You know the scene. Orders travel by email, stock lives in a shared spreadsheet, supplier quotations are rebuilt from memory each time. Then the quote for an off-the-shelf system arrives, with per-user licences and add-on modules, and the project stops there. For years the sensible answer was to accept it. It no longer is. Custom operations software now costs a fraction of what it cost five years ago, because AI-assisted programming has cut the work required to build it. This dossier sets out where business process automation produces measurable returns, and how to judge whether it pays in your company.
Key points
- The digital gap is structural: in 2025, 41% of small EU enterprises used an ERP against 89% of large ones.
- The real cost of a packaged system is not the licence. It is fitting your company to someone else's workflow.
- Controlled studies put the time saved in writing code at 35-55%, which brings a dedicated application back within reach.
- The processes worth automating first are high volume with stable rules: quotations, stock, procurement, shipping.
- For niche trade, finance and logistics operations, no market product exists at any price.
The digital gap the figures reveal
The numbers describe a clean fracture. Eurostat data for 2025 show ERP systems in 41% of small EU enterprises and in 89% of large ones. CRM adoption runs at 25% against 65%. Business intelligence tools sit at 11% against 69%. Just over half of all EU enterprises, 53%, used any specialised e-business software at all.
The digital gap by company size
EU, 2025 — share of enterprises using each tool
Chart data
| Item | Small enterprises (10-49 employees) | Large enterprises (250+ employees) |
|---|---|---|
| ERP | 41% | 89% |
| CRM | 25% | 65% |
| Business intelligence | 11% | 69% |
Italy follows the same pattern. The Istat survey published in 2025 records ERP use at 48.8% among firms with 10 to 49 employees, against 85.9% among firms above 250. For CRM the figures are 21.1% and 56.5%.
That 58-point gap on business intelligence is the most telling figure. It does not measure access to software. It measures who can read their own numbers and who cannot. A company that does not know which customer earns money and which loses it is not being frugal. It is flying blind.
The picture is not one of neglect. Research from the Politecnico di Milano SME observatory, released in 2026, found that more than half of Italian SMEs increased digital spending in 2025. Yet 76% have neither invested in artificial intelligence nor plan to. Spending grows while direction is missing. That is precisely the space where a focused project creates advantage, because firms of your size are not moving.
Why the packaged system costs more than it looks
The list price is the visible part, and the least important. The true cost of a generalist system has three components that never appear in the quotation.
The first is adaptation. A standard package imposes its own workflow. If your supplier qualification runs in three steps and the software expects five, you pay the difference in working hours, every day, for years.
The second is per-seat licensing. Zylo's 2025 SaaS Management Index puts average software-as-a-service spending at 4,830 dollars per employee, up 21.9% year on year. Companies under 500 employees run 152 separate applications on average. Unused licences account for substantial sums even in mid-sized organisations.
The hidden cost of per-seat licensing
Chart data
| Indicator | Value |
|---|---|
| Average SaaS spend per employee | 4,830 $ |
| Applications in use | 152 |
| Annual spend growth | 21.9% |
The third is fragmentation. When CRM, warehouse and shipping come from three vendors, data is reconciled by hand or through integrations that must be maintained. Every manual step is a point of failure. Each vendor also upgrades on its own schedule, so an integration that worked in March can break in October without anyone touching your side of it.
There is a fourth cost, harder to see. A packaged system decides what you are allowed to record. Anything it does not foresee ends up in a notes field, where it cannot be searched, counted or reported. Over a few years that is where most of your operational knowledge quietly accumulates.
Custom operations software reverses all three. The process comes first, so training shrinks. There are no per-user licences, because the application is yours. Modules talk to each other by design rather than by later integration.
What actually changed in the cost of building
Until a few years ago the arithmetic was simple and unfavourable. A dedicated system meant months of analysis and development at software-house rates. That cost has compressed, and the measurements are public.
The controlled study by Peng and colleagues, published in 2023, compared developers with and without a programming assistant. Those using one finished the task 55.8% faster. Three randomised experiments run by Microsoft, MIT and Accenture across 4,867 developers found a 26% average increase in tasks completed per week. Benefits were largest for less experienced developers. McKinsey's 2023 research estimated 35-45% reductions in the time to write new code and 45-50% in documentation.
How far the cost of writing code has fallen
Chart data
| Indicator | Value |
|---|---|
| Faster task completion | 55.8% |
| More tasks completed per week | 26% |
| Less time to write new code | 45% |
What changed in the cost of building an application
- 2020Fewer than 25% of new applications use low-code platforms (Gartner)
- 2023First controlled study: tasks completed 55.8% faster with an AI assistant
- 2024DORA: AI adoption improves quality and review speed, but not release stability
- 202584% of developers use or plan to use AI tools
- 2026Gartner projected 75% of new enterprise applications on low-code
Chart data
| Year | Event |
|---|---|
| 2020 | Fewer than 25% of new applications use low-code platforms (Gartner) |
| 2023 | First controlled study: tasks completed 55.8% faster with an AI assistant |
| 2024 | DORA: AI adoption improves quality and review speed, but not release stability |
| 2025 | 84% of developers use or plan to use AI tools |
| 2026 | Gartner projected 75% of new enterprise applications on low-code |
Adoption is now general. The Stack Overflow Developer Survey 2025 reports that 84% of developers use or plan to use AI tools, and that 51% of professionals use them daily.
One honest qualification belongs here. The same survey finds only 3.1% of developers highly trust the accuracy of these tools. Google's DORA report for 2024 associates higher AI adoption with an estimated 1.5% fall in delivery throughput and a 7.2% fall in stability. In plain terms: writing code got cheaper, designing and testing did not. The saving concentrates on construction, not on understanding the process.
The four processes that repay first
Not everything should be automated. The best candidates share three traits: high volume, stable rules, and data that already exists inside the company. Four areas almost always qualify.
Commercial relationships. A CRM built on your sales cycle records the deal, not just the contact. If you sell abroad, the record must carry the customer's language, the customs regime, the usual incoterm and the quotation history. No generalist CRM does that without paid customisation.
Warehouse. Real-time stock removes two opposite costs: tied-up capital and lost sales. For a firm with a few thousand item codes, a dedicated module linked to transport documents beats a full WMS that is never fully configured.
Procurement. Requests for quotation, bid comparison, contract deadlines. This is where manual error costs most, and where traceability also protects you in a dispute.
Shipping and documents. Packing lists, certificates of origin, proforma invoices. Each document draws on the same data, retyped several times.
Packaged system and custom application compared
| Process | Off-the-shelf package | Custom application |
|---|---|---|
| Commercial relationship | Generic contact record; language, incoterm and customs regime as free text | Deal record with language, incoterm, customs regime and quotation history |
| Warehouse | Full WMS, often only partly configured | Stock linked to the transport documents actually issued |
| Procurement | Catalogue purchasing module, customisation charged separately | RFQs, bid comparison and deadlines built on the real process |
| Shipping and documents | Fixed-format exports, data retyped | Packing lists, certificates and proformas generated from the same data |
| Cost model | Per-user licence, add-on modules, rising subscription | One-off development cost, plus maintenance and hosting |
| Ownership | Vendor's; data portability to be negotiated | Client's, with exportable code and data |
On these four areas, business process automation is not an IT project. It is a reorganisation of work that happens to use software. That is why the custom operations software service always starts from the real process rather than a feature list.
Niche operations, where no product exists
One category of work has no market answer at any price: niche operations. These are processes affecting a few hundred companies worldwide, often in a single sector and a single region.
Take supplier due diligence before payment. Verifying a producer means checking registry data, VAT status, certifications and banking details in a defined order, and keeping the evidence. The method behind how to verify an Italian manufacturer is a checklist that no packaged CRM contains, and that a small application can enforce on every deal.
Documentary finance works the same way. A letter of credit has deadlines, required documents and tolerances that admit no approximation. An application that derives the document list from the credit text, then flags the deadlines, costs a few weeks of development. It prevents discrepancies worth tens of thousands of euro.
Logistics repeats the pattern. Alternative routings, customs transits and country-specific health certificates generate data that standard tools cannot capture in structured form, as anyone running quality control and logistics support discovers quickly.
The working rule follows. The more specific your process, the less likely a suitable product exists, and the stronger the case for building one. Niche used to be a penalty in software. Today it is the ideal case.
How to judge the return honestly
Treat published automation returns with caution. Many come from platform vendors rather than independent studies. An internal estimate is worth more than any percentage read elsewhere.
Work through it in order. Pick the candidate process and measure, over two weeks, how many hours it truly absorbs, corrections included. Multiply by the fully loaded hourly cost. Add the cost of last year's errors: returns, penalties, reshipments, discounts granted to fix a problem. That total is the gross annual benefit.
An example makes the method concrete. A twelve-person firm measures forty hours a month on preparing export quotations. At a fully loaded 35 euro an hour, that is roughly 16,800 euro a year. If the application removes sixty per cent of it, the gross benefit approaches 10,000 euro. Against 12,000 euro of development and 1,500 a year in maintenance and hosting, the payback lands a little past the first year. These are illustrative figures, not a price list. The point is that the calculation runs on four numbers you already hold, rather than on a percentage taken from a vendor brochure.
Then book the real costs. Initial development is only the first. Add maintenance, hosting, training and the item almost nobody counts: your own time spent explaining the process to whoever builds the tool. A project that repays within twelve months is sound. Beyond twenty-four, cut the scope.
Two further costs deserve a line each. Retiring the spreadsheets the tool replaces takes real hours, and so does cleaning the data before it moves across. Budget for both, or the first month in production will feel like a failure when it is only a migration.
Three risks belong in the contract, not in a technical annex. The first is ownership of the code, which must be yours in writing. The second is data portability, in an open format you can export at will. The third is dependence on one person, which documentation and repository access remove. Custom operations software is an asset only if you can walk away with it.
Where to start: four moves
Digitalisation usually fails for one reason. It starts from the software instead of the process. The sequence that works runs the other way.
First, map. Draw the path of one order on paper, from enquiry to delivery. Mark every handover and every point where information is retyped. Two hours and a large sheet are enough.
Second, choose one process. The one with most repetitions and most errors. A project that opens on three fronts never reaches production.
Third, build the minimum version. The first release covers the normal case, not the exceptions. Exceptions are added once the main flow is in daily use and producing results.
Fourth, measure and extend. After three months, compare actual hours against your baseline. Only then decide whether to widen the scope.
A note on sequencing for exporters. If you are still deciding where to sell, business process automation is premature. Market choice comes first, then the tool that supports it. The same applies to sourcing: reading a guide to Italian medical device manufacturers tells you what data your system will have to hold before you commission it.
Conclusions
The gap between small and large firms in management software has not closed. The economic reason behind it has. Building a dedicated application now costs a fraction of what it did before AI-assisted programming, while per-seat pricing keeps rising. The calculation has flipped, above all for companies running specific processes that no market product covers. The practical next step is not choosing software. It is measuring, over two weeks, what your most repetitive process really costs you. That single number decides everything else.
Data current to September 2026.
