Contents8
- What exactly did Santander announce?
- Where does the value come from?
- Why do so few companies see a return on AI?
- How can a smaller company measure AI ROI?
- Which project should you start with for a fast return?
- Should you be wary of the numbers big companies announce?
- What this means for your business
- Frequently asked questions
On October 28, 2026, when it published its results for the first nine months of the year, Banco Santander put a number on the value created by its artificial intelligence projects: €194 million, made up of extra revenue, avoided costs and lower credit losses. The bank now reports this figure every quarter. For a small or mid-sized business, the amount itself is out of reach and beside the point. The method is what matters: Santander shows that AI ROI in business can be measured, as long as you decide what you will count before the project starts.
What exactly did Santander announce?
In its third quarter results press release, the bank says it kept rolling out AI across its businesses and generated €194 million of value so far this year. The results presentation for analysts explains what the figure covers: a year-to-date total that adds up incremental revenue, cost savings and lower loan-loss provisions, that is, the money a bank sets aside to cover loans that may not be repaid.
The number is growing fast. Santander reported €35 million in the first quarter, then €84 million at the end of the first half according to its first half institutional presentation. From those running totals, we work out that roughly €110 million was added in the third quarter alone. The bank was aiming for more than €200 million for the year and has set a target of over €1 billion in AI value between 2026 and 2028.
Keep it in proportion, though. The same release reports underlying profit of €11.2 billion for the nine months. By our calculation, the value attributed to AI is under 2% of that. It is not a transformation of the accounts. It is a lever that is starting to carry weight.
Where does the value come from?
Santander described its main workstreams in June 2026, in a piece titled "Santander turns its AI-first strategy into measurable impact". Three families of use cases stand out, and they apply to almost any company:
- Internal process automation. The bank says it has more than 280 automation agents in production, covering credit, fraud, customer identity checks and operations. In Brazil, handling card fraud claims is reportedly about 95% faster, with up to 90% automation and an error rate below 1%.
- Customer service. In the UK, Santander is starting to roll out AI on its phone channels for card-related queries, with a target of around 240,000 calls a year resolved through self-service, or 40% of the annual volume of those calls. That is a target, not yet a result.
- Tools for employees. AI tools were opened up to all 185,000 staff, up from around 40,000 active users before.
On October 19, 2026, the bank added a point about running costs. In its release on its collaboration with NVIDIA, it says that in experimental use cases the cost per interaction fell from about €5 to about €0.03, and that a customer service test cut costs by more than 50% by switching to a more efficient model. The release is explicit that these are experiments, not bank-wide figures.
Why do so few companies see a return on AI?
Santander is still the exception. McKinsey's annual survey, "The state of AI in 2026", run from May 4 to June 8, 2026 with 1,719 respondents in 97 countries, finds that 37% of organizations see a positive contribution from AI to their operating profit (EBIT), essentially unchanged from 2025. Only about 6% attribute at least 5% of EBIT to AI. And one in five says it is limiting AI use because of operating costs.
Yet eight in ten respondents say AI has improved their own productivity. That is the gap: an employee who saves twenty minutes a day writing emails does not create a visible line in the accounts. Saved time evaporates unless someone decides what it is for.
Our take: the difference between Santander and most companies is not technology. According to its own communication, the bank relies on market models from OpenAI, Anthropic and Google. The difference comes down to three choices. It targets specific processes rather than diffuse usage. It measures in euros, with a definition that stays stable from one quarter to the next. And it publishes the number, which forces its teams to justify it. These are the same reasons so many AI projects never make it past the pilot stage.
How can a smaller company measure AI ROI?
Santander's framework translates well to a company of 30 or 300 people once you simplify it. Three sources of value, each with a metric you can track without a finance team:
| Source of value | Example use case | Metric to track |
|---|---|---|
| Avoided costs | Agent that answers tier 1 support questions | Requests resolved without a human, multiplied by the cost of a request handled by the team |
| Extra revenue | Sales agent that qualifies and follows up leads | Meetings booked and revenue signed on leads handled by the agent |
| Reduced risk | Automated checks on invoices, contracts or files | Errors caught before sending, disputes and penalties avoided |
| Freed-up time | Search across internal documents | Hours saved, only if they are reassigned to a defined task |
Four rules keep the calculation honest:
- Measure before you deploy. Without average handling time, monthly volume and the current cost of a request, no gain can be proven. Two to four weeks of tracking is enough.
- Count the full cost. Licenses, model usage, integration with your tools, maintenance and the time spent steering the project. We break these down in our analysis of the real cost of an AI agent.
- Only count saved time if it is reused. An hour saved and not reassigned has no accounting value. Decide in advance what the team will do with it.
- Keep the same definition over time. Santander adds up revenue, costs and provisions using a fixed rule. Changing the method at every review makes the numbers useless.
Which project should you start with for a fast return?
A repetitive, high-volume process that is already measured, where mistakes are cheap. Tier 1 customer support often ticks every box: questions repeat, volume is easy to count, and a human takes over when the agent is unsure.
That is the case at Peps Digital, a software company serving home healthcare providers, for which we built an assistant that answers product questions from the documentation. According to its CEO, 80% of support requests are now handled without human intervention, across more than 1,200 questions a month, with a 92% satisfaction rate. The ROI calculation then becomes simple: the volume absorbed by the agent, compared with the cost of that volume handled by the team. Our approach is described on our page about AI agents for customer support.
On the other hand, we advise against starting with a company-wide assistant rollout in the hope of measuring the gain later. It is useful, and Santander did it too, but that is not where the bank finds its euros. It finds them in targeted processes such as fraud claims or customer service.
Should you be wary of the numbers big companies announce?
Read them with method rather than dismissing them. Santander's figure is self-reported, and neither the release nor the presentation breaks it down by use case or explains how value is attributed to AI rather than to other factors. Lower provisions, for example, can depend as much on the economy as on a better credit scoring model.
It is the same reflex we applied to PayPal's customer service results, which we analyzed in early November: a volume or value figure is a signal, not proof. The right questions for any published case are always the same. What was the starting point? Which costs were deducted? What is a result, and what is still a target?
What this means for your business
Pick a process, not a technology. List two or three repetitive tasks that eat time every week, put a number on their volume and current cost, and keep only the one where the gain is easiest to prove.
Write the calculation rule before any development. One line is enough: "value = requests resolved by the agent multiplied by the cost of a human-handled request, minus the full cost of the agent". It will save you arguments at the end of the project.
Review the numbers every quarter. Like Santander, track the same measure over time. If the curve is not rising after two quarters, stop or redirect, rather than paying for a tool nobody can show a return on.
Frequently asked questions
How do you calculate the ROI of an AI project?
Add up the measured gains (avoided costs, extra revenue, reduced risk) over a period, subtract the full cost of the project over the same period (development, licenses, usage, maintenance), and divide by that cost. The most important step is measuring the starting point before you deploy, otherwise the gain remains an estimate.
How long does it take for an AI project to pay off?
It depends on the volume handled and the starting cost. An agent that absorbs hundreds of repetitive requests a month can cover its cost within a few months, while a tool used occasionally may never do so. That is why we recommend starting with a high-volume process.
Is Santander's €194 million all cost savings?
Only partly. According to the bank's presentation, the figure adds up extra revenue, cost savings and lower loan-loss provisions. The split between those three sources is not published.
Can a small business get a return on AI without a data team?
Yes, if the project is focused. A support or automation agent plugs into your existing tools and can be measured with simple metrics such as requests resolved or handling time. A data team becomes necessary when you want to industrialize dozens of use cases, which is Santander's scale, not that of a first project.
Want to look at your case together? Book a slot, we'll set aside 30 minutes to find the process where AI will pay off fastest for you and set the rule for measuring its ROI.
Published by Anthony Cohen
← All insights