The question boards keep asking us is not whether to deploy AI. It is whether the ones already deployed are earning their keep, and how anyone would honestly know.
I keep an eye on the three Wall Street firms that have put the most money and the most people through an internal AI programme, because they publish numbers that mid-market buyers do not get out of their peers. Those numbers landed again this year. None of them tell the story you would get from a vendor keynote.
What they tell a buyer is quieter and more useful than the headlines. The rollouts that worked worked because they were pointed inside first, measured against work the firm was already doing, and left to run long enough that the honest answer about productivity could emerge later rather than sooner. If you are evaluating a platform or an agency this quarter, that is the lens, not the hype one.
The three disclosures a buyer should read before signing anything
Three internal GenAI rollouts in financial services have published enough detail in 2025 and 2026 to be read side by side: JPMorgan Chase’s LLM Suite, Morgan Stanley Wealth Management’s AI programme, and Goldman Sachs’s pilot of the autonomous coding agent Devin. All three are internal, aimed at the firm’s own staff rather than at customers. All three come with named executives on the record. All three publish at least one number you can compare.
Those three choices matter, because the picture that emerges is the opposite of the one that got AI into the budget cycle in the first place. These are not customer-facing moonshots. They are workforce enablement programmes, built behind the firewall, measured against the work the firm would have done without them, and talked about with a care that leaves room to be wrong.
Jamie Dimon’s 2025 letter to shareholders, published 7 April 2026, is the clearest frame for the lot of them. “AI, data and technology are key to the future, as is solving for how to implement AI properly and fast,” he writes, before the one line most of the industry reporting pulled out: “while I hesitate to use the word transformational, it is.” Then the sentence that actually matters for anyone writing an AI budget, from the same letter: “AI will affect virtually every function, application and process in the company.” The admission under that is unusual: the pace of AI adoption, Dimon says, “will likely be far faster than prior technological transformations, like electricity or the internet.” (JPMorgan Chase 2025 Annual Report, Chairman & CEO Letter, 7 April 2026.)
Here are the three, in order of ambition and in reverse order of certainty.
JPMorgan put almost half its workforce on one AI surface before it put a customer on one
JPMorgan Chase’s LLM Suite is a proprietary platform that routes employees to third-party large language models (OpenAI and Anthropic) through the bank’s own interface and governance. The thing to notice is where it was pointed. It did not go onto a customer-facing surface. It went onto employees’ desks, inside the firewall, with a path for them to build small internal applications on top of it.
Derek Waldron, JPMorgan’s Chief Analytics Officer, gave McKinsey the sharpest single number any Wall Street firm has published on internal AI use so far. In the McKinsey interview “JPMorgan Chase’s Derek Waldron on building an AI-first bank culture”, he says: “A little under half of JPMorgan employees use gen AI tools every single day,” and that staff use them “in tens of thousands of ways specific to their jobs.” Waldron frames the training ambition around a workforce “more than 300,000-strong,” which lines up with the 320,000+ employee figure Dimon’s letter uses elsewhere.
Three things in that disclosure matter more than the headline percentage. First, “every single day” is a different metric from “has access.” A seat count would be larger and less honest; daily use is what rolls through to actual productivity. Second, the “tens of thousands of ways” line is a confession that the use cases are long-tail, not top-down. Nobody at head office specified them. Third, “a little under half” is still a floor to grow from, not a ceiling to be proud of, which is where Waldron takes the rest of the interview with the internal “AI Made Easy” training programme.
A little under half of JPMorgan employees use gen AI tools every single day.
Derek Waldron, Chief Analytics Officer, JPMorgan Chase, to McKinsey.
The transferable lesson for a buyer: the firm with the most internal AI use on record did not start by picking a customer-facing use case. It started by giving its own people a safe, governed way to use the tools at all, and let the use cases surface from the inside. That is the opposite of how most mid-market programmes are scoped.
Morgan Stanley measures its AI by whether the advisor still opens it next week
Morgan Stanley Wealth Management went narrower on purpose. The firm named one audience, its roughly 15,000 Financial Advisors, and gave them two tools: AI @ Morgan Stanley Assistant (September 2023), a chatbot that fronts the firm’s internal research, and AI @ Morgan Stanley Debrief (June 2024), which joins a client Zoom meeting with consent, takes the notes, drafts the follow-up email and files a Salesforce record.
The adoption number is on the record. In the June 2024 launch note, Morgan Stanley writes that the Assistant has been fully rolled out since September 2023 and that “to date, 98% of Financial Advisor teams have adopted the Assistant” (Morgan Stanley press release, “AI @ Morgan Stanley Debrief launch,” 26 June 2024). That is as high as these numbers go in practice, and it is tied to a specific population of roughly 15,000 people with a visible, measurable job to do.
Jeff McMillan, the firm’s Head of Firmwide AI, is the one framing worth keeping. In the same release he describes AI as “an efficiency enhancing interaction layer that sits between our colleagues and the many applications they interact with such as execution and order entry, CRMs, reporting tools and risk analysis.” The sentence is doing work. It positions AI as a layer that joins up tools that already exist, rather than a replacement for them. That is a buyer-safe posture, and it is the posture that produced the 98% number.
An efficiency enhancing interaction layer that sits between our colleagues and the many applications they interact with.
Jeff McMillan, Head of Firmwide AI, Morgan Stanley.
What is still unsolved, and what Morgan Stanley says plainly: productivity gain is not proven yet, only usage. The firm is careful to publish adoption rather than hours saved or revenue uplift. That is the right order to publish them in. A tool with 98% adoption and no measured gain is a problem that can be fixed. A tool with a published 20% uplift and 12% adoption is a different, worse problem.
The transferable lesson: pick one audience that already shares a workflow, give them a tool that attaches to the workflow at a single point (the client meeting, the research question, the Salesforce entry), and measure whether they keep opening it. Everything else is a derived number.
Goldman Sachs is piloting an AI engineer, which is a different word than deploying one
Goldman Sachs is the one to read for honesty about where AI is not yet. Marco Argenti, the firm’s Chief Information Officer, told CNBC in July 2025 that Goldman had begun piloting Devin, an autonomous coding agent built by Cognition AI, inside its engineering function. In Argenti’s words: “We’re going to start augmenting our workforce with Devin, which is going to be like our new employee who’s going to start doing stuff on the behalf of our developers.” (CNBC via NBC, “Goldman Sachs is piloting its first autonomous coder,” July 2025.)
The number worth clipping sits in the same interview. Argenti said: “Initially, we will have hundreds of Devins and that might go into the thousands, depending on the use cases.” Against a 12,000-strong engineering workforce, “hundreds, maybe thousands” is a careful ratio. The firm is not promising a replacement, and it is not promising a timeline. It is promising a pilot with a defined upside and a clear human supervisor.
The work Devin is being asked to do is the second honest signal. Argenti is clear that it is routine, often tedious engineering: updating legacy code to newer language versions, the kind of task that is well defined, has a known test outcome and is not where senior judgement earns its margin. That is the right place for a first bet on an agent, because the failure mode is contained and the measurement is cheap.
Initially, we will have hundreds of Devins and that might go into the thousands, depending on the use cases.
Marco Argenti, Chief Information Officer, Goldman Sachs, to CNBC.
The transferable lesson: name the task narrowly, keep the agent on the side of the work where mistakes are visible and reversible, and keep a human on the review line. The firm that most has to be careful about engineering risk picked exactly the task where an agent is least risky, and said so publicly.
What the three disclosures share, and what they do not
Three firms, three different ambitions, one pattern. Each put AI onto the inside of the firm first, named one function or workflow, and bolted the tool to work the firm was already doing well. Each published the number that could be falsified (daily usage, team adoption, pilot headcount), and declined to publish the ones that cannot be, yet (dollars saved, revenue added, headcount offset). Each is led by a named executive who is willing to be wrong on the record. That last point is doing more work than it looks like it is.
What the three do not share is any claim that the pilot phase is over. Dimon writes that AI “will affect virtually every function”; he does not write that it has. Morgan Stanley’s own Head of Firmwide AI has told reporters, on the record elsewhere, that measuring the productivity impact will take at least a year. Argenti is piloting Devin with a human in the loop at every step. If you came out of a vendor pitch this week with a sentence like “the leaders are already past measurement,” that sentence is not what the leaders are saying.
What this means on a mid-market WooCommerce estate, in six moving parts
The three cases above are the top of the market. The build section is the one most of our readers actually care about, because it maps the pattern onto a mid-market WooCommerce estate and prices it honestly. A buyer running a 50 to 500 person brand with WordPress and WooCommerce as the content and commerce layer can replicate the shape of what JPMorgan, Morgan Stanley and Goldman did. Not the scale. The shape.
- One internal AI surface, not many. One place employees go to use governed models against company data, modelled on LLM Suite rather than on individual browser tabs. In practice this is a small internal application, hosted next to the WordPress estate, routing to a vendor API with a session log the firm can inspect. Effort: two to three weeks for a usable first version, then an ongoing operational line because the governance never stops being work.
- One concrete workflow, not many. Pick the Morgan Stanley equivalent on your estate: the one workflow a defined group of people repeat every week, that already runs through WordPress or WooCommerce (product description production, customer service response drafts, campaign brief generation, support ticket triage). Attach the AI tool to that workflow at one point, not to five.
- Content and catalogue data, structured. The reason LLM Suite works is that the firm’s own data is reachable in a shape the model can use. On WooCommerce this means product attributes, variations and categories that are modelled properly rather than stuffed into free-text descriptions. The maturity of WordPress is doing real work here, taxonomies, custom fields and REST endpoints that have been in production for a decade.
- A governance layer, not a prompt library. Logging who used what, with which input and which output; a redaction pass on anything client-identifying before it leaves the perimeter; a review step on anything customer-facing. On a WordPress estate this layer is usually a lightweight custom plugin with capability checks, plus a sensible choice of hosting that keeps the inference traffic on a known path.
- A measurement rule you commit to before you start. Pick the number you will publish internally in six months, and promise not to change it. “Percentage of the target team using the tool in a given week” is the Morgan Stanley equivalent. “Hours saved per task” is not, because it is a self-reported number that drifts. The point of fixing the metric early is that it binds you to honesty when the result is thin.
- A pilot scope narrow enough to fail cheaply. This is the Goldman principle. Pick a task where mistakes are visible and reversible, and where a human in the review line is cheap. On WooCommerce, product description drafts and support macro drafts fit. Automated order refunds, pricing changes and inventory writes do not.
Rough budget to pilot a programme in this shape on an existing WooCommerce estate, with the hosting and store already in production: in the region of 60 to 120 engineering hours for the first workflow and the governance layer, plus a running vendor API spend that for a 20 to 50 person team sits in the low hundreds of euros a month. The honest line in the quote is the one that says the first workflow is cheap and the second one shares the governance it paid for.
Where the pattern lands for a buyer evaluating a partner this quarter
If a prospective agency or platform partner is telling you the AI question is already settled, they are selling the wrong thing. Three of the firms with the most to spend, the most to lose and the clearest need to produce a result in public are each in the pilot or scale-out phase of exactly the kind of internal programme a mid-market WooCommerce estate can borrow the shape of. None of them is claiming it is done. All three are claiming it is worth doing.
What we would do for an organisation sitting in that chair this quarter is boring on purpose: pick the one workflow where the measurement is honest, build the governed internal surface it runs on, bolt it to the WordPress and WooCommerce data that already exists, and commit to the number that will tell you later whether the pilot earned the next one. If the right partner is in the room they will push back on the scope before they quote the price. If they will not, that is also useful information.
Last modified: October 20, 2026
