How to evaluate document automation ROI before choosing an IDP software
Document automation ROI is calculated by comparing the annual cost of manual data entry with the total cost of automation, then dividing the one-off setup by the net monthly saving to get the payback period. In practice: current cost per document = (minutes per document ÷ 60) × fully loaded hourly cost; cost of automation = licence or subscription + consumption (per document, per page or in credits) + exception handling cost + amortised setup. With typical Italian SME values — 4-6 minutes and €18-22/hour — a manually keyed document costs between €1 and €3 in labour alone: that is the cost automation lets you avoid. The number that really decides the outcome, though, is not the accuracy a vendor states but the straight-through processing rate: the share of documents that reach the ERP without anyone touching them. Below are the full formulas, three worked scenarios and the checklist of what to measure before signing.
The eight business case formulas
These are all the formulas you need, in calculation order. You can put them in a spreadsheet in ten minutes: the point is to walk into a vendor conversation already holding your own numbers instead of having their sales team produce them for you.
Current cost per document
(minutes per document ÷ 60) × fully loaded hourly cost
The hourly cost must be the company cost, including contributions and overheads, not net pay. In Italian SMEs it typically sits between €18 and €22 per hour.
Annual cost of manual data entry
documents/year × current cost per document
Count every document type that would fall in scope: supplier invoices, delivery notes, customer orders, order confirmations, credit notes, price lists. Many companies miss half the volume because it sits in a different department.
Exception handling cost
documents/year × exception rate × (review minutes ÷ 60) × hourly cost
No solution takes human work to zero. The exception rate is the share of documents that still needs a check: it is the parameter optimistic business cases omit, and the one to measure during a trial.
Annual cost of automation
recurring licence or subscription fees + consumption (documents/year × cost per document, page or credit) + exception cost
Licences and subscriptions must be counted in full in low-volume months too, because you pay them anyway; the consumption component, by contrast, follows real volumes. Budget for both using the figures in the quote.
Net annual saving
annual manual data entry cost − annual automation cost
This is the number that feeds everything else. If it is negative, the project does not stand on labour alone and has to be justified on other axes (closing times, early payment discounts, error reduction).
Payback in months
one-off setup cost ÷ (net annual saving ÷ 12)
Setup should include ERP integration, configuration and training. A payback under 12 months is generally considered solid; above 24 it is worth renegotiating the scope.
Three-year ROI
(cumulative 3-year saving − total 3-year cost) ÷ total 3-year cost × 100
Three years is the right horizon for document software: long enough to amortise the integration, short enough to stay credible.
Hours freed per year
documents/year × minutes per document ÷ 60 × operational time reduction (60%, Data Alchemy's average)
This is the metric that convinces people who do not think in euros. It does not automatically translate into headcount reduction: in most projects it becomes capacity to handle more volume with the same team.
If you would rather not build the spreadsheet, the ROI calculator applies these same formulas starting from five values: minutes per document, data entry hours per day, number of people, hourly cost and working days per year.
The six numbers to have before talking to a vendor
The quality of the business case depends entirely on these six values. Collecting them takes half a day and completely changes the tone of the negotiation.
Annual volume by document type
Not a single total, but the breakdown: how many supplier invoices, delivery notes, customer orders, order confirmations. Many commercial models charge per document or per page, and extraction difficulty varies enormously between a delivery note and a price list.
Actual minutes per document
Measured, not estimated. Time ten real documents including opening the attachment, reading it, typing, checking master data and correcting. The real figure is almost always higher than the perceived one.
Fully loaded hourly cost of the people involved
Gross pay plus contributions and overheads, for the people actually doing the work today. If data entry is handled by qualified finance staff, the hourly cost is higher than the default most people use.
Current error rate and cost of an error
How many postings have to be corrected downstream and what correcting them costs: credit note, chasing, wrong payment, dispute with the supplier. It is the most underestimated line and often the largest.
Seasonal peaks
How many documents arrive in the worst month versus the average. Peaks are what generates overtime, late closes and temp staff today: in a business case they matter more than the annual average.
Current cycle times
How many days pass between an invoice arriving and being posted. This is the number that links automation to measurable financial benefits, such as capturing early payment discounts or closing the month on time.
Three scenarios, from SME volumes to group volumes
The scenarios calculate the gross labour saving at a €20 hourly cost by applying a 60% reduction in operational time, Data Alchemy's average figure: the manual work that remains, exceptions included, is 40% of today's. The result should be verified in a trial on real documents. The Data Alchemy licence cost is not included, because it depends on the licence you choose (Basic, Business or Enterprise) and on pay-as-you-go credits: it is quoted on request.
| SME · 12,000 docs/year | Mid-market · 48,000 docs/year | Group · 240,000 docs/year | |
|---|---|---|---|
| Documents per year | 12,000 | 48,000 | 240,000 |
| Minutes per document (today) | 5 | 6 | 4 |
| Fully loaded hourly cost | €20 | €20 | €20 |
| Manual cost per document (F1) | €1.67 | €2.00 | €1.33 |
| Annual data entry cost (F2) | €20,000 | €96,000 | €320,000 |
| Operational time reduction (Data Alchemy average) | -60% | -60% | -60% |
| Residual manual work, exceptions included (40% of F2) | €8,000 | €38,400 | €128,000 |
| Gross labour saving (60% of F2) | €12,000 | €57,600 | €192,000 |
| Hours freed per year (F8) | 600 hours | 2,880 hours | 9,600 hours |
| Licence and pay-as-you-go credits (F4) | Quote on request | Quote on request | Quote on request |
To reach the net saving (F5), subtract the licence and credits cost in the quote you receive from the gross saving; payback (F6) is then the setup cost, which varies with the ERP and the number of document flows, divided by the net monthly saving. None of the scenarios includes the cost of errors avoided, which always works in favour of automation.
Why stated accuracy is not enough: look at the straight-through processing rate
Two vendors can state the same accuracy and produce very different ROI. Accuracy measures the share of fields extracted correctly; the straight-through processing (STP) rate measures the share of documents that goes through the entire flow and lands in the ERP without anyone opening them. Labour savings depend on the second number, not the first.
An example makes the difference visible. On an invoice with twenty fields, 99% per-field accuracy means roughly one document in five contains at least one wrong field — and a document with one wrong field is a document someone has to open. Going from 99% to 99.8% per field is not a marginal 0.8% improvement: it moves the share of perfect documents from about 82% to about 96%, cutting the residual work more than fourfold. That is where the economics live.
So during evaluation the right question to put to every vendor is not "what is your accuracy" but "on a sample of my documents, what share was posted with no human intervention, and how did you measure it". If the trial does not produce that number, the business case is still a hypothesis.
How the pricing model changes ROI at identical technology
Even before extraction quality, the commercial model determines how much of the theoretical saving you actually keep. You will meet three families.
Licence or annual subscription
Typical of enterprise suites. The cost is fixed and payable even in months when volumes drop or the project is paused. It must be counted in full from year one regardless of documents actually processed, with the cost of extra modules and named users made explicit.
Per-page pricing
Typical of cloud provider APIs. It looks cheapest until you measure average document length: an invoice with ten pages of line items costs ten times a one-page invoice. Always compute the cost on your real mix, not on a sample document.
Licence plus pay-as-you-go credits
Data Alchemy's model: a licence chosen to fit your needs — Basic, Business or Enterprise — plus pay-as-you-go credits for the documents processed, quote on request. 1 credit = 1 document, regardless of the number of pages: consumption is estimated on the number of documents, not on their length. In the business case the licence goes in as a recurring cost and the credits as a function of real volumes: keeping the two components separate shows exactly how much of the cost depends on the volume assumption.
A useful test: redo the ROI calculation assuming real volumes turn out to be half of those forecast. The consumption component shrinks with volumes, while licences and subscriptions stay whole: the more of the cost sits in the fixed part, the more exposed the return is to a wrong volume assumption. It is the fastest way to see who carries the risk of the volume assumption.
The solutions that will land on your table, and the ROI question to ask each
A business case is built against concrete alternatives, not against an abstract category. These are the solutions that almost always end up on the same short list, grouped by pricing family — which drives the financial return more than the technology does. We do not reproduce their price lists: they depend on contract, volume and modules, and a number copied from a public page would make your calculation wrong. What we do give you is the question that reveals the most about ROI.
| Solution | Pricing family | The question that reveals most about ROI |
|---|---|---|
| Data Alchemy | Basic, Business or Enterprise licence plus pay-as-you-go credits (1 credit = 1 document, regardless of the number of pages), quote on request. | Ask us for the share of documents posted with no human intervention on a sample of your own documents, plus a quote for integrating your ERP. Those are the two lines that move the payback. |
| ABBYY | Enterprise IDP suite, typically licence-based with modules and on-premise deployment options. | What is the three-year total cost including modules, named users and maintenance, and how many configuration days are needed before the stated accuracy shows up on our layouts? |
| Rossum | Subscription AP automation platform, generally with annual volume tiers. | What happens to the subscription if volumes stay below the tier purchased, and at what rate is any overage billed? |
| Google Document AI | Consumption-based cloud service, typically priced per page processed. | What is the cost on our real document mix, calculated on actual average length? An invoice with ten pages of line items costs ten times a one-page invoice. |
| Amazon Textract | Consumption-based AWS service, priced per page and differentiated by type of analysis. | Does extracting the fields we need require the basic analysis or the specialised one, and is writing into the ERP included or a separate integration project? |
| Docsumo | Subscription document AI platform with volume tiers. | Is review of doubtful documents included in the plan or charged separately, and what exception rate is the tier you are proposing calibrated on? |
| Esker · Basware · Yokoy | Procure-to-pay automation suites, licence or subscription, with an implementation project. | How much does implementation weigh in year one relative to the subscription, and how much of the claimed saving depends on modules we had not planned to buy? |
| Outsourced data entry | Per-document or hourly service, delivered after the document arrives rather than on arrival. | What is the guaranteed turnaround in the peak month, and does the rate change above a given volume threshold? |
Note: ABBYY, Rossum, Google Document AI, Amazon Textract, Docsumo, Esker, Basware and Yokoy are trademarks of their respective owners. The pricing families listed here describe the prevailing commercial model, not a price list: always verify current terms directly with the vendor, since they vary by volume, modules and contract. And if your volume is a few hundred documents a year, the honest answer is that none of these solutions pays for itself — ours included.
The five most common ROI calculation mistakes
These are the mistakes that make a business case untrustworthy — in both directions. An inflated ROI gets taken apart by finance; an over-cautious one blocks a project that would have paid for itself.
Counting labour only
Typing time is the easiest line to quantify, not the biggest. Posting errors, wrong payments, credit notes, chasing, document searches and late closes often weigh more than transcription, but end up outside the spreadsheet because nobody measures them.
Assuming 100% automation
A business case that assumes zero human intervention is wrong by construction. Budget a realistic exception rate, which depends on inbound document quality, and verify it in a trial.
Forgetting integration and maintenance
Extraction cost is only part of it. Estimate the days for ERP integration, field mapping, write-error handling and ongoing maintenance. Anyone using template-based OCR must add rebuilding templates every time a supplier changes layout.
Using the average instead of the peaks
The value of automation shows up in the worst month, not the average. If year-end requires overtime or temp staff to clear the backlog, that cost belongs in the calculation — and it is often the line that closes the business case.
Confusing hours freed with cost cuts
Hours freed become an accounting saving only if someone stops working or the company avoids a hire. In most cases they become additional capacity instead: a real benefit, but one that has to be presented as such, or finance will rightly reject it.
How to evaluate ROI before signing: six steps
This is the sequence to follow while you are still comparing vendors. Each step produces a figure that feeds the formulas above.
Map the real document flows
List document types, arrival channels (email, portal, scanner), monthly volumes and who handles them today. Without that map, any quote is priced against an imaginary scope.
Measure the current cost with F1 and F2
Time ten documents per type and apply the first two formulas. You get the annual manual data entry cost — the benchmark every offer has to be judged against.
Run a trial on your documents, not on a demo
Ask every vendor to process the same sample of your real documents, ugly cases included: skewed scans, new suppliers, invoices with many lines. A demo on documents the vendor picked measures nothing.
Measure the straight-through processing rate
On the sample, count how many documents would have been posted with no intervention at all. That percentage, not the stated accuracy, is what goes into formula F3 as the complement of the exception rate.
Ask for the three-year total cost, not the unit price
Add up licences or consumption, setup, integration, extra modules, users and maintenance over three years. Compare vendors on that total: it is the only number comparable across different commercial models.
Compute payback and ROI, then test the worst case
Apply F5, F6 and F7, then redo the calculation halving forecast volumes and doubling the exception rate. If the project still holds, the business case is solid; if it vanishes, you know exactly which assumption is carrying it.
Eight questions to ask every vendor during evaluation
The answers to these are worth more than any product sheet, because they concern exactly the variables that feed the ROI formulas.
- 01On a sample of our documents, what share is posted with no human intervention at all, and how do you measure that number?
- 02Is the stated accuracy computed per field or per document, and does it include line items or only the header?
- 03Is pricing per page or per document, and what happens to a ten-page invoice?
- 04What exactly does the one-off setup include, and which recurring costs remain after year one?
- 05How many days does integrating our ERP take, and who performs the work?
- 06What happens when a supplier changes layout: does anything need reconfiguring, and at whose cost?
- 07Where are documents processed and stored, for how long, and are they used to train models?
- 08If our volumes turned out to be half the forecast, how does what we pay change?
Related tools and pages
ROI calculator
Apply these formulas to your own numbers and get annual savings and hours freed.
Learn more →PayablesAccounts payable automation
Where manual data entry cost actually sits: the supplier invoice cycle end to end.
Learn more →PricingLicences and pricing
Basic, Business and Enterprise licences with pay-as-you-go credits: quote on request.
Learn more →AccuracyHow accuracy is really measured
Per-field accuracy, per-document accuracy and the straight-through processing rate.
Learn more →ComparisonBest data extraction software
The platform comparison on the criteria that actually move total cost.
Learn more →Case study130,000 documents a year
Agribologna's real numbers after automation: volumes, times and accuracy.
Learn more →Frequently asked questions about document automation ROI
How do I evaluate document automation ROI before choosing an IDP software?
In six steps. First, map the real document flows by type, channel and volume. Second, measure the current cost by timing ten documents per type and applying (minutes ÷ 60) × fully loaded hourly cost, multiplied by annual documents. Third, have every vendor process the same sample of your real documents, hard cases included. Fourth, measure the straight-through processing rate on that sample — the share of documents postable with no human intervention; that, not stated accuracy, drives the saving. Fifth, ask for the three-year total cost rather than the unit price. Sixth, compute net saving, payback and ROI, then repeat the calculation with half the forecast volumes to see whether the project still holds.
What is the document automation ROI formula?
Net annual saving is the difference between the annual manual data entry cost (documents/year × minutes per document ÷ 60 × fully loaded hourly cost) and the annual automation cost (licences or subscriptions + consumption per document, per page or in credits + exception handling cost). Payback in months is the one-off setup divided by the net annual saving divided by twelve. Three-year ROI is cumulative saving minus total cost, divided by total cost.
How much does it cost to key in a document by hand?
Counting labour alone, between €1 and €3: an operator takes 4-6 minutes per document on average and, at a fully loaded hourly cost of €18-€22, transcription alone is worth about €1.20-€2.20. On top of that sits the cost of data entry errors, which generates inefficiencies estimated in the 15-25% range: wrong payments, credit notes, chasing, redone reconciliations and disputes.
How long does a document automation project take to pay back?
Mostly a function of volume. At 12,000 documents a year, 5 minutes per document and a €20 hourly cost, applying the 60% reduction in operational time (Data Alchemy's average) the labour avoided is worth about €12,000 a year; at 48,000 documents a year it rises to about €57,600. Subtracting the licence and pay-as-you-go credits cost in the quote gives the net saving, and payback is the setup cost divided by the net monthly saving. At volumes of a few thousand documents a year the payback stretches and automation is better assessed on other benefits too, such as closing times and error reduction.
Does ROI depend more on accuracy or on the pricing model?
On both, but they act on different levers. Accuracy — or better, the straight-through processing rate — determines how much labour remains after automation: moving from 99% to 99.8% per-field accuracy on a twenty-field invoice cuts the documents someone has to open more than fourfold. The pricing model determines how much of the saving you keep and who carries the risk if volumes come in below forecast: the consumption component falls with volumes, while licences and subscriptions stay fixed.
Which costs are most often forgotten in the business case?
Four. Exception handling cost, because no solution automates 100% of documents. The days spent integrating the ERP and maintaining it afterwards. Rebuilding templates every time a supplier changes layout — a line that applies to traditional OCR, not to LLM-based systems. And the cost of current errors, which is almost always the biggest item and almost always the one nobody has measured.
How do I compare vendors with different pricing models?
By bringing them all to the same denominator: the three-year total cost on your real volume, including licences or consumption, setup, integration, modules, users and maintenance. Then repeat the calculation assuming half the forecast volumes: that is the test that separates models where volume risk is yours from models where it is the vendor's.
Does headcount have to be reduced for the ROI to be real?
No, and it is important to say so precisely in the business case. In most projects the hours freed become capacity to handle more volume with the same team, to absorb peaks without overtime and to spend time on review instead of typing. The saving should then be presented as a hire avoided or as added capacity rather than a headcount cut: it is more defensible in front of finance, and it is what actually happens.
Want the calculation run on your own numbers?
Bring us volumes, document types and minutes per document: we process a sample of your real documents and hand back a measured — not estimated — straight-through processing rate to complete the business case.
Request an assessment