AI automation ROI is a measurement, and one afternoon is enough to produce it: time a real document through the old path, time it through the new path, price the run, and multiply by a volume your own finance team supplies. Any ROI line the client cannot verify gets cut from the business case.
The last time webvise ran this stopwatch, an 11-page construction tender went from roughly an hour of hand-typing to about 3 minutes 30 seconds, measured live in an on-site workshop with the estimator watching.
You have probably sat through an automation pitch where the savings slide was large and nobody in the room could check a single input. This article shows the method webvise uses instead: one measured baseline from a real client, a five-way classification for every business-case input, and run costs priced in cents. Copy it, fill in your own volume and labor rate, and the result is a number finance can audit.
- One measured baseline beats ten projections. At MP Bau, hand-typing an 11-page tender took about an hour; the automated path takes about 3:30 minutes and costs $1.50 to $2.50 per run.
- Classify every input: Known, Customer-supplied, Assumption, Estimate, or Unknown. Anything invented is deleted before the case is presented.
- The formula is method only: hours saved × loaded hourly cost × yearly frequency, minus build and run costs. Volume and labor rate come from the client, never from the vendor.
- Run costs live in cents. One gateway with a curated model list targets 20 to 30 cents per routine run and a two-digit euro amount per month, compared to €100 to €200 per seat for AI subscriptions.
- Most published ROI numbers cannot be verified. Refuse them, including any webvise number that was never measured.
If you want this measurement run on one of your own workflows before any budget commitment, webvise's AI automation service starts with exactly this timing session. Your documents, your workflow owner in the room.
Time one real document through both paths
webvise ran this measurement at MP Bau, a German construction company, during an on-site workshop. The estimator picked a real 11-page tender from a current project, the kind he keys in by hand. The old path takes him roughly an hour. The automated path converted the same document in about 3 minutes 30 seconds, with him watching the run.
The run itself cost between $1.50 and $2.50 depending on the model, straight from the gateway logs. That figure went into the business case exactly as logged. Nobody had to trust a vendor slide, because the person who owns the workflow saw both numbers happen.
That is the whole first step, and it works for any document workflow. Pick one real document, time the old path with the person who does the work, run the new path on the same document, and write down the run price. A baseline measured this way survives every challenge in the budget meeting, because every challenger can rerun it.
The formula is a method, and two inputs are never the vendor's
The frame itself is standard: yearly savings = hours saved per run × runs per year × loaded hourly cost, minus the one-time build and the per-run costs. Nothing about that equation is proprietary. The argument lives in where each input comes from.
Two inputs never come from the vendor. Runs per year is the client's own count, and loaded hourly cost comes from the client's finance team, salary plus employer costs. When webvise drafts a business case, both fields stay empty until the client fills them, and the case says so.
As method, with placeholders you must replace: the measured tender delta is about 56 minutes per document. At 100 documents a year that is roughly 93 hours, and at a €45 loaded hourly rate about €4,200 a year before build and run costs. The 56 minutes is measured. The 100 and the €45 are placeholders, and an honest case labels them that way.
Most ROI numbers published for AI automation fail this sourcing test. They arrive as finished totals with no baseline measurement, no run price, and no way for the buyer to rerun anything. A number you cannot rerun is marketing, and it belongs in neither your business case nor your board deck.
Classify every input before it enters the case
webvise sorts every business-case input into one of five classes: Known, Customer-supplied, Assumption, Estimate, or Unknown. The working rule behind it: business cases without invented ROI. An input that does not honestly fit the first two classes gets labeled, and the label ships with the case.
| Business-case input | Class | Where the number comes from |
|---|---|---|
| Minutes per tender, old path | Known | Timed live in the workshop: about 60 minutes for 11 pages |
| Minutes per tender, new path | Known | Measured on the same document: about 3:30 |
| Cost per run | Known | Gateway logs: $1.50 to $2.50 per conversion |
| Documents per month | Customer-supplied | The client's own count |
| Loaded hourly cost | Customer-supplied | Client finance: salary plus employer costs |
| Adoption in the first months | Estimate | Stated as an estimate until usage data replaces it |
| Error and rework rate | Unknown | Labeled Unknown until measured, never assumed to be zero |
This discipline cuts both ways: it also bounds what webvise itself may claim. One stopwatch number I can reproduce beats ten I cannot, and prospects notice the difference.
Price the run in cents, then compare it to the seat
Run cost is the input teams guess at most and measure least, and it is the easiest one to log. webvise routes every client automation through one gateway with a curated model list, so each run gets a logged price. The tender conversion costs $1.50 to $2.50 on the strongest models. Routine runs target 20 to 30 cents.
During development, a test token budget of about €50 covers the experimentation. In production, the monthly AI cost of a client automation is typically a two-digit euro amount, with no per-seat licenses. Compare that to AI subscriptions at €100 to €200 per user per month, which scale with headcount whether anyone processes a document or not.
The full cost side, build price included, is broken down in what AI automation costs in 2026. For the ROI calculation, the run price is the line that turns hours saved into net savings, so log it per run from day one.
Measure quality with the same stopwatch
Speed is half a business case. On a quantity takeoff from construction plans, the same client's estimator measured a surface by hand at 95 m², and the model read 92 m² from the plan. That comparison, 92 m² AI versus 95 m² manual hand measurement, went into the evaluation as a measured quality delta, right next to the time saving.
Publishing the deviation makes the review gate concrete. A 3 m² gap on one measurement tells the estimator how much checking the output still needs, and the business case can price that checking time honestly. A case that claims zero errors has skipped a measurement, and finance should ask why.
The two document workflows behind these numbers, tender conversion and quantity takeoff, are walked through in AI document automation for small businesses, including what the setup looks like in production.
Run the measurement this week. Pick one document workflow, time one real document through both paths with its owner present, log the run price, and classify every remaining input before anyone builds a slide. If the net number stays positive with the labels showing, you have a case. If it only works after an assumption gets promoted to a fact, you caught that before it cost you a build.
webvise builds AI automations for exactly these document workflows and measures them the same way, before the build and after. If you want a candidate workflow timed and priced before you commit budget, get in touch.
Development practices are aligned with ISO 27001 and ISO 42001 standards.