iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
A document-automation pilot can post a 90% faster turnaround and still cost many times more per finished document than the manual process it replaces. That is the central claim of an audit described by Richard Ewing in an opinion piece listed by CIO on October 8, 2026. The case involves an anonymized mid-sized enterprise, and the figures below are the author’s account of that audit, not independently verified numbers. They are still useful as a checklist, because the failure mode they describe is common: the pilot measures speed and uptime, while the production economics are measured in a different budget, at a different volume, on messier inputs.
What the audit reported
The enterprise used automation to review vendor onboarding agreements. Before the pilot, a clerk checked five or six clauses, verified vendor details, and filed each document in about four to five minutes. Ewing estimates the fully loaded labor cost of that manual process at roughly $0.80 per document. The automated pipeline reportedly cost $12 to $14 for the same kind of file.
The pilot itself looked strong on the metrics it tracked. Turnaround time fell by 90%, and the system stayed up. The problem, as Ewing tells it, was where the costs sat. Computing bills, data lookups and third-party model fees were paid from a central innovation fund, so the business unit that owned the process did not see them at first. Once production costs were allocated to the department, the author says the economics turned negative within sixty days at full transaction volume.
Free tools Windows power users keep installed
One-click scans. No signup required.
The manual baseline
Any cost-per-task comparison starts with the baseline. In this account, the baseline is a clerk spending four to five minutes on a document at a loaded cost of about $0.80. That figure implies a loaded labor rate of roughly $9.60 to $12 per hour, which is a useful sanity check when you are estimating your own review time. If your manual baseline cannot be expressed as minutes multiplied by a loaded rate, you do not yet have a baseline.
#1 Best Overall
Where the automated cost came from
The $12 to $14 per document figure covers the automated pipeline as the author describes it: infrastructure compute, database and reference-data lookups, and third-party model charges. The article does not break that figure into line items, so it cannot tell you which component dominated. It does tell you that the total was roughly fifteen to seventeen times the manual cost for each document, before any human review was counted.
Cost per completed task: the arithmetic
The most important correction to a pilot dashboard is to divide total cost by completed, accepted outputs, not by API calls, seats or documents attempted. The table below sets the reported manual figures beside the reported automated figures and an illustrative blended figure that adds human exception handling. The blended row is arithmetic on the author’s numbers, not a figure the article states.
Rank #2
| Measure | Manual clerk process | Automated pipeline (as reported) |
|---|---|---|
| Time per document | About 4 to 5 minutes | Not stated |
| Cost per document, fully loaded | About $0.80 (author’s estimate) | $12 to $14 per file (author’s account) |
| Share of documents needing human review | Not applicable | About 40% of daily transactions fell below the confidence threshold (author’s account) |
| Human time per exception | Not applicable | About 10 minutes, roughly twice the manual baseline |
| Illustrative blended cost per document, including exception labor | About $0.80 | About $12.60 to $14.80 (calculated from the figures above; assumes the pipeline cost applies to every document and exceptions cost about $1.60 to $2.00 of labor each) |
The blended range is sensitive to two assumptions. If the pipeline cost is lower for documents that pass the threshold, the blended figure falls. If exception documents consume extra model passes, it rises. Either way, the order of magnitude does not change: a process that costs $0.80 by hand and well over $12 per document by machine is not a saving unless the automated cost falls by more than an order of magnitude.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsHuman exceptions are a second bill
Automation rarely removes human work. In this case, roughly 40% of daily transactions fell below the system’s confidence threshold and went to a person. Each exception took about ten minutes, because staff had to inspect the source document and the system’s partial output side by side. That is double the manual time for the whole task, so the exception queue is worse than the old process for those files.
Rank #3
Two practical points follow. First, measure the exception share on production inputs for a full business cycle, not on the demonstration set. Second, measure the time per exception as a separate line, because reviewers often spend longer on a half-finished machine output than they would on a clean file they process from scratch.
Messy inputs multiply the work
The production queue described in the audit contained low-resolution scans, rotated photocopies, handwritten notes and conflicting payment terms. The author says these inputs triggered extra extraction passes, more reference-data lookups and more validation steps. Those steps consumed compute and money without reliably resolving the ambiguity, so the system paid for more processing and still sent a large share of files to people.
This is the part of the pilot most likely to mislead. A clean demonstration set usually has fewer retries, fewer lookups and fewer exceptions than a real queue. If the pilot used born-digital PDFs and the production queue is mostly phone photographs of paper, the cost per task measured in the pilot will understate the cost in production.
When the pilot fund stops paying
The budget move is where many AI cases quietly change. Infrastructure and model costs charged to a central innovation fund do not appear in the department’s operating numbers, so the business case looks better than it is. The author’s point is that the department’s cost should be recalculated once those charges are allocated to the process that causes them.
Best Value
A workable check has three parts. Identify every charge that the pilot pays from outside the process owner’s budget. Allocate those charges to documents processed, using the same unit as the manual baseline. Then recalculate the cost per completed task at expected production volume, not at the pilot’s volume. The question to ask is not whether the tool is cheap, but whether the organization is replacing a payroll cost with a consumption meter whose size depends on how messy the inputs are.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Two price trends that measure different things
Published figures about AI costs are often quoted as if they describe the same thing. They do not, and the audit’s argument depends on keeping them apart.
| Source and date | What it measures | Scope and limits |
|---|---|---|
| Stanford HAI, AI Index, 2025 | Inference price for a model at GPT-3.5-equivalent level on MMLU: about $20 per million tokens in November 2022 and $0.07 per million tokens in October 2024, a reduction of more than 280-fold | A benchmark-level price per token for one capability level. It is not the cost of a completed business task. |
| Gartner press release, August 17, 2026 | A forecast that inference cost per agentic workflow will rise more than fivefold through 2028 | A forecast, not an observed outcome. It attributes the increase to greater workflow complexity and token use, even as model prices fall. |
| Ewing, CIO opinion piece, October 8, 2026 | Cost per document in one anonymized audit, with a 40% exception rate and a 10-minute exception time | The author’s account of one enterprise. The figures were not independently verified. |
Gartner’s Will Sommer, Senior Director Analyst, put the implication for product leaders this way in the same release: “Product leaders cannot rely on more efficient token economics to rationalize AI costs.” He also said, “Each successive generation of AI capability will necessitate more, and often more expensive, tokens. There is no reliable, economical one-size-fits-all model on the horizon.” Cheaper tokens and costlier workflows can both be true at once, because a workflow can use more calls, more context and more retries than a single query.
How to run the check on your own workflow
- Define the completed task. Write down what counts as a finished, accepted output, such as one onboarding agreement with all clauses checked and vendor data filed. Count only those outputs in the denominator.
- Set the manual baseline in loaded dollars. Multiply minutes per task by a fully loaded hourly rate that includes benefits and overhead. Record the result as a single number per task.
- Collect the automated cost per task from billing data. Include compute, storage, database or lookup calls, third-party model charges and any tooling billed separately. Use the invoice period that covers the same tasks you counted.
- Count the workflow steps per task. Record model calls, retries, reference-data lookups and validation passes for a sample of documents, grouped by input quality. Clean and messy inputs should be separate rows.
- Measure human intervention on production inputs. Record the exception share and the minutes per exception over at least one full business cycle. Multiply by the loaded rate and add the result to the automated cost.
- Allocate every charge to the process. Move any cost held in a central fund, shared account or innovation budget into the process’s cost base, then recalculate the cost per completed task.
- Repeat at expected production volume. Use the volume you expect after rollout, and check how the cost changes as the input mix shifts toward the messy end of the queue.
What the evidence does and does not establish
- The per-document costs, the 40% exception rate, the ten-minute exception time and the 90% turnaround improvement come from one anonymized audit described by its author. They are not verified against invoices or payroll records, and they should not be read as industry averages.
- The Stanford figures describe falling benchmark prices for a fixed capability level. They do not show what a multi-step business workflow costs.
- The Gartner figure is a forecast for agentic workflow inference cost through 2028, published August 17, 2026.
- Neither external source independently confirms the audit’s numbers, and neither compares named vendors or products. The argument is about how to measure cost, not which platform to buy.
The practical lesson for a pilot review is to keep the dashboard metrics you already have and add a cost-per-completed-task line beside them, built from the full bill and the real exception queue.
Quick Recap
“
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
