A POV on how retail and small business customers will manage money in North America over the next three to five years, and what it means for banks, challengers, and fintechs.

The airline analogy that says everything

Air Canada owns Aeroplan. It is Air Canada’s single most defensible customer relationship. On Point Loyalty’s 2026 ranking values the program at US$7.4 billion, the eighth most valuable airline loyalty program in the world. Air Canada’s entire market capitalization sat near C$5.7 billion this spring. The loyalty program is worth comfortably more than the airline that owns it.

And yet.

If you have Aeroplan points, you can fly Lufthansa, United, Singapore Airlines, Turkish, Air New Zealand, and the rest of a 45-airline partner network. You never leave Aeroplan. Aeroplan does the coordination for you. The alliance handles interlining, settlement, and currency conversion. You do not choose an airline in the old sense. You have a primary relationship with a loyalty currency that shops for you across airlines.

Star Alliance was formed in 1997. Oneworld and SkyTeam followed within three years. Within a decade, the frequent traveller had learned to think in alliance networks rather than individual carriers. Some airlines won by joining an alliance and pooling supply. Some won by refusing to align and being uncompromising in one segment. What died was the middle: undifferentiated full-service carriers with weak alliance positions and nothing distinctive of their own. When aggregation arrives, it sorts on differentiation.

That is the model consumer finance is about to run.

The alliance layer for money will not be built by banks. It is being built by ChatGPT, Claude, Perplexity, and the next generation of AI orchestrators that follow them. The customer’s primary relationship will be with the orchestrator. The banks, insurers, lenders, and brokerages become the fleet the orchestrator books against.

This POV is about how that transition plays out over the next three to five years in North America, what customer behaviour has been telling us all along, and what each type of player (big banks, challengers, fintechs, orchestrators) can defensibly own on the other side. It covers retail and small-business finance. It does not cover institutional money. That is a different essay.

A note on numbers. Sourced figures are linked. Where a number is my estimate, it says so.

---

Listen on Notebook LLM or the player below.

---

What customer behaviour has always been

The industry has been building against a fantasy of the customer for thirty years. The fantasy is that people want to see their money.

They do not.

They want to know they are okay this week. They want to hit the goal they set six months ago. They want to be treated well when they make a mistake. That is the entire behavioural repertoire of retail personal finance. Ask people what they want from their bank, and some version of those three answers has come back for decades. And every product built on the assumption that customers wanted a beautiful dashboard has died the same way.

Mint signed up more than twenty million registered users over seventeen years. By 2021, its monthly actives had fallen to roughly 3.6 million. Intuit shut it down in March 2024. The interface was fine. Looking at your money does not change what you do with your money.

There is a name for the mechanism. Karlsson, Loewenstein, and Seppi called it the ostrich effect, and Sicherman and colleagues later confirmed it in account login data: when markets fall, investors check their portfolios less often. The dashboard’s core assumption is that people will look. People look when looking feels good and avoid it when it matters most.

The behavioural finance literature has been making the larger point for two decades, in two landmark studies that are usually blurred into one. Thaler and Benartzi published Save More Tomorrow in 2004: employees precommitted today to saving more from tomorrow’s raises, and inertia worked for them rather than against them. Of those offered the plan, 78% joined, and their savings rates rose from 3.5% to 13.6% over 40 months. Madrian and Shea ran the companion experiment with pure defaults: automatic enrollment took 401(k) participation from roughly 37% to 86%. Two mechanisms, one lesson. Interventions that move money act on the customer’s behalf.

Fitbit, Whoop, Oura, MyFitnessPal all built beautiful information layers for health. Endeavour Partners found a third of American owners stop using their wearable within six months, and more than half of all-time owners no longer use one. Health tracking carries more emotional charge and more visible outcomes than personal finance, and even that is not enough to make information-display behaviour stick.

The survey data has consistently misled the industry because it measures the customer’s ideal self. A person surveyed about financial dashboards says they want visibility, control, and clarity. The same person, watched for 90 days, opens the app 4 times in the first week and never again. Sheeran’s meta-analytic work on the intention-behaviour gap puts a number on it: stated intentions explain about a quarter of the variance in people’s behaviour. The rest is context, defaults, and friction.

The customer’s ideal self reads dashboards.
The customer’s actual self wants someone else to handle it.

One honest amendment to the dashboard obituary. Dashboards did not die. They shrank to a hobby. YNAB charges about $100 a year and thrives on its devotees. Monarch raised $75 million in 2025 serving Mint refugees. A minority will always pay to look at their money, the way a minority bakes bread from scratch. The mass market wants the bread. This has been true forever. The industry ignored it because there was no interface capable of handling it. Now there is.

What the current LLM layer is catching

ChatGPT reached 900 million weekly active users in February 2026. On May 15, 2026, OpenAI launched personal finance inside ChatGPT, connected to bank, brokerage, and card accounts through Plaid across more than 12,000 institutions. It shipped as a read-only preview for Pro subscribers in the US and expanded to Plus users on June 25. OpenAI says more than 200 million people already come to ChatGPT every month with money questions. Perplexity launched Portfolio in March 2026 with read-only brokerage access, then extended the Plaid connection to bank accounts, credit cards, and loans in April, with its Computer agent building net-worth trackers and budgets on request. Perplexity says 75% of its users already ask finance-related questions each month. And Intuit partnered with OpenAI in November 2025 so ChatGPT users will be able to take actions through Intuit’s apps: check credit card approval odds, submit the application, get a tax estimate, book a live tax expert, all inside the chat.

(Anthropic shipped ten finance agent templates the same month OpenAI launched consumer finance. Those are institutional tools for pitchbooks and KYC files, outside this essay’s scope. Note the velocity anyway.)

TD’s own year-over-year comparison is the sharpest single data point: 10% of Americans asked AI a financial question a year ago; 55% do today.

How many American adults use AI for money depends entirely on how you ask the question. Wells Fargo, asking about financial advice, found 19% and 38% of Gen Z. FNBO, asking about help with personal finances, found 46%. TD, asking about help managing money, found 55%. TD’s own year-over-year comparison is the sharpest single data point: 10% of Americans asked AI a financial question a year ago; 55% do today. The spread across surveys comes from question wording. The direction is the same in every one of them: up, fast. And the usage is not idle. In the Wells Fargo data, two-thirds of people getting AI advice acted on it, and 90% of those called the result worthwhile.

The Canadian number is lower, mostly because write access to bank data is gated behind CDBA Phase 2, which is about a year out if the timeline holds. Adjusting the US figures for product availability, my estimate puts Canadian usage somewhere around 30% of adults. That is an inference, labelled as one. I expect it to roughly double over the next two years and roughly double again by 2030. Also mine.

The small-business picture is further along, with one qualifier: coverage keeps dropping. The Goldman Sachs 10,000 Small Businesses survey published in March 2026 found that 76% of respondents currently use AI, with 93% of users reporting a positive impact and only 14% saying AI is fully embedded in core operations. The qualifier: the sample is 1,256 participants in Goldman’s own program, a growth-selected cohort. NFIB’s broader Main Street sample found about a quarter of small businesses using AI at all. Both numbers are true, and the gap between digitally engaged firms and the median plumber is itself part of the story. Intuit’s May 2026 AI Impact Report, built on 34,000-plus survey responses and anonymized data from 5.3 million QuickBooks businesses, sits between them: 77% of US businesses now use AI regularly, up from 48% in July 2024. On time saved, Intuit’s disclosed figures: 45% of customers using the AI-powered bank feed report saving 12 hours a month on bookkeeping, and its December 2025 customer survey puts the average at 6 hours a month across the broader base. Blue J and CPA.com surveyed more than 1,000 US tax professionals in June 2026 and found 60 percent now use AI for tax research at least weekly, up from 33% a year earlier.

Small businesses adopt faster than retail consumers for a specific reason. They have a greater coordination burden and no accountant on the payroll. AI is the only affordable way to close the gap between what they need to do and what they know how to do.

But raw usage is not the interesting number. The interesting number is the funnel.

Here is my working model, labelled as such. At any point in time, there are three concentric circles of AI-in-finance users. The outer circle is anyone who has ever asked the AI a financial question. The middle circle is anyone who uses it monthly for financial tasks. The inner circle is anyone who lets it move money. Right now, roughly a third of the outer circle becomes the middle circle, and roughly a quarter of the middle circle becomes the inner circle. By 2031, I expect both conversion rates to roughly double, because trust builds. Those conversion rates are my estimates, not survey findings.

Trust is the whole game, and the real survey numbers draw the gradient better than any single pair of figures could. TD’s February 2026 survey: 62% of Americans trust AI to provide honest, reliable information, and 18% would trust it to make financial recommendations on its own. YouGov, December 2025: 56% would trust AI to flag unusual transactions, 16% would trust it to move money even with approval, and 10% would trust it to make financial decisions automatically. HSBC’s 2026 survey of affluent investors found the same pattern at the top of the wealth ladder: 73% use AI for finance and investing, 12% named it the most influential input in their most recent decision, and half want a hybrid of AI and a human adviser. Trust falls in a smooth line from watching to advising to acting. The distance between the top of that gradient and the bottom is the entire commercial opportunity in AI finance, and closing it is the work of the next five years.

Here is the important point about the current generation of LLMs. They are the intermediate step. They are still single-player. They still require the customer to prompt them. They still place the coordination burden on humans. They are, in behavioural terms, a very well-informed search bar with better memory. The customer still has to decide what to ask, which account to reference, and what to do with the answer.

Even at that intermediate level, somewhere between a fifth and half the country is already using them for money, depending on your definition. Three-quarters of the most growth-minded small business owners are.

That tells you the customer behaviour is not a demand-side problem. The customer has been waiting for this. What the customer has been waiting for is not what the industry has been building.

What the agentic layer will do next

The next generation of these products, arriving in the 2027 to 2030 window, is agentic in the full sense. It monitors your finances in real time. It surfaces the right decision at the right moment. It negotiates on your behalf. It executes with delegated authority. It handles the multi-institution coordination that the current generation still leaves to humans.

The specific promise, in customer terms: never open a dashboard, never call a bank, never fill out a mortgage application, never shop for insurance, never file a tax return. Speak to the layer in natural language when you want to. Otherwise, it runs your financial life the way a team of experts would if such a team were affordable.

For small businesses, the promise is even more concrete. The AI runs the ledger, expense management, payroll, tax filing, working capital decisions, and invoicing follow-up. Intuit’s agent suite already automates categorization, reconciliation, and invoice chasing, and its own numbers show that nearly half of AI bank-feed users get 12 hours a month back. The remaining human work is exception handling and judgment. That is what the founder does. Everything else the machine does.

The reason this is technically feasible now, but not in 2020, is that LLMs can maintain context across multiple institutions and over long time horizons. They can integrate structured data from brokerages, bank feeds, credit files, tax records, and insurance policies. They can reason about tradeoffs the way a financial advisor does, at a marginal cost approaching zero. And they can act, once the delegated-authority protocols and the write-access APIs are in place.

Pricing shifts from subscription toward outcomes, with one behavioural asterisk. Lambrecht and Skiera documented the flat-rate bias: people prefer a predictable flat fee even when paying per use would be less expensive. Expect hybrids. A flat fee for peace of mind, outcome-based pricing layered on top.

And one honest cost, because delegation cuts both ways. Prelec and Loewenstein showed that decoupling payment from consumption raises spending. The less it hurts to pay, the more we spend. Fully frictionless money is also how people drift. The same layer that catches your overdraft can anesthetize your spending, so good orchestrators will have to engineer some pain back in, on purpose.

The reasons it is not fully commercialized yet are trust, liability, and regulation. Not technology.

The inattention economy

Before the precedents, follow the money: the current system makes money from you not paying attention.

US banks collected about $5.8 billion in overdraft and NSF fees in 2023, per CFPB reporting, down from more than $12 billion a few years earlier only because the regulator started squeezing. Credit card late fees ran roughly $14 billion in 2022. Add the spread earned on deposits customers forget to move, the subscriptions they forget to cancel, and the out-of-pattern fees they never contest. Thaler has a word for friction engineered to produce this revenue: sludge. A meaningful slice of retail banking profit is a tax on inattention.

Thaler has a word for friction engineered to produce this revenue: sludge. A meaningful slice of retail banking profit is a tax on inattention.

Now the study that reframes the whole interface debate. Stango and Zinman found that people who merely answered survey questions about overdraft fees went on to overdraw less for as long as two years afterward. No dashboard. No alert system. Just a few minutes of forced attention. Attention alone changes financial behaviour, and a display cannot create attention. It can only demand it. That is why displays fail.

An AI orchestrator is outsourced attention. It watches so the customer does not have to. Its business model, whatever the pricing wrapper, is capturing a slice of recovered inattention rent: the overdraft that never happens, the idle cash that finally moves, the subscription that dies, the refinance that gets done the month it makes sense.

That is also the clearest explanation of why banks cannot build it. The orchestrator’s core function is deleting revenue the bank currently books. Deleting that revenue is the value proposition, and a bank cannot ship a product whose success shows up as a hole in its own fee line.

Where this has already happened

The airline alliance model is the closest analog in North America, but the more instructive precedents are elsewhere.

China. WeChat Pay and Alipay together handle roughly 90% of third-party mobile payments in a market whose annual volume is measured in the tens of trillions of dollars. For hundreds of millions of Chinese consumers, the primary financial relationship is a super-app, and the banks provide the plumbing. The relationship, the identity, the interface, and the daily habit sit inside a layer the banks do not own. This happened over roughly a decade, without an AI interface. Substitute AI for QR codes and the pattern is directionally the same and probably faster.

India. UPI plus Aadhaar plus a growing set of agent-driven financial products. Consumers still choose HDFC or ICICI or SBI at the account level. What has changed is the layer above the account. Groww, Paytm, Cred, and the RBI’s account aggregator framework have made this the default consumer experience for anyone under thirty-five. The bank is still there. The bank just does not own the relationship anymore.

Latin America. Nubank started as a credit card and led primary-institution incidence across Brazil by the fourth quarter of 2025, at roughly 30%, per NPS Prism by Bain. It crossed 112 million Brazilian customers in January 2026 and became Brazil’s largest private financial institution by the central bank’s own count. The transition from single product to primary relationship took about a decade. Mercado Pago is running the same play. Every regional bank in Latin America is now a supplier to a super-app.

Africa. M-Pesa did this in Kenya starting in 2007 on feature phones, without a smartphone and without a legislative framework built for it. As of the first quarter of 2026, Kenya’s mobile money subscriptions reached 53.4 million, penetration exceeded 100% of the population, and M-Pesa held an 89% share. If you are wondering whether consumers will trust a non-bank to hold and move their money, M-Pesa answered that question at national scale fifteen years before ChatGPT existed.

United Kingdom. The control group. The UK mandated open banking in 2018: standardized APIs, regulated access, the exact plumbing North American regulators are now building. Seven years on, 15.16 million users — nearly one in three UK adults — are connected, up 34% year over year, on Open Banking Limited’s own count (measured as user connections per brand, not deduplicated individuals). And the usage is real: HMRC tax payments, Just Eat and Tesco checkouts, sweeping between accounts. The rails changed how Britons pay. What they did not produce is a primary money-management relationship. The UK proved the bottleneck was never the pipes. It was the absence of an interface that could reason, and that piece arrived in 2023, not 2018.

Every market that has completed the full transition has ended up with a primary financial relationship that sits above the bank, and a bank that serves as a supplier of custody, settlement, and lending. The direction of travel is set. What varies is the mechanic and the timeline.

All four traversals happened without an AI interface. AI is not necessary for aggregation. AI is what makes aggregation compress in a mature-card-network market that would otherwise resist it. That is the North American case, specifically.

North America has been slower for three reasons. The card networks work well enough to have blunted the demand for a super-app. Regulation has been more protective of incumbents. And the population is more heterogeneous, so no single dominant super-app emerged the way WeChat did in China.

None of those three reasons hold in an AI orchestrator world. The AI does not need the card network to be broken. It does not need regulation to catch up, though regulation helps. And it does not need cultural homogeneity, because it adapts to the customer, not the market.

The airlines are not the only precedent worth naming

Travel. Kayak, Expedia, and Priceline turned airlines and hotels into suppliers. The customer relationship moved to the aggregator. Airlines and hotels still exist, still make money, and still compete on product. They just do not own the pricing conversation with the customer.

Media. Netflix, Spotify, and YouTube turned studios and labels into suppliers. The relationship moved to the interface. Studios still make content. Labels still sign artists. They just do not own the customer.

Retail. Amazon turned brands into suppliers. The customer relationship moved to the platform. Brands still make products. Amazon owns the browse-buy-review-return conversation with the customer.

Grocery. Instacart, DoorDash, and increasingly Amazon Fresh turned supermarkets into suppliers. Not all the way. Not everywhere. But the direction is set.

Ben Thompson named this pattern a decade ago: aggregation theory. Commoditize supply, own demand, and the customer relationship migrates to whoever controls discovery. Finance fits the pattern last because finance demands more from the interface than travel or media ever did, and only LLMs can supply it.

But the pattern has a second half that the aggregation-always-wins crowd skips. Some suppliers fought back and won ground. Airlines pulled bookings back to their own channels, and Southwest never listed on the OTAs at all. Hilton ran “Stop Clicking Around” in 2016 and wired direct booking into its loyalty economics. The split was not random. Differentiated suppliers with their own loyalty currency kept direct customer relationships. Commodity suppliers got aggregated. Same sort, one layer up. Every bank should be asking which side it lands on. Almost none are.

The counterexample worth taking seriously

Anyone arguing banks cannot build the new layer has to get past Zelle. The big US banks watched Venmo, formed a consortium, and shipped their own P2P network in 2017. Zelle now moves more than a trillion dollars a year, more than Venmo and Cash App combined. Banks built a customer-facing layer and won it. This essay’s claim has to survive that.

Here is the sharper version. Banks have repeatedly won pipes: Zelle, Interac, the card networks themselves. Pipes are neutral by construction. A transfer is a transfer, and there is nothing to recommend. What banks have never built is an intelligence layer: a product whose job is to form a view and, when the owner’s product loses on the merits, recommend against its owner. Zelle never has to tell a Chase customer to leave Chase. An orchestrator has to, or it is not an orchestrator.

So the live question is whether financial orchestration behaves like P2P rails, where the consortium won, or like search, where no publisher consortium ever stood a chance. I think it is search, for one reason. Rails are a coordination problem, and banks are good at coordination. Recommendation is a conflict-of-interest problem, and a consortium cannot solve a conflict it is made of. But this is the strongest objection to the whole thesis, and it belongs inside the essay rather than left for a banker to raise in the comments.

The three moats (plus one)

Here is where the industry conversation gets interesting. The airline analogy is not a warning that everyone becomes a supplier. Some airlines became suppliers. Others became the alliance leader. Others became the low-cost carrier. Others became the premium brand.

The same fanout is available in finance, and the moat you can defend depends on what you already are.

The big bank moat: custody, trust, and regulatory position.

Big banks are not going to become orchestrators. They cannot. Their business model requires them to sell their own products, and the inattention economics above make it worse: the orchestrator’s core function is to delete fee revenue that the bank currently books. Finn by Chase launched in October 2017 and shut down by mid-2019, less than two years in. The post-mortems were blunt: Finn was neither fully digital nor independent of Chase’s retention interest, and it could not credibly recommend against Chase. The reason was structural.

What big banks can defend is the position of the recommended custody and settlement partner. When the orchestrator says to move $10,000 to a high-yield savings account, some institution has to hold it. When it says, “Here is your mortgage,” someone has to underwrite it. When it says settle this payment, someone has to clear it. The big banks that survive as the recommended suppliers are the ones that build the deepest integrations with the orchestrators, offer the best terms at the point of decision, and hold a trusted brand that makes customers feel safe about letting an AI move their money.

Knowing the customer is no longer the moat. The orchestrator will know the customer better than the bank does within three years (my projection). The moat that remains is the AI, which keeps the bank on the shortlist because the bank is consistently the best product at the point of decision, and the customer trusts the bank’s name enough to accept the recommendation.

Analog: British Airways in the Oneworld alliance. The premium supplier the alliance keeps recommending because customers trust the brand and BA delivers the product.

The challenger bank moat: the vertical action layer.

Challenger banks (KOHO, Neo, Simplii, Wealthsimple’s cash product, and their US analogs Chime, Current, Varo) have a different structural position. They are already digital-first. They are already unbundled from the traditional bank stack. They can move faster and take more product risk than a big bank can.

What challengers can defend is the position of the best-in-class product for a specific job. The AI orchestrator does not care about your brand. It cares about which product delivers the best outcome for a given task. If Wealthsimple offers a higher-yield savings product with better fee structure and faster funding than any big bank, the orchestrator recommends Wealthsimple for savings. If KOHO offers a better everyday spending experience with tighter loyalty integration, the orchestrator recommends it for spending.

The challenger’s moat is being the product the orchestrator picks. That is a very different moat from being the customer’s primary bank in the old sense. It requires vertical excellence over breadth, operating margin over cross-sell, and the ability to win a product-level bake-off against every other challenger in the same category.

Analog: JetBlue or Southwest. The category winner in a specific segment.

The fintech moat: the workflow that owns the moment.

Fintechs (Plaid, Ramp, Brex, Melio, Wave, TurboTax, Credit Karma, and the emerging AI-native players) sit in a third position. They are workflow layers. They own a specific moment in the customer’s financial life (small-business bookkeeping, expense management, tax preparation, payment infrastructure), and they own the software the customer uses at that moment.

What fintechs can defend is the workflow the orchestrator has to route through. If QuickBooks owns the small business ledger, any AI orchestrator that helps a small business owner will route through QuickBooks. If Plaid owns the bank connection layer, any orchestrator that needs bank data will route through Plaid. The moat is being the mandatory infrastructure at a specific moment in the workflow.

The fintech moat is the most interesting because it can coexist with the others. A fintech can be the workflow layer the big bank uses to serve the customer, the challenger uses to compete, and the orchestrator uses to execute. It sits horizontally under all three.

Small business is where the fintech moat is deepest. Intuit’s position with QuickBooks, TurboTax, and Credit Karma is the most durable fintech position in North America because it owns three of the most-used workflows in a small business owner’s life, and every AI orchestrator that helps a small business owner has to route through at least one of them. Its OpenAI partnership makes it the first supplier wired for actions inside the biggest orchestrator. Wave in Canada, Xero globally, and FreshBooks each hold different segments of the same territory.

Analog: Sabre, Amadeus, and Travelport. The GDS layer that every airline and every travel site has to run through. Nobody sees them. Everyone pays them.

The orchestrator moat: the interface itself.

And then there are the orchestrators. Perplexity, Claude, ChatGPT, and the next generation of AI-native players that will emerge from them (like Fin’s Intercom play). The moat here is the interface layer itself. The relationship with the customer. The habitual point of entry into the customer’s financial life.

There will be fewer of these than any other category. Two to five in North America over the five-year horizon (my count, not a sourced forecast). Winner-take-most economics. This is the alliance leader position, and it is the position no bank or fintech can defend against a well-funded AI company with a good interface.

The neutrality problem

One assumption in the moats deserves its own stress test: that the orchestrator recommends the best product.

Every aggregator named above started neutral and ended up selling the slot. Amazon’s advertising business runs north of $50 billion a year. Booking sells rank. Google is Google. The early signals in AI are mixed: OpenAI announced in January 2026 that it will test ads inside ChatGPT’s free tier and the new $8 Go tier, with Plus, Pro, and business customers excluded. Perplexity has experimented with sponsored placements. Anthropic has committed to keeping ads out of Claude. Which model wins is an open question, and it matters more than most of the regulatory questions.

The disclosure fight is already in the launch language. OpenAI’s own framing promises that ads will not dictate ChatGPT’s answers and will not appear in regulated topics like health or politics. That commitment, day one, is the fight — because it will be tested, litigated, and priced against every quarter of ad revenue that follows.

If the recommendation slot goes to auction, “best product wins” degrades into “best bidder wins,” and a pay-to-rank orchestrator recreates the bank’s conflict of interest one layer up. Two consequences follow. For challengers: vertical excellence stays necessary and stops being sufficient. Treat orchestrator distribution the way smart airlines treated the OTAs. Take the volume, keep a direct door, and never let the aggregator own all of your discovery.

For customers and regulators: neutrality disclosure at the recommendation layer becomes the fight that ad disclosure was for search. The orchestrator that can prove its recommendation is unbought will own the trust position, and in a category where trust is the whole game, that may decide the winner.

The small business dimension deserves its own section

Everything above applies to retail personal finance. It also applies to small business finance, but faster and harder.

Small businesses have a greater coordination burden per customer than retail businesses. A retail customer might touch three financial institutions. A small business owner touches ten to fifteen (my count: bank, credit card, payroll, invoicing, expense management, tax software, insurance, retirement, business credit, benefits, and often two or three merchant processors). Coordination burden does more than annoy. It suppresses action. Huberman, Iyengar, and Jiang found that every 10 additional fund options in a 401(k) plan reduced participation by roughly 2 percentage points. Choice load is a behavioural tax, and a founder holding fifteen financial relationships pays it daily. The orchestrator’s value is proportional to the burden it absorbs. Call it three to five times the retail case. That multiple is my estimate.

Small businesses also have thinner margins on advisory. They cannot afford a bookkeeper, a CFO, and a lawyer. The AI’s competition here is a spreadsheet, an unpaid spouse, or nothing at all. No bookkeeper is losing this account, because no bookkeeper had it. That is a very different competitive landscape from retail.

And small businesses have already voted. 76% of Goldman’s small-business cohort currently use AI. 60% of tax professionals use it weekly for research. Nearly half of QuickBooks’ AI bank-feed users report getting 12 hours back a month.

The small business orchestrator, when it arrives in full, will be the equivalent of a CFO, bookkeeper, tax advisor, insurance broker, and commercial banker, running continuously on the founder’s data, at a price point below what any one of those services costs today. That product is technically feasible now. Intuit, Ramp, and the AI-native entrants are racing to shape it. The winner has not been decided.

Where each player wins under each regulatory scenario

Regulation matters. It shapes the pace, the sequence, and the enforceable claims. But it does not change the direction. Customer behaviour wins either way. What regulation does is decide who captures the value the customer creates.

One correction to the record before the matrix, because the ground moved this spring. Canada’s open banking law is no longer pending. Bill C-15 received Royal Assent on March 26, 2026, enacting the full Consumer-Driven Banking Act. The legislative risk is gone. What remains is implementation risk: the Bank of Canada said in March it would be premature to commit to a launch date; Phase 2 write access is targeted for mid-2027 and contingent on the Real-Time Rail being live at scale, and the RTR’s official target is late 2026.

Here is the matrix.

Best case for institutions and users (implementation lands on schedule)

Timeline: CDBA Phase 1 read access will be operational by late 2026 or early 2027. RTR launches near its late-2026 target. Phase 2 write access lands mid-2027. OSFI’s E-23 model-risk guideline takes effect May 2027. The rewritten US 1033 rule finalizes in 2027. UK Open Banking evolves into Open Finance by 2027-28.

What it means for users: Full data portability. Customers can consent to an AI orchestrator seeing and acting across every institution. Trust standards become table stakes. Insurance products emerge for AI-error liability. The trust gradient closes faster because the regulatory floor is high enough to prevent the worst outcomes.

Big banks: Compete on being the recommended supplier. Invest in API quality, product competitiveness, and trust brand. The banks that lean into being the best supplier keep their customers. The banks that resist API access get routed around by orchestrators and slowly lose share, then all at once.

Challengers: Win on product excellence in specific verticals. The regulatory infrastructure lets them compete on level terms with the big banks because the interface layer is neutral. Challengers with strong single-vertical products outcompete generalist big banks in that vertical.

Fintechs: Win on workflow ownership. The regulatory framework makes their APIs more valuable. Plaid becomes more critical. Intuit becomes a plausible small-business orchestrator entrant, given its data position on both sides of the ledger and its head start on actions within ChatGPT.

Orchestrators: Win the customer relationship layer. Consolidation to two to five players. Outcome-based pricing emerges as a differentiator among orchestrators. Winner-take-most dynamics play out.

Worst case for institutions (implementation slips, but customer behaviour continues)

Timeline: The Bank of Canada takes through 2027 to stand up Phase 1. RTR slides into 2027 and drags write access toward 2028. The US 1033 rule stays stuck in litigation and rewrite. Open banking effectively parked.

What it means for users: Slower adoption, but adoption continues. In markets where the AI orchestrator is already live, customer behaviour does not wait for the regulator. Customers who want AI financial services get them via workarounds. Screen scraping continues until the ban bites. Voluntary bank partnerships with orchestrators emerge outside the regulatory framework. The trust gradient closes more slowly, so adoption of action delegation runs a year or two behind schedule. But the direction of travel is set. Regulation shapes the speed and the mechanics. It does not reverse the flow.

Big banks: This is superficially the best case for them, but only superficially. They keep the interface for longer. But they still lose share to the challengers and orchestrators that partner voluntarily. The banks that use the delay to strengthen their position with suppliers win. The banks that use the delay to double down on the app-as-primary-interface lose more than in the fast scenario, because they invest against the wrong strategy for longer.

Challengers: Slower growth. But the ones that rent distribution inside the interfaces where customers already live gain relative ground. Wealthsimple’s April 2026 partnership with X, which turns a stock tag into a one-tap path to a Wealthsimple trade, is the shape of the move: go where the attention is, rather than waiting for the regulator to finish the pipes.

Fintechs: Win either way. Slower growth in the delayed scenario, but their structural position is undamaged. Plaid, Intuit, and the AI-native infrastructure players continue to be the workflow layer every other player needs.

Orchestrators: Slower growth in Canada. Roughly on schedule in the US, where the 1033 rule is stalled but the market already runs on Plaid’s voluntary connections. Global orchestrators are largely unaffected by North American regulatory delays because they are already global products.

The scenario nobody wants to name: regulation lands but adoption stalls

This is the third scenario, and it deserves its own row. It is what happens if a high-profile AI financial failure occurs early enough to reset consumer trust for a generation. A Klarna-style reversal, but bigger. A regulator responds by imposing mandatory human-in-the-loop requirements. Or a viral consumer story about an AI moving money to the wrong account.

There is a named mechanism under this scenario. Dietvorst, Simmons and Massey called it algorithm aversion: people abandon an algorithm after seeing it err once, even when it still outperforms the humans they retreat to. And the precedents are already on the books. Klarna claimed in 2024 that its AI assistant was doing the work of 700 agents, then spent 2025 admitting quality had slipped and hiring humans back. Knight Capital showed the speed problem in 2012: one bad deployment, $440 million gone in 45 minutes. Automated money fails at machine speed. Moffatt v. Air Canada showed where liability lies: Air Canada’s chatbot invented a bereavement-fare policy, and in 2024 a British Columbia tribunal ordered the airline to honour it, rejecting the argument that the chatbot was a separate legal entity responsible for its own actions. Courts will pin the agent’s words on the principal. (Yes. Air Canada again.)

Probability: This is my authored estimate, not a sourced projection. I put it at roughly 20 to 30% over the next two years, on the basis that Klarna has already shown this pattern at a moderate scale, and the current generation of consumer-facing AI products has a non-zero rate of visible errors. Reasonable people can disagree with the number. Anyone modelling this seriously should run the scenario at multiple probability weights.

What it means for users: Two-year delay on trust closure. Faster maturation of the trust and liability frameworks, because the failure forces the industry to build them. The thesis survives it with a delay. And Dietvorst’s follow-up work found the repair: give people even modest control over the algorithm’s output and adoption recovers. That is why serious agentic products ship as confirm-the-exceptions rather than full autopilot, and why the weekly exception review is the trust architecture rather than a transitional interface.

Big banks: This is the best case for them. A trust event that makes customers cautious about AI orchestrators is a gift to institutions with existing trust equity. The banks that use the pause to build supplier readiness win. The banks that read the pause as permission to keep doing what they were doing lose.

Challengers: Slower. But the ones that positioned as trust-forward AI gain. The we-do-this-safely position becomes valuable.

Fintechs: Compliance workflow becomes more valuable. Fintechs that can prove trust standards to institutional partners gain.

Orchestrators: Consolidation happens faster. The players that survive the trust event dominate the recovery. Weaker orchestrators either fold or get acquired.

What is different this time

Every time the industry has faced a new interface (online banking, mobile banking, robo-advisors, PFM apps), the argument was the same. Customer behaviour is changing. And every time, the customer behaviour changed less than the industry predicted.

There are three specific reasons this time is different.

One. The interface finally does the reasoning.

Every prior interface required the customer to translate what they wanted into what the interface understood. Type the query. Click the category. Open the right app. Know which button to press. LLMs are the first interface that translates in the other direction. The customer says what they want in the language they already have. The system does the rest.

That is a category change. It removes the structural assumption underlying every previous consumer finance product: that the customer would learn to use the tool. They will not. They never have. The tool has to learn to use the customer.

Two. The customer behaviour has been there the whole time.

The one-fifth to one-half of American adults using AI for money is not a demand curve building from zero. Neither is three-quarters of Goldman’s small-business cohort. These are demand curves that have always been there, only now being captured for the first time by a product that speaks the way the customer thinks. TD’s own tracking went from 10% to 55% between surveys. Online banking and mobile banking each needed the better part of a decade to cross half of American adults. The customers were already using AI for other things (writing, research, coding, therapy) and applied it to their money the moment the products allowed them to. The friction sat on the product side the whole time.

Three. The competitive dynamic favours the new entrants for the first time.

In every prior wave, the incumbents had structural advantages that new entrants had to overcome: brand, trust, distribution, capital, regulatory relationships. Digital-only challengers spent a decade fighting these advantages.

In the AI orchestrator wave, the structural advantages flip. The orchestrator does not need brand equity in banking. It borrows the customer’s existing trust in the orchestrator’s AI. It does not need distribution. The customer is already there. It does not need capital in the traditional sense. The marginal cost of serving another customer is near zero. And it does not need regulatory relationships in banking, because it does not do banking. It routes to banks that do.

For the first time in fifty years, the entrants have the structural advantage in a category that consumer banking has controlled since its inception. One caveat rides along: structural advantage says nothing about structural virtue. See the neutrality problem above.

What each type of player should do next

The player-type-times-scenario matrix says what each type does in each regulatory scenario. The larger strategic move is simpler.

Big banks: Stop trying to be the interface. That fight is over. Start being the best supplier. That fight is winnable, and the prize is durable revenue on the other side of the transition. Build API quality first. Product competitiveness second. Trust brand third. Every dollar spent on the app-as-primary-relationship is a dollar spent defending a position that is already lost.

Challengers: Pick your vertical. The generalist challenger bank category will not survive the orchestrator transition. The vertical winners will. Wealthsimple has picked investing plus adjacent savings. KOHO has picked spending plus habitual cashback. Chime picked salary redirect and overdraft. Each of them is now defensible. The challengers trying to be miniature big banks are not. And hedge the neutrality risk: take orchestrator volume, keep a direct door.

Fintechs: Own the moment. If you are Plaid, own bank connectivity. If you are QuickBooks, own the small business ledger. If you are TurboTax, own tax. The temptation to expand horizontally is real, and it is usually wrong. Depth in one moment beats breadth across the customer’s life.

Orchestrators: Move faster. The consolidation curve favours early movers with strong interfaces. The regulatory delay in Canada is a competitive advantage for anyone who can partner directly with institutions to work around it. The trust event is coming, and Dietvorst already published the design brief: keep the customer in visible control of exceptions. The orchestrator that has invested in safety, transparency, and provably unbought recommendations before the event wins the recovery.

Everyone: Stop building dashboards. The dashboard era is ending, and it was always a bad idea. Build defaults. Build mechanisms. Build the automatic behaviour that the dashboard used to confirm. Save More Tomorrow raised its savings rate from 3.5% to 13.6% because it acted in the customer’s best interest. Madrian and Shea’s defaults increased participation from 37% to 86% because they acted on behalf of the customer. Every product decision in the next three years should be measured against that standard.

Where this ends up

By 2028, the default interface for retail personal finance in North America is conversational. The dashboard survives as a compliance artifact and a receipt of what the automated behaviour did. The bank app becomes something you open to confirm, not to manage.

By 2028, the same is true for small business finance. The AI orchestrator runs the ledger, expense management, payroll, tax filing, and working capital decisions. The founder speaks to it once a week and confirms exceptions. The QuickBooks plus TurboTax plus Credit Karma stack, or whichever product wins the small business orchestrator race, becomes the AI-native fabric Intuit has been building toward since it sunset Mint.

By 2031, the primary financial relationship for most North American adults under 45 is with an AI orchestrator. Banks continue to exist. They continue to be profitable. They become suppliers to the relationship, and ownership sits with the layer above.

Airlines still exist, still fly planes, still make money. Aeroplan still owns the customer. And the fates on the supplier side played out exactly as the moats predict. Emirates never joined an alliance and thrived because it holds a position. Southwest stayed off the aggregators and thrived, because Southwest owns a position. The carriers that disappeared had neither a seat at the coordination layer nor anything a customer would cross the street for.

You can own the layer. You can own a position the layer has to route to. What you cannot survive is being an undifferentiated supplier inside someone else’s interface.

Consumer finance is about to run the same play. The winners already know which of the two things they are building. The losers are still building a better dashboard.