Paving the Cow Path

Picture a cow crossing a field. She isn’t going anywhere in particular. She swings wide around a boulder, cuts through a gap in the fence and wanders down to the creek. People follow her tracks, carts follow the people, and one day someone paves it. The boulder got hauled off decades ago, but everyone still takes the long way around a rock that isn’t there, and they’ve gotten very good at it.

Back offices have roads like that, and AI is the best paving crew ever built. Point it at the way work gets done today and it’ll smooth every curve, including the ones that only exist because of the boulder. This issue is about walking the route before you call the crew.

Also inside: Build This has Claude read the sell-side inbox before you get in. Asked This Week takes on whether to replace two retiring staff or see what AI can do. What Moved has Schwab’s Enterprise credit deadline and Goldman pulling AI coding tools from its traders.

This is issue 6 of AI Under Management. It runs every other week for people at wealth and asset management firms who want real work out of this without creating a problem for themselves. If you know someone who’d want it, send it their way.

BUILD THIS

The sell-side morning digest

By the time the market opens, the inbox on the investment side is full of broker notes, morning calls and trade-desk color, and much of it has nothing to do with what you own. Build a scheduled task that reads it before you get in. Connect your email, set it to run on weekdays before the open, and give it your coverage list. None of this needs client data. The task should tell Claude to:

Read everything. Every broker and trade-desk email since the last run, in every folder your email rules sort them into, attachments included.

Lead with the top ten. The ten items that matter most for your coverage, ranked, each with one line on why it matters.

Then the patterns. Themes more than one broker is raising, and a short section on positioning and sentiment from the trade-desk notes.

Then the rest. A table of news on each name you cover, followed by a one-line summary of every email with a link back to the original.

Stay in its lane. Report what each broker said and who said it. No views or recommendations of its own.

Keep a record. Save each digest to a dated file so you can look back at what was said and when.

How it’s gone. An analyst I work with built a version that checked every note against his own scoring and risk reports. It was so thorough it took forever to read. Once he told it to give him only the ten things that mattered most to him, it earned its place: on a typical morning, two to four of those were things he’d have missed. A colleague’s first version skipped broker mail his rules had filed into other folders until he told it to check every folder and open every attachment. Once yours works, share it as a skill so teammates can build their own for their coverage.

THE MEMO

Paving the Cow Path

In July 1990, Michael Hammer published a Harvard Business Review article with a blunt title: “Reengineering Work: Don’t Automate, Obliterate.” His complaint was that companies had poured money into computers only to speed up old ways of working, much of which shouldn’t have existed at all. His line for it was “It is time to stop paving the cow paths.” AI has made paving cheaper and faster than it’s ever been. The cow paths are still there.

The research hasn’t moved much either. In McKinsey’s March 2025 State of AI survey, redesigning workflows had the biggest effect on whether a company saw an earnings impact from generative AI, out of 25 attributes tested. Only 21% of respondents at companies using it said they’d fundamentally redesigned even some workflows. That survey spans every industry. Looking only at advisory firms, Kitces Research found little relationship between technology investment and revenue productivity, at firms running an average of 12 applications.

Walk the route first

When I sit down with an operations team and walk through a process one step at a time, a few kinds of things turn up. A report gets exported, reshaped by hand and pasted into another file that exists mostly out of habit. A procedure still refers to tools the firm has moved on from. Two people confirm the same number at different points without knowing the other one did. None of it is anyone’s mistake. Each step made sense when it was added.

So before you hand a process to AI, ask four questions about it.

Is every step needed? Follow each step’s output and find who uses it. A step whose output nobody reads is the cheapest saving there is, and it doesn’t need a skill.

Can it be simpler? Two files holding the same information can be one, and every hand-off between them is a chance for the two to drift apart.

Is the data clean? AI works from what you give it. If an exception lives only in someone’s inbox, or the same detail is kept in two systems that don’t match, the automation will be wrong faster and more confidently than a person would be.

Is every system still earning its place? Each tool is another login, another export and another place the numbers can disagree. Now that custodians and platforms are connecting directly to AI, some of what gets retyped from a website can be read from the source instead.

Then pave it

None of this means cutting the checks that matter. Anything that moves client money or creates a record keeps its review. The goal is one check that works instead of several that check each other.

It helps to have someone outside the process asking the basic questions, starting with the simplest one: what would break if we stopped doing this? Someone who has run a process for years has every reason to trust it. A new hire doesn’t, and neither does a well-built skill.

AI on a clean process is better than AI on an antiquated one. Walk the route, take out the curves that only exist for the boulder, and make sure the data feeding it is right. What you pave after that is a process your people trust, and one that’s ready for client data to flow straight from the source.

ASKED THIS WEEK

Two of our support staff retire next year. Do we hire replacements, or see what AI can take on?

Decide after they’ve written down what they do, not before. The bigger risk in a retirement isn’t the empty seat. It’s what leaves with the person: the exception nobody documented, the custodian quirk, the account that always gets handled by hand.

So use the time you have. Have each of them walk through their work while they’re still here, and turn the repeatable parts into skills they help build and check. They’re the best reviewers you’ll ever have for it, because they know what right looks like.

Then look at what’s left. Judgment calls, client contact and the exceptions that should stay with a person. That’s the job you hire for, and it may not be the job that’s leaving.

One thing to start early: if the work touches client data, the setup has to be right before any of it moves into AI.

Have a question you’d like answered here? Reply to this email.

WHAT MOVED

Schwab’s Enterprise credit runs through November

At Schwab IMPACT, Alison Dooher, Schwab’s head of AI for Advisor Services, told reporters that advisors signing up for the Enterprise version of Claude for Financial Advisors by Nov. 30 will get a $4,800 credit, equal to a year of the 20-seat minimum, according to Wealth Management. She said it’s meant to make the seat cost net neutral for smaller firms. There’s also a smaller-teams option that requires showing certain security measures. What it means for you: if the seat minimum is what kept you on a lower plan, run the numbers this month, and get the credit’s conditions in writing before you sign.

Goldman Sachs pulls AI coding tools from its traders

Bloomberg reported on Nov. 2 that Goldman limited Anthropic’s Claude Code and Microsoft’s GitHub coding tools to its engineers, under internal guidance updated in October. Traders had used them for months to monitor client flows and build models. What it means for you: Goldman didn’t stop using AI. It drew a line at people building software they can’t maintain. In a firm where everyone wears several hats, the same line works: anything that runs as code, like a portal, a dashboard or a script, needs an owner who can read it. The rest belongs in Claude itself and in skills, where a person reviews the output every time.

Next issue in two weeks. Forward this to whoever at your firm gets handed the AI question, and reply if you want to talk about any of it.

Tim Diemer
Partner and COO, Bridge AI Consulting

This is general information about published rules and products, not legal or compliance advice. Run your own facts past your own counsel and CCO.

Know a firm that should be working with us? Refer them to Bridge AI and we’ll pay you 10% of their first year.