What proprietary AI has PNC built in-house?
PNC is building its own AI stack down to the chips. CEO Bill Demchak told the Morgan Stanley conference in June 2026: 'We are building our own AI factory. We will have our own GPU compute', with PNC less reliant on 'burning external tokens' than on 'our own large language models'. American Banker reported that this means acquiring data centers, buying Nvidia chips and building large and small language models for banking problems such as blocking fraud and running call centers. In September 2026 president Mark Wiedman said PNC owns its data centers and GPUs and is 'bringing in our own proprietary LLMs. Actually SLMs, small language models', and CFO Rob Reilly said open-weight models run in PNC data centers while frontier models are reached through hyperscalers. On October 15, 2026 Demchak said GPU chips are already in its data centers. The software is built in-house too: about 115 agents built and tested the July 2026 PNC Mobile app, and agent-assisted development made the app and PNC Total Rewards about 40% more efficient to build, with 5 to 10 times the productivity expected from agent-centered development. A PNC digital assistant is planned for later in 2026. A spokesperson says the aim is to 'avoid vendor lock-in' and reduce reliance on usage-based pricing, while still mixing third-party and internal models by use case; head of data and automation Ned Carroll calls it 'us owning our intelligence'. PNC has not named its models, their sizes or base models, or the capital cost of the build.[1][2][3][4][5]
Data as of Oct 22, 2026 · PNC profile updated Oct 13, 2026 · 15 sources cited on this page · 140 banks profiled on this site · Markdown
Which AI models, platforms and agents has PNC built itself, and what do they run on?
From executives' statements and press reports. PNC has given no product names for its models; rows describe them as the bank does.
| Platform or model | What it does | Built on | Since | Status |
|---|---|---|---|---|
| AI factory (own compute) | Data centers and GPUs owned by PNC so that AI runs on 'our token inside of our own data centers' rather than external tokens | Nvidia chips in PNC data centers; a new Eastern facility under construction and a Midwestern hosting site secured | June 2026 (announced) | GPU chips installed by October 2026; cost shows up as depreciation[1][6][3] |
| Proprietary small language models | Task-specific models for banking problems; American Banker cites blocking fraud and running call centers; 'you don't need the whole kit and caboodle for most of the problems we have to solve' | PNC-built; base models and sizes not disclosed | 2026 | In use, per the president (September 2026)[1][2] |
| Own large language models | Demchak: PNC will rely less on external tokens than on 'what we will do internally for our own large language models' | PNC compute; not disclosed | June 2026 | Being built[1] |
| Open-weight models hosted in-house | Run in PNC data centers 'with no data sharing' | Open-weight models; families not named | 2026 | In use[2] |
| Model harness | Routes each task to the cheapest adequate model; Carroll will not put an API-accessed foundation model into a highly scaled operational process because it adds variable expense | Frontier models through hyperscalers plus PNC's own and open-weight models | 2025-2026 | In use[2][7][8] |
| Agentic development system | About 115 AI agents handled development and testing of the new app, with humans reviewing output; agent assistance made the app and PNC Total Rewards about 40% more efficient to build; 5 to 10 times expected from agent-centered development | Built in-house, 'not relying on vendors' | 2025 (agent assistance); July 2026 (app) | In production[4][2] |
| Teammate AI in branches and service | Helps staff answer customer questions, such as explaining an overdraft fee, quickly and accurately; learns from how staff use it | Not disclosed | 2026 | In use[7] |
| PNC digital assistant | Customer assistant to follow the new app, 'a slightly different flavor', letting clients choose how to interact | Not disclosed | Planned for later 2026 | In development[4] |
Who leads and builds PNC's in-house AI?
| Person or team | Role | Since |
|---|---|---|
| Bill Demchak, CEO | Set the AI factory course: 'We are building our own AI factory... That's a big deal'; wants the work done 'with your own personnel instead of vendors' | June 2026[1][3] |
| Mark Wiedman, president | Owns-the-compute message: own data centers, own GPUs, proprietary small language models; in-house software 'can actually alter the relationship with vendors' | September 2026[2] |
| Rob Reilly, CFO | Agentic development and harnessing done in-house; open-weight models in PNC data centers | February and September 2026[2][9] |
| Deborah Guild, head of technology | Advanced the national data-center strategy behind the compute build, per the 2026 proxy | March 2026[6] |
| Ned Carroll, head of data and automation | Model choice and data; separate risk policy for generative AI apart from model risk management; worked on the FSSCC data nutrition label | 2025-2026[1][5] |
| Alex Overstrom, head of retail banking | Led the agent-built PNC Mobile app and the coming digital assistant | July 2026[4] |
| Employees | A company-wide 'agent hack' asked staff to hack their processes with agents; two tellers reached the final with a working prototype | 2026[7] |
Why is PNC building AI itself instead of buying it?
Cost first. Demchak said in June 2026 that 'any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens', and that the most efficient tool 'oftentimes isn't the $35 token, it's the $1.50 token, or may well be our token inside of our own data centers.'[1]
A PNC spokesperson told American Banker the bank is positioning itself to 'avoid vendor lock-in': 'We are focused on reducing long-term reliance on external, usage-based pricing models. Over time, that includes building more internal capabilities and selectively developing models or model components to support specific business use cases.' The same statement said PNC continues to 'use a mix of third-party models and internally developed capabilities, depending on the use case'.[1]
Reilly gave two further reasons in September 2026: PNC's 'general high control nature of our own data', and efficiency from fitting each task to a cheaper approach, which he said vendors often do not do. Carroll put the data point more sharply: 'This is about us owning our intelligence... OpenAI and Anthropic would love for us to open up all of our data to them and let them train their models on it. But that's our proprietary knowledge.'[2][5]
On October 15, 2026 Demchak said long-term success needs core infrastructure 'without vendor support, without compute costs coming from third parties and with your own personnel instead of vendors.' Four days later at Money20/20 he drew the limit on agents: no bank will let agents transact inside accounts 'without a human in the loop'.[3][10][11]
What has PNC not disclosed about its proprietary AI?
PNC has not named its small or large language models, their sizes, their base models or the open-weight families it hosts, nor the hyperscalers it uses for frontier models. Demchak said on the Q3 2026 call that the bank has 'been purposefully quiet' about the work.[2][3]
No capital figure for the data centers or GPUs is published; the cost appears as rising depreciation. Technology spend is about $3.5 billion a year, expected to rise about 10% in 2026 with AI about 20% of the increase, and PNC declined to comment on what it invested in the agent-built app.[3][4]
What does the PNC profile record on this topic?
The rows below are the entries of the PNC AI strategy profile that bear on this page, as of Oct 13, 2026; the profile holds the full record of 26 dated moves, 14 use cases and 37 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Agentic software development | The system that built the new PNC Mobile app. | In production | [12] |
| In-house fraud-blocking models | One of the named targets for the bank's own language models. | Announced | [1] |
| Call-centre models | Small and large models for call-centre tasks, built in-house. | Announced | [1] |
| Open-weight models in PNC data centres | A harness picks the cheapest adequate model; frontier models are reached through hyperscalers, open-weight models run in-house. | In production | [2] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Sep 14, 2026 | Barclays conference: agents at the centre of development, 5–10x | Wiedman: using agents as the centre of software development, rather than as assistants, shows 5 to 10x productivity, work of ten weeks done in a week or less; Reilly: own data centres, own GPUs, proprietary small language models, open-weight models in-house, frontier models through hyperscalers. | [2] |
| Jul 8, 2026 | About 115 AI agents built the new app; a PNC digital assistant is planned | Banking Dive: the agents handled development and testing with human review of their output; about 8 million customers use the app, mobile active users up 8% year over year; the release says the platform is designed for embedded generative-AI assistance and that frontier models were used to harden the development process. | [4][13] |
| Jul 7, 2026 | New mobile app built with agentic development | PNC's agentic development system and patented data-streaming microservices architecture. | [12][14] |
| Jun 9, 2026 | 'We are building our own AI factory' | Demchak at the Morgan Stanley US Financials Conference: own GPUs, own models, less reliance on external tokens. | [1][15] |
| Feb 10, 2026 | BofA conference: the 'Big Five', $1.5 billion of AI-addressable spend | CFO Rob Reilly names software, retail operations, AML compliance, the client care centre and commercial loan processing (Midland); the new mobile app was '100%' agentic coding; on vendors, 'everybody, and a lot in-house'. | [9] |
Does PNC build its own AI models?
Yes, by its own account. President Mark Wiedman said in September 2026 that PNC is 'bringing in our own proprietary LLMs. Actually SLMs, small language models', running on GPUs in data centers it owns, and CEO Bill Demchak said in June 2026 the bank is building its own large language models. PNC also hosts open-weight models in-house and uses frontier models through hyperscalers. It has not named any of its models.
What is PNC's AI factory?
Demchak's name, from June 2026, for PNC's plan to own its AI production: acquiring data centers, buying Nvidia chips for its own GPU compute and building large and small language models for specific banking problems. By October 15, 2026 GPU chips were installed in PNC data centers, with the cost showing up as rising depreciation.
Does PNC still use OpenAI or other AI vendors?
Yes. A PNC spokesperson said the bank uses a mix of third-party models and internally developed capabilities depending on the use case, and CFO Rob Reilly said frontier models are reached through hyperscalers. The stated direction is to reduce reliance on usage-based pricing and avoid vendor lock-in, not to stop using vendors.
How did PNC use AI agents to build its mobile app?
PNC built about 115 AI agents that handled software development and testing for the new PNC Mobile app, launched in July 2026, with humans reviewing their output. The bank says agent-assisted development made the app and PNC Total Rewards about 40% more efficient to build, and expects 5 to 10 times the productivity once agents sit at the center of development.
Why is PNC building AI in-house?
Cost and control. Demchak says token costs can erase AI's productivity gains; Reilly cites control of PNC's own data and the efficiency of matching each task to a cheaper model; and Ned Carroll says the bank's IP is embedded in its data, which it will not open to model providers for training.
- Why PNC is building its own 'AI factory'
- PNC at Barclays conference: growth, AI and branch push (transcript)
- PNC (PNC) Q3 2026 Earnings Call Transcript
- PNC's new mobile banking app built on personalization
- Do banks need a nutrition label for their AI model data?
- The PNC Financial Services Group, Inc. 2026 Proxy Statement (DEF 14A)
- The AI-Enabled Bank: Reimagining Retail Banking from Branch to Mobile (Leaders session with Ned Carroll, PNC)
- How banks are looking to tame their growing AI bills
- PNC Financial at Bank of America Conference: Strategic Growth in 2026 (transcript)
- PNC Quietly Building AI Agents That Don't Rely on Tech Vendors
- PNC CEO: AI agent-only transactions will never happen
- PNC's New Mobile App Delivers an Intuitive, Personalized Experience to Help Clients Seamlessly Manage Their Day-to-Day Financial Lives
- PNC's New Mobile App Delivers an Intuitive, Personalized Experience to Help Clients Seamlessly Manage Their Day-to-Day Financial Lives
- PNC Bank rolls out new mobile banking app
- PNC's solution to rising AI costs: build its own AI
When PNC moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning