample we could publish, and the number that matters for your decision is the one you replace them with.
Why does the saving only count fully automatable tasks?
Because a build that pays back only if you count optimistic hours is a build that does not pay back. Partial tasks still need a human to review, and human tasks should not move at all — so neither is credited. If the payback works on the conservative number, it works.
What is not in the build cost?
The internal time to specify it, the month of running both processes in parallel, and the ongoing cost of somebody owning the automation when it breaks. Automation is not free after launch; it moves the work from doing to supervising, which is cheaper but not zero.
Related
Find out what you’re losing.
Before you commit to anything, we tell you exactly what you’re losing and what it costs to stop it. Two weeks. Fixed fee. Credited in full against any build you go ahead with.
- Fee
- $1,500–$3,000, fixed
- Duration
- Two weeks
- Credited
- In full, against any build
- You supply
- Read access + one 45-minute call