On August 3, PwC published its 2026 Financial Services Workforce AI Survey, based on responses from 1,004 director-level-and-above leaders at US financial services firms with $500 million or more in revenue. Respondents were split evenly across banking and capital markets, insurance, asset and wealth management, and private equity.
It's a US survey. It's still directly relevant to any Canadian financial institution reading this, because the workforce assumptions US firms are making right now will shape vendor roadmaps, model pricing, talent markets, and regulatory expectations that Canadian lenders will run into by the end of 2027.
The one-line summary of the report: financial services firms are moving aggressively on AI, most think they're still not moving fast enough, and the workforce plan behind the aggression is thinner than the aggression itself would suggest. PwC's own framing is unusually blunt for a Big Four survey report: firms "may say they're planning for an AI-enabled workforce, but most really aren't. They're often planning for a smaller workforce, and hoping AI fills the gap." Those are not the same exercises.
The findings are worth walking a board through:
90% of FS executives say firms need to get more comfortable moving quickly in the age of AI.
70% say their own organization is moving faster to remain competitive.
77% say their organization is not moving fast enough to keep pace with AI innovation. Those last two numbers coexist inside the same firms; the aggregate mood is speed plus dissatisfaction with the speed.
78% expect their workforce to shrink by at least 20% over the next five years.
42% have done enterprise-wide workforce modeling for AI. Only 50% of that group has modeled the impact of redesigning processes or workflows around AI.
91% are increasing compensation for employees with AI skills. 58% will tie compensation directly to AI-enabled productivity. 86% agree AI skills training is more valuable than an MBA for many new hires.
77% say most of their AI investments are not delivering measurable ROI.
41% cite fragmented or low-quality data as the biggest barrier to scaling AI across the workforce.
90% say employees using AI outside centrally governed tools (shadow AI) is creating regulatory risk. 35% say it's happening at their firm to a significant extent.
27% say the CEO and board own the risk of material harm from AI agents. 16% say technology. 15% say risk and compliance. 12% say the business unit. 8% say the human who acted on the AI output. There is no consensus.
Two findings sit side by side and deserve to be read together. 78% of firms expect their workforce to shrink at least 20% inside five years. 77% of firms say most of their AI investments are not yet delivering measurable ROI. The first is a confident forecast. The second is an admission that the mechanism supposed to produce the first isn't proven yet at their own firm.
Add a third number: only 42% have done any enterprise-wide workforce modeling, and only half of that group has looked at what happens when you actually redesign processes around AI. So, in most firms, the confident five-year workforce forecast is being made on the back of no formal modeling of how the workforce would need to change. That is not a plan. That is a hope with a spreadsheet attached.
The report is polite about this and offers the useful reframe: "There's a difference between modeling how many people you can cut and designing the workforce you'll actually need." The distinction sounds obvious. It is obvious. The survey shows it isn't being made.
The survey is US-based. Three findings translate directly to Canadian lenders, credit unions, and fintechs, and a couple translate with a wrinkle.
41% of leaders in the survey named fragmented or low-quality data as their number-one barrier to scaling AI. That is one place where Canadian firms are structurally slower than US peers, not faster. Canadian lenders operate under PIPEDA, Quebec's Law 25, OSFI's B-10 third-party risk guideline, and (as of this year) OSFI's agentic AI bulletin. Data movement, retention, cross-border processing, model-of-record documentation, and third-party inventory obligations all bite before the fun modeling work starts. The firms that will actually get AI ROI in Canada in 2026 and 2027 are the ones that treated data hygiene, lineage, and privacy-safe experimentation as strategic groundwork two years ago, not as compliance overhead.
90% of PwC's respondents said employees using AI tools outside centrally governed options is creating regulatory risk. 35% said it's happening at their firm to a significant extent. For a Canadian lender, shadow AI touches every OSFI expectation currently on the table: model risk management, third-party risk under B-10, data privacy under PIPEDA and Law 25, and the accountability expectations in OSFI's agentic AI bulletin. A well-meaning underwriter pasting borrower details into a consumer LLM to "just check a wording" is a regulatory finding waiting to happen. Canadian firms should treat the 35% "significant extent" number as a floor, not a ceiling, for their own baseline until they've measured it.
The survey's most striking non-answer: no role has clear ownership of AI agent risk. 27% CEO/board, 16% technology, 15% risk/compliance, 12% business unit, 8% the person who executed the output. In the US, that's a debate. Under OSFI's agentic AI bulletin, it's a required answer, and "no consensus" is not a defense.
Regardless of vendor, stack, or headcount plan, three lenses keep working when a report like this lands on a Canadian CEO's desk.
PwC's headline advice is to buy AI talent, upskill AI talent, pay more for AI talent. All of that is true, and all of it puts the specialized signal in your firm's future workforce, which is precisely the layer most exposed to poaching, replacement, or automation. The alternative worth noting is that the signaling signal that actually matters in lending is the lender's own credit policy and the corrections its underwriters make to model outputs, not the model beneath any tool. Agents trained on that signal let a smaller team do more, and the moat stays inside the lender rather than in a workforce a competitor can hire from you next quarter.
The 41% "fragmented data" finding is the tell that most firms are still trying to solve AI with more data pipelines. In a PIPEDA + Law 25 + B-10 environment, a privacy-safe synthetic twin of a lender's own book lets strategy be pressure-tested for channel shifts, credit policy changes, and product-migration scenarios without moving raw customer records. It also happens to be the fastest way to get from "we're modeling workforce impact" to "we're modeling process redesign," which is the gap PwC identifies inside the 42% who've modeled anything at all.
The report's honest admission that most AI investments aren't yet delivering measurable ROI has one very common cause: firms tried to modernize the plumbing at the same time as they tried to deploy the agents. Ripping and replacing a working LOS to hang AI on top of it is the wrong sequence. Agents that sit on top of the system a lender already runs are easier to inventory, easier to audit under the agentic AI bulletin, and much faster to prove ROI on because the baseline is the current process, not a still-being-built one.
Three exercises that any CEO, COO, or head of underwriting can run without hiring a consultant:
The plan-vs-hope test. Take your current five-year workforce projection and ask: which pages describe the roles, skills, and pipelines you'll actually need, and which pages describe headcount you plan to remove? If the second set is longer, you're in the 78% and you don't yet have an AI-enabled workforce plan; you have a reduction plan.
The shadow AI census. Ask front-line teams (anonymously) which AI tools they've used for work-related tasks in the last 30 days. Compare the answer to your enterprise-approved list. Whatever gap you find is the size of your OSFI conversation.
The accountability sentence. Write, in one sentence, which role at your firm is accountable for a material-harm outcome from an AI agent's decision. If the sentence isn't clean, PwC's finding of no consensus applies to you, too, and the agentic AI bulletin will draw that out during your next OSFI review anyway.
None of the above needs a vendor. All three change the quality of the AI conversation on your executive team from "we're doing a lot of AI things" to "we know what we're actually doing, what we're not, and what we owe a regulator."
PwC's report is worth the 20 minutes it takes to read in full. The most useful thing about it is that it does not try to sell a smoother story than the data supports. 90% moving fast, 77% not fast enough, 78% shrinking headcount, 77% no measurable ROI, 42% doing the actual modeling: that is a candid picture of an industry that has decided the AI question and is still working out the AI answer.
For Canadian lenders and financial institutions, the takeaway is not the workforce number. It is the sequence. Design the workforce you need before you cut the one you have. Fix the data and the governance before the modeling. Put the specializing signal inside your credit policy where a competitor can't hire it, not in a workforce where they can. Every one of those choices is easier to make in August 2026 than in August 2027.
1,004 director-level-and-above leaders at US financial services firms with at least $500M in revenue, surveyed May 12 to 22, 2026. Respondents were evenly split across banking and capital markets, insurance, asset and wealth management, and private equity. Full methodology and findings.
Yes, the sample is US firms. The workforce, vendor, and regulatory pressures it describes still translate directly to Canadian financial institutions, and in a few cases (shadow AI, agent accountability, data quality) the Canadian regulatory environment makes the issues sharper, not softer.
Shadow AI is employees using AI tools outside their firm's centrally-approved and governed options. 90% of PwC's respondents said it creates regulatory risk. In Canada, shadow AI touches PIPEDA, Quebec's Law 25, OSFI's B-10 third-party guideline, and OSFI's agentic AI bulletin all at once. A single underwriter pasting borrower details into a consumer LLM is a regulatory finding waiting to happen.
The unique judgment a lender applies to raw credit information: which corrections underwriters make to model outputs, which exceptions get approved, which policy tweaks are made after loss data comes back. That signal is what distinguishes a lender from a vendor's baseline model. It should be captured, versioned, and used to train the lender's own agents, not lost into ad hoc human decisions or generic vendor tools.
Some are, some aren't. The honest picture: the firms that started on data hygiene, model risk management, and privacy-safe experimentation two years ago are now going fast. The firms treating Consumer-Driven Banking and the agentic AI bulletin as compliance projects rather than strategic groundwork are the ones the PwC-style ROI gap is going to catch up with first.
Culture. The report captures employees are hesitant (44% concerned about job security, 43% use AI only when required, 40% feel overwhelmed, 34% cited change fatigue) but doesn't quantify the effect of communicating a clear workforce plan on those numbers. Anecdotally, the firms that have named what the future looks like, and how their people fit into it, are the ones seeing hesitancy drop the fastest.