VACB 49th Annual Convention · Charlottesville · October 2026
AI & Community Banks: Friends or Foes?
Practical AI, emerging risks, and the questions leaders should be asking
VACB 49th Annual Convention & Trade Show · Boar’s Head Inn
AI & Community Banks: Friends or Foes?
Chris Busse · SingleStone · 10:00 am
On screen while the room fills: the Boar’s Head on an October afternoon. Two minutes, looping: flags, chimney smoke, falling leaves, a gust, geese heading south, a doe and her fawn. Advance when you’re introduced; the title slide’s walk-in plays then, not on load.
Open on the question in the title, the one every bank leader has heard in a board meeting: is AI going to help us or hurt us?
Let it sit for a moment, then advance. The signs are about to come down.
Act I
Friend or foe?
It’s the question in the title of this session, and one bank boards are asking right now.
Three acts. The first takes the title question seriously for about five minutes, then answers it.
A closer look
Neither. It isn’t a person.
Friend and foe describe relationships, and AI can’t have one with your bank. It feels no loyalty and holds no grudge.
- It is “an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions.”[1]
- A language model stitches together language it has seen “without any reference to meaning.”[2]
- So it can be fluent and wrong at once. NIST’s word for it is confabulation.[3]
The humanoid we picture when we say “AI” dissolves into what’s actually there: weighted connections, trained on data, producing likely outputs.
Treating it like a person feeds both fears, that it’s on our side and that it’s out to get us. Neither describes a model. Whether it helps or hurts your bank depends on what you use it for, and how.
Not a foe
Most AI pilots don’t crash. They never leave the ground.
You’ve seen the headline: “95% of AI pilots fail.” Look under the cowling.
- MIT’s own report puts the gap down to approach rather than model quality or regulation.[1]
- What breaks is mechanical: brittle workflows, and nobody owning the result.[1]
- RAND’s most common root cause: leadership “misunderstanding how to set the project on a pathway to success.”[2]
Treat the 95% with care. It comes from a small, preliminary, self-reported study.[1]
The “95% of AI pilots fail” line went everywhere in August 2025. The report itself says “95% of organizations are getting zero return,” calls itself preliminary, and notes its figures are “directionally accurate based on individual interviews rather than official company reporting.” Critics, including Wharton’s Kevin Werbach, questioned the number.
What survives the scrutiny is the diagnosis, and it’s mechanical. MIT’s words: the gap “does not seem to be driven by model quality or regulation, but seems to be determined by approach,” and what breaks is “brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.” In practice that means tools bolted onto processes nobody changed, with nobody owning the result. (Part of MIT’s “learning gap” is that the tools don’t retain feedback. That’s a fair criticism, and one more reason to choose carefully.) RAND found leadership misunderstanding the most common root cause, from 65 interviews a year earlier.
The aircraft is fine. The engine was never properly fitted. Nobody would call that sabotage.
But people can misuse it
The threat is whoever’s holding the tool.
- FinCEN has seen “an increase in suspicious activity reporting … describing the suspected use of deepfake media in fraud schemes targeting their institutions and customers.”[1]
- An Arup employee sent about US$25 million after a video call with deepfaked colleagues. Arup: “We can confirm that fake voices and images were used.”[2]
- Deloitte projects generative AI could enable US fraud losses of $40 billion by 2027, up from $12.3 billion in 2023.[3]
Even here, the danger “comes not from the technology used to create it, but from people’s natural inclination to believe what they see.”[1]
Behind every deepfake is a person pulling the strings. The AI didn’t decide to rob Arup; someone did.
FinCEN’s alert (FIN-2024-Alert004) gives red flags and asks institutions to use the key term “FIN-2024-DEEPFAKEFRAUD” in SARs. Make sure your BSA team has it. The FBI’s December 2024 PSA adds: “Criminals obtain access to bank accounts using AI-generated audio clips of individuals and impersonating them.” The last line on the slide is DHS, quoted in FinCEN’s alert. Deloitte’s $40B is a projection, not a count.
A tool can’t choose
A person decides, and a person answers for it.
“A computer can never be held accountable. Therefore a computer must never make a management decision.”[1]
- Lawyers who filed ChatGPT’s invented cases were sanctioned: “existing rules impose a gatekeeping role on attorneys.”[2]
- For banks, the agencies are blunt: “A banking organization has ultimate responsibility…”, whatever a vendor or an AI did.[3]
As then-FTC Chair Lina Khan put it in 2023, “There is no AI exemption to the laws on the books.”[4]
Everything after this rests on one idea. AI can draft, sort, flag and recommend, but it can’t be held responsible. A named person signs for every decision it touches, and the law already assumes as much.
The IBM line is attributed to a 1979 training page; the original was reportedly lost, and IBM’s archive couldn’t locate it, so say “attributed.” Mata v. Avianca (S.D.N.Y. 2023): $5,000 in sanctions; the court added that “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance.” The banking quote is from the interagency third-party proposal published Sept. 15, 2026 (comments due Nov. 16); it restates the 2023 guidance’s principle. Khan’s statement is from 2023; the laws it refers to still apply.
Act II
Choosing the tool
Pick the job first, then the tool, then the controls.
Act two is the practical part: the toolchest, finding the job, build or buy, and the controls that go with each.
So if AI is a friend…
AI is the friend who has every tool.
It will sell your bank one off the shelf, or help you build your own.
If we’re going to talk about AI as a friend at all, this is the only version that holds up: the neighbor with the fully stocked garage, and a workshop out back for anything that isn’t in it.
Pause here. Next slide is the reveal.
Every tool.
Build some. Buy some. Match each one to a job.
Let the tools land. Say nothing for a second.
This is the market a community bank faces right now: chat assistants, copilots, document extraction, fraud models, underwriting models, agents. Most are a contract away, and many are arriving inside products you already own.
The real risk
AI is easy to come by. Choosing it well is the hard part.
“I suppose it is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail.”[1]
“Can AI do this?” will nearly always get a yes. The better question is which tool suits which job.
The tools are already in the building: 49% of banks say they have deployed generative AI.[2]
Generative AI is the shiniest hammer anyone has handed a bank in a decade. It will happily drive a screw, badly.
The Cornerstone figure is self-reported deployment, from 416 bank and credit union executives at mostly $250M–$50B institutions; 49% is the bank figure.
Practical applications
Start with the job, not the tool.
Three jobs AI does well at a community bank:
- Efficiency. Loan-file intake, call and meeting summaries, first drafts of policies and procedures.
- Customer service. After-hours answers, call routing, and an assistant at your front-line banker’s elbow.
- Better decisions. Triage for fraud and BSA alerts, first drafts of credit memos, early warning across the portfolio.
AI for customer support is now a top opportunity for 47% of community bankers, up 16 points in a year.[1]
Put the hammer down. Nobody builds a branch by buying tools first. You draw the blueprint, decide what each room is for, and then pick the tools. Three rooms here: efficiency, customer service, better decisions.
Pick jobs with a person already in the loop and a clear measure of “better”: hours saved, calls answered, alerts cleared. The CSBS figure is the share of community bankers who see AI for customer support as a promising opportunity over the next five years, up from its first appearance in the 2024 survey.
Finding the job
Walk the process before you pick the job.
Follow one loan file from application to close. Mark:
- Waits. Where the file sits in a queue. The waiting usually dwarfs the work.[1]
- Re-keying. The same data typed into a second system.
- Rework. The file goes back for something missing.
- Judgment. Where a person makes the call. AI can help, but the call stays with the person.
The job for AI is usually a handoff or a re-key, not the whole process.[2]
RAND’s first root cause of failed AI projects: stakeholders “misunderstand — or miscommunicate — what problem needs to be solved.” The remedy is decades old: lean’s value-stream map. Toyota calls it genchi genbutsu, go and see. Sit with a loan processor for two hours and draw what actually happens, which is rarely what the procedure manual says. On a value-stream map the waiting usually dwarfs the work.
Mark four things: where the file waits, where data gets typed twice, where it goes back for something missing, and where a person makes a call. The first three are candidate jobs for AI. The fourth is where AI assists and a person stays accountable.
NANDA (MIT, 2025, preliminary): the organizations getting returns started from “narrow but high-value footholds”; budgets tilt to sales and marketing, but “back-office automation often yields better ROI.” That is the small line on the slide, in full.
Finding the job
Draw the walls: one job, one room.
Your bank already has walls: to deposits a customer is an account holder; to lending, a borrower; to BSA, a subject.[1] Give AI a job that fits in one:
- One room. One owner, one vocabulary, and only that room’s data.
- Through a door. At a handoff you’ve mapped, rather than reaching into every room.
- Worth doing. High volume, a person already checking the work, and a clear measure of better.
Software designers call these bounded contexts.[2] Pilots that sprawl across rooms stall on tools that “integrate poorly.”[3]
Eric Evans’ Domain-Driven Design gave software teams a way to carve up a big, messy business: bounded contexts. Inside each, words mean one thing, one team owns the data, and the rules are clear. Between them are explicit handoffs.
Your bank already works this way. To deposits a customer is an account holder; to lending, a borrower with a credit file; to BSA, a subject with a CIP record and a risk rating. Martin Fowler notes “Customer” is the classic word that means different things to different parts of a business.
A first AI job should live in one room, with one owner and one set of data it may see, or sit at a door you’ve already mapped. Tools that try to span the whole bank are the ones NANDA found “integrate poorly” and stall. Next: once you know the job and the room, build or buy.
Build or buy
Your friend can sell you a tool, or help you make one.
AI made small tools cheap to build. It did not make them cheap to own.
- Buy what every bank needs: core, fraud, document intake. Your diligence is the control.
- Build what’s yours alone, and keep it small: an internal assistant, a workflow, a report. Then you’re the vendor, with the testing, change control and support that comes with it.
Lean toward buying: MIT found internal builds “fail twice as often.”[1] Most community banks already do.[2]
This is how your bank already thinks about software. AI changes the cost of the “build” column (a capable person can now make a small internal tool in an afternoon), but not what building obliges you to do. The day you build it, you’re the vendor: you test it, you control its changes, you support it, and an examiner will ask about it like any other system.
MIT’s figure: externally sourced tools reached deployment about 67% of the time, internal builds about 33%. It’s the same small, preliminary study as before, weighted toward larger firms, so read it as a direction rather than a measurement. Treasury (2024) found smaller institutions “solely rely on third-party or core service providers” for AI capability. Next: what buying obliges you to do, and then, for the banks already down this road, where to start.
If you buy
Buying the tool doesn’t sell off the risk.
“Use of third parties does not diminish banks’ responsibility to manage risks.”[1] Ask the vendor:
- Is our customers’ data used to train your models?
- How will you tell us when the model changes?
- Can we explain its output to an examiner?
- How do we leave?
Your core’s AI roadmap is your AI roadmap. The September 2026 proposal comes with a joint statement on community banks and their core providers.[1]
Treasury (2024) describes a “widening capability gap” between the largest and smallest institutions in building AI in-house. For most banks in this room, AI arrives through the core and a handful of fintech partners.
The 2023 interagency third-party guidance is still in effect. On Sept. 11, 2026 the agencies proposed a revision and, alongside it, the FDIC, OCC and Fed issued a joint statement on community banks’ engagement with core service providers. Comments are open, and both are worth reading before your next core renewal.
Show of hands
Raise your hand if your bank…
- has people using a chatbot at work, approved or not.
- has switched on an AI feature your core or a vendor shipped.
- is running an AI pilot right now.
- has built an AI tool of its own.
Keep it up if you could name the person who owns it.
This one is for the room, so take your time. Ask each one, let hands go up, and say what you see (“about half the room…”). Expect nearly every hand on the first: most banks have someone using a chatbot whether or not there’s a policy. The second catches banks that think they haven’t started: AI that arrived inside a product they already own.
The last line is where it turns: ask people to keep their hands up only if they could name the owner. Hands come down. That gap is what the next slide is about.
If you’ve already started
Already started? Run it back through the plan.
Nobody needs to start over. Take stock of what’s already in the building.
- Inventory. Every AI tool in use, including pilots, vendor features and personal accounts.
- Map. Tie each one to a job and a room. If no one can say which, you’ve learned something.
- Tier. Check that the controls match the risk.
- Decide. Keep, fix or retire each one, and name an owner for every keeper.
The tools came in ahead of the plan: 78% of AI users bring their own to work, and 80% at small and mid-sized companies.[1] Retiring a pilot is a perfectly good outcome.
For the hands that went up: none of this means stop. It means running what you have back through the same steps: walk the process it sits in, draw its room, tier it, and read the contract or the code. The ones that pass get an owner. The ones that don’t get fixed or switched off, and switching one off is a decision you can show an examiner. There’s no shame in it.
Start with the personal accounts: they’re the tools with no contract and no policy. The Microsoft and LinkedIn figure is from a 2024 survey of 31,000 knowledge workers in 31 countries. It isn’t bank-specific, but I’d be surprised if bankers were the exception.
Emerging risks
Match the tool to the risk.
- Pencil. Drafting and summarizing, with a person editing everything.
- Drill. Customer-facing chat and copilots on internal data. You own what it says.[1]
- Crane. Credit, fraud, BSA, pricing. These need validation, fair-lending testing, and adverse-action reasons that are “specific.”[2]
Built or bought, the tier follows the job. And the 2026 model-risk guidance puts generative and agentic AI out of scope,[3] so if you build one, designing those controls falls to you.
Air Canada argued its chatbot was responsible for its own words. The tribunal called that “a remarkable submission” and held the airline to what the bot said. A customer-facing assistant is the bank talking.
SR 11-7 was superseded on April 17, 2026 by SR 26-2 / OCC 2026-13 (the FDIC joined). The new guidance says generative and agentic AI “are not within the scope,” with an RFI to come. The CFPB’s adverse-action circulars were withdrawn in 2025, but Reg B’s specific-reasons requirement is unchanged.
Emerging risks
Every tool gets only the data its job needs.
The safeguards didn’t change when the tools did.[1]
- Any tool. Keep an approved list, and keep customer information out of consumer AI and personal accounts.
- Bought. Put it in the contract: no training on your data, retention you control, notice of a breach.
- Built. You’re the custodian: the least data the job needs, with access controls and logs.
- Every tool. Treat prompts and outputs as records.
And expect the fraud to come down-market, to “smaller institutions.”[2]
The Interagency Guidelines require every bank to “ensure the security and confidentiality of customer information,” and to require its service providers, by contract, to do the same. Built, bought, or a free chatbot on someone’s phone, the Safeguards guidelines don’t care how the tool arrived. The room you drew on the bounded-contexts slide sets the limit: a lending assistant sees lending data, not the whole core.
The quick win at most banks is a policy line and an approved-tools list: which AI tools staff may use, and what may never be pasted into them. For vendors, it’s contract terms. For what you build, you hold the keys yourself: least privilege, access logs, retention.
On fraud: Treasury’s words are that it “would likely shift to more vulnerable corners of the sector like smaller institutions.” As the largest banks use AI to block more of it, attackers move to institutions with fewer defenses: deepfaked voices on the wire-transfer line, synthetic identities at account opening.
Governance
Keep a human in the cockpit.
“Loop thinking creates blame cultures. Cockpit thinking creates ownership cultures.”[1]
- Inventory every AI tool, including the ones inside vendor products.
- Give each one a named pilot, and report to the board.
- A person decides anything that affects a customer.
- Watch the gauges: re-test when the vendor changes the model.
Keep it proportionate. The OCC says its guidance “does not … require community banks to perform annual model validation.”[2] Examiners oversee AI with the rules they already have.[3]
“Human in the loop” makes a person a checkpoint, someone who watches for mistakes. “Human in the cockpit,” my SingleStone colleague Tripti Mishra’s phrase, makes them the pilot: they fly the aircraft, and they own where it lands. That’s what examiners and boards will expect to see.
And it can be proportionate. A community bank doesn’t need a model-risk department; it needs an inventory, an owner per tool, a person in the loop where customers are affected, and a board that hears about it.
GAO (2025): regulators “primarily oversee AI using existing laws, regulations, guidance, and risk-based examinations.” The information-security guidelines already require the board to approve and oversee the program, and AI tools that touch customer data fall inside it.
Act III
What comes next
Lessons from working with dozens of executives and their companies. They all lead to the same place.
Act three is less tidy on purpose: lessons from working with dozens of executives at banks and other regulated companies. Some are small. Watch where they end up.
What comes next
Everything in the AI industry has a six-month shelf life.
- Revisit. Decisions made six months ago were made on old facts.
- Re-test. Something that didn’t work last spring may work now.
- Date it. Put a review date on every AI decision.
The length of task AI can finish on its own has doubled about every seven months since 2019.[1]
The models, the vendors and the prices all turn over that fast. Your data governance doesn’t, and we’ll come back to that. The tools a bank looked at last spring are not the tools on the market now, and the ones that disappointed you then deserve another look, because the capability underneath them keeps improving.
METR measured the length of task a frontier model can complete with 50% reliability: it has doubled roughly every seven months since 2019. The paper has since been revised and retitled (“…Long Software Tasks”); the trend holds.
What comes next
There’s no magic wand in the chest, and you’ll likely buy AI more than once.
- The market is messy. Gartner counts about 130 real agentic-AI vendors among the thousands claiming the label.[1]
- It keeps moving. Expect to re-buy, re-contract and re-test as models and vendors change.
- Buy the exit. Make sure you can take your data with you and leave on reasonable terms.
Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027.[1]
Everyone would like to buy AI once: one platform, one contract, done. Nobody is selling that. The market is crowded with relabeled products (Gartner calls it “agent washing”), it’s consolidating, and it changes every quarter.
So plan to switch. That’s the “How do we leave?” question from the vendor slide, asked again at every renewal. Gartner’s reasons for the cancellations: escalating costs, unclear business value, inadequate risk controls.
What comes next
You won’t know your new normal until you’re living in it.
No one planning a transformation really knows what the new operating reality will look like.
- Small. Take steps you can finish in a quarter.
- Reversible. Favor pilots, short terms and data you can take back.
- Measured. Agree up front what “better” means, so you’ll recognize it when it arrives.
This one comes from experience rather than a study: an enterprise never really knows what its new operational reality looks like while it’s planning any kind of big transformation. AI only makes that more true.
So don’t plan the destination in detail. Plan to learn: small steps, each one reversible, each one with a measure agreed before you start.
What comes next
The slop is rising inside your own walls.
AI-written memos, decks and replies that look finished and say nothing.
- It moves the work. The sender saves time, and the reader spends it: nearly two hours apiece.[1]
- It costs trust. Every obviously machine-written memo makes the next one harder to take seriously.
- The fix is a norm. Whoever sends it stands behind it, and says when AI wrote the draft.
40% of U.S. employees received “workslop” in the past month.[1] Merriam-Webster’s 2025 word of the year: slop.[2]
I don’t mean the internet. I mean your own inbox: the colleague’s memo that is plainly a first draft from a chatbot, sent on without a read. The researchers define workslop as AI-generated content “that masquerades as good work, but lacks the substance to meaningfully advance a given task.”
Their survey of 1,150 U.S. full-time employees: 40% received some in the last month; about 15% of what they receive qualifies; each instance cost nearly two hours to deal with. You don’t need a ban, just a norm: if you send it, you own it, and you say when AI wrote the first draft.
What comes next
The bad guys are ahead of us.
Fraud rings, ransomware crews, scammers, insiders, hostile states. None of them waits for a committee, a contract or an examiner.
- They move first. A new capability reaches them long before it clears your procurement.
- They come down-market. Treasury expects them to move toward “more vulnerable corners of the sector like smaller institutions.”[1]
- It’s working. Reported internet-crime losses reached $16.6 billion in 2024, up 33%.[2]
That’s why the controls in Act II matter, and why your defenses will need AI as well.
In Act I we met one person misusing the tool. This is the whole field, and the point is the asymmetry: an attacker adopts a new model the week it ships. A bank adopts it after a risk assessment, a contract, a policy and a board report, as it should.
The FBI’s IC3 logged 859,532 complaints and $16.6 billion in losses in 2024, a record. Treasury expects AI-enabled fraud to drift toward smaller institutions as the large banks harden. The deepfake and synthetic-identity examples from Act I belong here too.
What comes next
People are saying the quiet part out loud. So who trains the juniors?
“In the next few years, we expect that this will reduce our total corporate workforce.”[1]
- The bottom rungs. Employment of 22–25-year-olds in AI-exposed jobs is 19% below trend, mostly because fewer are being hired.[2]
- Your bench. Today’s analysts, processors and tellers are tomorrow’s chief lender, BSA officer and COO.
- Design the apprenticeship. If AI writes the first draft, have juniors check it, and teach them what they’re checking for.
Executives are now saying in public what they used to say in private. Andy Jassy’s memo to Amazon employees, June 2025, is the plainest example.
The Stanford Digital Economy Lab’s payroll study is the one to watch: in its August 2026 revision, employment of 22–25-year-olds in AI-exposed occupations stands 19% below where it would be had it kept pace with less-exposed peers (up from 13% when they first measured it), mostly through reduced hiring. Experienced workers show no comparable gap.
For a community bank this is a succession question. If the entry-level work goes to AI, you have to design how people still learn the craft.
It’s the data
The first hard question comes on day one: why did it do that?
Someone will ask it early: an examiner, a customer, or a member of your board.
- What it saw. The inputs, and the data behind them.
- What it said. The output, and what a person did with it.
- Which version. The model, the prompt and the settings on that date.
- Who approved it. The owner, and the decision on record.
Reg B still wants “specific” reasons for a denial.[1] Whatever you don’t record at launch, you can’t explain later.
From here on, every question in this act is a question about data.
Banks expect the observability questions to come at the next exam. They come the first week: a customer complaint, a denied application, a board member who saw a strange answer. Answering them takes a flight recorder (inputs, outputs, model and prompt versions, approvals) switched on at launch rather than bolted on afterward.
It’s the data
Most of your AI governance answers are already in your data governance.
- What does it see? Classification and access.
- Where did that come from? Lineage.
- Is it right? Data quality.
- How long do we keep it? Retention.
- Who answers for it? The data owner.
RAND’s second root cause of failed AI projects: the data isn’t there.[1] Gartner expects 60% of AI projects without AI-ready data to be abandoned.[2]
When a bank sits down to write an AI governance policy, most of the questions it ends up asking are questions its data governance already answers, or should. What may this tool see? Classification and access. Where did this answer come from? Lineage. Is it right? Quality. How long do we keep the prompts and outputs? Retention. Who answers for it? The data owner.
In RAND’s interviews with data scientists and engineers, the second root cause of failure was that the organization lacked the data to do the job. Gartner (Feb. 2025): through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data.
It’s the data
And doing AI governance will improve your data governance.
Every AI tool you roll out puts your data to the test.
- It finds the gaps. The stale field, the missing owner, the spreadsheet nobody documented.
- It forces decisions. Who may see what, for how long, and why.
- It pays twice. Fix the data for the AI, and every report, model and exam that relies on it improves too.
63% of organizations don’t have, or aren’t sure they have, the data management practices AI needs.[1]
It runs the other way too. Point an assistant at your loan files and it will find, in an afternoon, every field your data dictionary says is required and nobody fills in. The AI didn’t create that problem; it just made it visible.
So treat AI governance and data governance as one loop: each AI rollout should leave your data governance better than it found it. The 63% is from Gartner’s Q3 2024 survey of 248 data management leaders.
Act III
It’s the data.
Working with AI is mostly about working with your data, and doing it right.
Let the dots pour out. Then say it simply: the tools will change every six months. The vendors will change. The bad guys will keep moving. What lasts is your data: what you have, how good it is, who owns it, who can see it, and what you can prove about it.
Get that right and every tool in the chest becomes safer and more useful. Neglect it, and no tool will make up the difference.
For the board table
Five questions leaders should be asking.
- Which jobs are we giving AI? Have we walked the process, and who gets hurt if it goes wrong?
- Is this the lightest tool that does the job?
- Are we building or buying, and what do our vendors and our core do with our customers’ data?
- Could we explain the result to a customer and to an examiner?
- Who owns each AI tool, and do we know what data it sees and keeps?
Leave this one up during Q&A. If a bank takes one thing home, it’s these five questions on the next board agenda.
AI & Community Banks: Friends or Foes?
Thank you.
Chris Busse
Chief Client Officer & Head of AI, SingleStone
VACB 49th Annual Convention & Trade Show · Illustrated with dotscene · presented with dotpreso
Thank VACB. Offer the five questions as the takeaway.