More than 90% of ING's generative-AI pilots reach production, by the bank's own account. The comparator its head of advanced analytics strategy uses for the industry is 30%.
That gap is not luck. It is architecture — and the decisions ING made before the GenAI wave arrived. ING serves around 41 million customers, carries roughly €1.05 trillion in assets, and runs one of Europe's largest mortgage books at about €370 billion. When a bank that size builds AI, the choices it makes — platform, governance, sequencing — become the industry's reference case.
The outcomes ING publishes are rare in a sector where more than 70% of AI use-case announcements carry no ROI disclosure at all (Evident Insights, August 2025): inbound contacts needing human support down 18% versus 2021; mortgage "time to yes" under one day for 80% of cases; KYC due diligence from weeks to seconds; an 82% straight-through rate across 245 retail journeys. But the real case study is not the metrics. It is the structural decisions that made them achievable — and the board-level critiques that even ING's success story cannot fully deflect. The agentic mortgage assistant entered pilot only in March 2026, at a bank that has been one of Europe's largest mortgage lenders for years. That timing gap is the sharpest question in this case.
of ING's (Gen)AI pilots are successfully moved into production. Its head of advanced analytics strategy puts the industry average at about 30%.
ING investor presentation, 18 Mar 2026; Computer Weekly, 13 Jan 2026
risks against which every AI system is evaluated, in a 20-step process, before deployment. An AI committee that includes the CEO reviews each use case.
ING newsroom, 17 Nov 2024 (Bahadir Yilmaz); Reshaping Work, May 2025
to design and deploy the generative-AI customer chatbot in the Netherlands with McKinsey/QuantumBlack. It assisted 20% more customers than the classic chatbot.
McKinsey & Company client case (project launched Sept 2023)
of the 117 AI use cases announced by 50 tracked banks in H1 2025 reported no specific ROI. ING publishes hard numbers — with gaps this case file names.
Evident Insights Banking Brief, 7 Aug 2025
The problem: scale meets structural friction
ING closed 2025 with nearly 41 million customers, a presence in more than 100 countries, total assets of roughly €1.05 trillion, more than 60,000 employees and a net result of €6.3 billion (ING FY2025 results, February 2026). Its mortgage book stands at about €370 billion, which Euromoney and ING itself describe as one of the largest in Europe (Euromoney, July 2026; ING, March 2026).
The operational reality underneath those numbers, as McKinsey's case study describes it: in the Netherlands alone, ING handled 85,000 customer interactions a week by phone and online chat. The classic rule-based chatbot resolved 40–45% of chats, leaving around 16,500 customers a week who needed a live agent (McKinsey & Company, case study c. January 2024). Know-your-customer checks consumed operations headcount at scale, and mortgage origination relied on manual document gathering and cross-system checks.
The strategic pressure was the operating paradox every large bank faces: costs had to fall while service quality and regulatory compliance both improved at the same time. That is the paradox AI had to resolve.
Foundation: three decisions made before the GenAI wave
ING's own account of why its pilots ship — and where the account is a claim rather than a proof.
One centralised AI platform
"From the beginning, we invested in having one platform," Marco Li Mandri, ING's head of advanced analytics strategy, told Computer Weekly — the platform the bank uses to build most of its AI and all of its generative AI (Computer Weekly, 13 January 2026). Models are standardised centrally and deployed across markets rather than rebuilt per country. ING presents this as the structural reason its pilot-to-production rate sits above 90%. It is a plausible reading, and it is ING's; no outside study has tested the causal link.
Governance as a design principle
"At ING, we have a 20-step process which evaluates any AI system for 140 risks," chief analytics officer Bahadir Yilmaz said in November 2024 (ING newsroom). An internal AI committee that includes the CEO reviews each use case for benefits and risks (Reshaping Work, reporting ING's Sheila Gemin, May 2025). ING has also implemented a data-driven quality-assurance framework to contain hallucinations and malicious manipulation of prompts in the customer-facing layer (Computer Weekly, January 2026).
Sequencing by domain depth, not by technology hype
ING's own framing runs: more than ten years of machine learning; generative AI taken "from pilots to global scaling"; agentic AI as "streamlined bold experiments" (ING investor presentation, March 2026). Its COO has described concentrating on five core areas — KYC, hyper-personalised marketing, wholesale lending, chatbots and software engineering — "rather than fifty" (Tech Monitor, August 2025). The four-stage reading of this sequence below is my synthesis of that record, not an ING statement.
Deployment: three phases
Phase 1 — foundational machine learning. More than a decade of ML in transaction monitoring, fraud, credit scoring and personalisation, on the platform that later carried generative AI (ING, March 2026; Computer Weekly, January 2026).
Phase 2 — 2023–2024, generative AI in production. The Netherlands chatbot, designed and deployed with McKinsey/QuantumBlack in seven weeks and released first to 10% of customers on testing days; it assisted 20% more customers than the classic bot, cutting wait times (McKinsey). More than 5,000 software engineers now use AI as a peer-programming assistant; 5,000 employees have been trained in data fluency and generative AI (Computer Weekly, January 2026). Generative AI was extended to KYC and compliance work (McKinsey interview with COO Marnix van Stiphout, 2025; Tech Monitor, August 2025).
Phase 3 — 2025–2026, agentic AI in core operations. Agentic KYC that can answer up to 80% of the data points required for regulatory checks without contacting the customer (BFSI Insider and Computer Weekly, July 2025). An agentic mortgage assistant that entered pilot in March 2026 for applications that would normally need manual assessment, with gradual rollout in the Netherlands from June 2026 (ING newsroom, 8 June 2026; Computer Weekly, 10 June 2026). Voice-agent trials in Spain and Germany for routine contact-centre tasks (FStech, December 2025).
Method: nail it, then scale it
The chatbot is the clearest window into how ING works. The brief came from an audit of the legacy bot's failure modes; the build was seven weeks; human-in-the-loop was designed in, with customers able to reach a live agent at any point; and quality assurance against hallucination and prompt manipulation was built as a framework, not bolted on afterwards (McKinsey; Computer Weekly, January 2026).
For KYC, data readiness was a stated prerequisite. "The bank needs its data actualised so it can run in the model," van Stiphout told Computer Weekly in July 2025 — infrastructure before agents. And when the EU AI Act arrived, Li Mandri says, no structural changes were required, because the compliance elements were already embedded in the 140-risk vetting (Computer Weekly, January 2026). That is ING's claim about its own readiness; it is consistent with the record, and it has not been independently examined.
Friction: four challenges in the record
1. ROI measurement. At the Reshaping Work conference in October 2024, ING's Sheila Gemin, global head of corporate technology, described the cost implications of tools like Copilot and an ongoing internal debate about measuring productivity; ING was cautious about expanding the rollout without clear ROI indicators (Reshaping Work, May 2025).
2. Data readiness as a hard prerequisite. Actualised data before agentic models — van Stiphout's condition for trusting agents in production (Computer Weekly, July 2025).
3. Hallucination and adversarial prompt risk. The dedicated quality-assurance framework exists because the risk was real in the customer-facing layer (Computer Weekly, January 2026).
4. Workforce change. "We do have to be mindful of the social consequences," van Stiphout said. ING launched a new data-fluency curriculum and frames the shift as moving people "from where they are today to a new type of job" (Tech Monitor, August 2025). He expects roughly 25% fewer people per operations process where AI is introduced — an expectation, not a measured outcome (BFSI Insider, July 2025).
Outcomes: the exhibits ING has published
All figures from ING's Morgan Stanley European Financials Conference presentation, 18 March 2026, unless stated.
| Metric (year end) | 2023 | 2024 | 2025 |
|---|---|---|---|
| Inbound contacts handled with human support, reduction vs 2021 | 43% | 26% | 18% |
| Chatbot deflection (chats handled without human support) | 75% | 54% | 46% |
| Digi-index: average straight-through rate across 245 retail journeys | 71% | 78% | 82% |
The two customer-service series are falling; the straight-through series is rising. ING gives no explanation for the first two in the presentation. That matters for the critique below.
Mortgage "time to yes" — from income application to irrevocable binding offer, excluding customer time — for 80% of cases.
ING, 18 Mar 2026
Customer due diligence for standard cases, down from days or weeks; up to 80% of required KYC data points answered without contacting the customer.
BFSI Insider / Computer Weekly, 1 Jul 2025 (COO van Stiphout)
reduction in customer complaints, listed under customer experience. Period not disclosed.
ING, 18 Mar 2026
(~3%) of incremental cost efficiencies targeted for 2026, operations-driven, with about 1,250 fewer operations FTEs expected. Not attributed to AI specifically.
ING, 18 Mar 2026
Analysis: three structural advantages
Platform before use cases
Building one AI platform before the generative wave — rather than retrofitting infrastructure mid-transformation — is the most plausible explanation for a pilot-to-production rate above 90%. Peers rebuilding data infrastructure while trying to deploy agents are paying a sequencing penalty ING avoided. This is the reading ING itself offers; the evidence is consistent with it, and it has not been tested against a counterfactual.
Governance as a velocity driver, not a velocity killer
A 20-step, 140-risk gate sounds like drag. In a supervised institution it is the opposite: it removes post-launch regulatory remediation, which is the true velocity killer. ING's statement that the EU AI Act required no structural changes is the strongest evidence the gate was well calibrated (Computer Weekly, January 2026).
Domain depth as the selection filter
Five areas rather than fifty concentrated the programme where ING has the deepest data: mortgages, KYC and the contact centre. That is what prevented the "GenAI everywhere" fragmentation that has produced near-zero return elsewhere — MIT's 2025 study found 95% of organisations getting zero measurable return on enterprise generative AI (MIT NANDA, State of AI in Business 2025).
The critique: four questions ING's narrative does not answer
The operator's view, not ING's. Every question is answerable from ING's own disclosures — if it chose to.
1. Sequencing lag on the highest-value product. ING is one of Europe's largest mortgage lenders. The agentic mortgage assistant entered pilot only in March 2026 — years after the generative-AI infrastructure was production-ready. If the governance model was the differentiator, what specifically delayed the highest-ROI use case this long? A board should require a post-mortem on that gap.
2. The deflection trend has no published explanation. Chatbot deflection fell from 75% at end-2023 to 54% to 46% at end-2025, and the inbound-contact reduction versus 2021 has narrowed from 43% to 18% over the same three years. The presentation offers no reason. Scope expansion to harder queries would be a reasonable explanation, but ING has not made it, and it publishes no CSAT or resolution-quality scores alongside the deflection figure. The same deck states that more than 80% of chats are resolved without human support in four countries — which suggests the 46% is a specific market, likely the Netherlands, without saying so. Until the two are reconciled, the narrative is not airtight for a board reading the trend.
3. No generative-AI-specific ROI is disclosed. ING's 2026 efficiency target — about €350 million, roughly 3% of costs, with some 1,250 fewer operations FTEs — is presented as operations-driven, not AI-attributed. The 25%-fewer-people figure is an expectation from the COO, not a measured result. For a bank with a €6.3 billion net result and more than 60,000 employees, that leaves the cost of transformation, and AI's share of the return, impossible to assess from outside. ING is better than the 70% of banks that disclose nothing (Evident Insights, August 2025). It is not yet auditable.
4. The denominator question on pilot-to-production. A 90%-plus conversion rate is compelling — but it measures pilots that were submitted. How many use cases were never submitted because the 140-risk gauntlet was judged too costly to run at all? That unseen numerator matters to any board deciding whether its own governance is appropriately calibrated or subtly over-engineered.
Operator takeaways: what to bring to your board
Build the platform before the use cases
ING's pilot-to-production rate is a platform outcome, not a talent outcome. If your AI team is rebuilding data infrastructure while deploying, you are already behind the compounding curve.
Treat governance as a deployment accelerator
A pre-production risk gate — even a rigorous one — costs less than post-launch remediation in a supervised institution. Design the gate to be thorough enough to be trusted and fast enough to be used.
Sequence by domain depth, not by vendor demo
ING's best outcomes are where its data is deepest: mortgages, KYC, the contact centre. Where is your institution's €370 billion equivalent? Start there.
Publish the series, not the snapshot
ING's 43% became 18% in two years. A single-year headline hides direction. If your board cannot read a sourced metric across time — and cannot separate AI's contribution from restructuring and cycle — it is funding a programme it cannot govern. Measurement discipline is a control function, not a communications choice.
The real ROI inflection at ING is still ahead — in the agentic workloads that entered production in 2026. The case is live.
Frequently asked
What percentage of ING's AI pilots reach production?
More than 90% of its generative-AI pilots, by ING's own account (investor presentation, 18 March 2026). Its head of advanced analytics strategy puts the industry average at about 30% (Computer Weekly, 13 January 2026) — his comparator, not an independent benchmark, and the 90% covers generative-AI pilots started from 2024.
How does ING govern AI risk before deployment?
A 20-step evaluation process checks every AI system against 140 risks before deployment (ING newsroom, November 2024), and an AI committee that includes the CEO reviews each use case (Reshaping Work, May 2025). ING says the EU AI Act required no structural changes because compliance was already embedded (Computer Weekly, January 2026).
What results has ING published from AI?
From its March 2026 investor presentation: inbound contacts needing human support down 18% versus 2021 (43% in 2023, 26% in 2024); chatbot deflection 46% at end-2025 (75% at end-2023); mortgage time-to-yes under one day for 80% of cases excluding customer time; 82% straight-through across 245 retail journeys; 50% fewer complaints. Its COO reports KYC due diligence in seconds for standard cases and up to 80% of data points answered without contacting the customer.
Why did ING build one AI platform before generative AI?
So models could be standardised centrally and deployed across markets without re-engineering — "from the beginning, we invested in having one platform," per Marco Li Mandri (Computer Weekly, January 2026). ING presents this as the structural reason for its pilot-to-production rate; the causal link is ING's claim and has not been independently tested.
What questions does ING's AI story leave unanswered?
Why the agentic mortgage assistant entered pilot only in March 2026; why chatbot deflection fell from 75% to 46% with no published explanation or quality scores; what AI specifically contributed, since no generative-AI-specific ROI is disclosed; and how many use cases never entered the 140-risk gate at all.
Get an independent read on your own pilot-to-production rate.
If your pilots are stalling where ING's ship, the useful first step is an honest diagnosis of which foundation — platform, governance gate or sequencing — is the real bottleneck.
Request a conversationSources
- ING, "Driving value through digitalisation, scalability and (Gen)AI," Morgan Stanley European Financials Conference, London, 18 March 2026 (PDF). Pilot-to-production >90%; inbound-contact, deflection and digi-index series; mortgage time-to-yes; complaints; 2026 efficiency target. ing.com (PDF)
- ING Group, FY2025 / 4Q2025 press release, February 2026. Customers, countries, employees, net result. ing.com
- ING newsroom, "ING's approach to AI explained in eight minutes!", 17 November 2024 (Bahadir Yilmaz: 20-step process, 140 risks). ing.com
- ING newsroom, "How ING uses agentic AI to speed up mortgage decisions," 8 June 2026. ing.com
- McKinsey & Company, "Banking on innovation: How ING uses generative AI to put people first," client case study (undated; project launched September 2023). mckinsey.com
- Cliff Saran, "Interview: How ING reaps benefits of centralising AI," Computer Weekly, 13 January 2026 (Marco Li Mandri). computerweekly.com
- Karl Flinders, "ING Bank transforming operations through agentic AI," Computer Weekly, 1 July 2025 (Marnix van Stiphout). computerweekly.com
- Computer Weekly, "ING increases use of AI in mortgage application process," 10 June 2026. computerweekly.com
- BFSI Insider, "ING Accelerates AI Rollout to Transform Operations," 1 July 2025. bfsiinsider.com
- Tech Monitor, "Training is everything in AI banking, says ING's Marnix van Stiphout," 12 August 2025. techmonitor.ai
- FStech, "All of ING's product fulfilment 'can and probably will' be impacted by agentic AI, says COO," 4 December 2025. fstech.co.uk
- Reshaping Work, "Navigating AI in banking: how ING balances innovation with data security," 20 May 2025 (Sheila Gemin; Reshaping Work Conference, October 2024). medium.com
- Evident Insights Banking Brief, "Exclusive: AI use cases surge," 7 August 2025 (117 use cases, 50 banks, H1 2025). evidentinsights.com
- Euromoney, "The world's best for mortgage/home loans 2026: ING Bank," July 2026 (€370 billion mortgages outstanding). euromoney.com
- MIT NANDA, State of AI in Business 2025 — The GenAI Divide (95% of organisations report zero return on enterprise generative AI).