See all articles

The financial advisor’s Claude AI playbook

A practical guide for financial advisors: which workflows to encode as AI skills, where the failure modes hide, and how skill libraries compound into real value.

By Devon Klumb, CFP®, BFA™

Updated Jul. 21, 2026 |

The financial advisor’s Claude AI playbook Jul 21, 2026 2:49:51 PM A practical guide for financial advisors: which workflows to encode as AI skills, where the failure modes hide, and how skill libraries compound into real value. An analysis: In the first piece in this series, I argued that AI middleware is being commoditized while durable value lives one layer down, in regulated execution. The natural next question is how to actually build on that frame. This piece is for the firms ready to answer it. Not a how-to-install. A thinking guide: Which workflows to encode, what good output looks like, where the failure modes hide, and how the work compounds as a strategic asset rather than a tooling expense. Less code, more practice design. AI for financial advisors is a workflow design problem Stop thinking like a buyer. Most advisors evaluate AI as a buying decision: Which platform, what features, what price. That’s the right frame for a CRM, but the wrong one for an AI operating layer. The right frame is practice design. What work in your firm is repeatable, document-heavy, pattern-driven, knowledge-dependent? What do you do every week with the same shape but different inputs? What do you delegate poorly because there’s no one to delegate to? Skills are not features. They are encoded versions of the firm’s existing processes. The advisors who get the most out of this aren’t the ones who buy the most plugins. They’re the ones who can describe their workflows clearly enough to encode them. Five categories of advisor work that map to this Category What it means Synthesis Pulling information from many sources into one coherent picture. Pre-meeting briefs, quarterly review packets, onboarding fact-finders. Anywhere the answer exists across systems but no one has time to assemble it. Drafting First-pass written work that a human refines. Client emails, plan summaries, investment proposals, meeting follow-ups. The skill produces version one; the advisor produces version two. Pattern recognition Surfacing things that need attention without telling the advisor what to do. Portfolio drift, tax-loss candidates, plan gaps, clients who haven’t been contacted in N days. The skill flags; the advisor decides. Routing and triage Daily prioritization. Which clients need a call. Which tasks should escalate. Which prospects deserve attention. Memory Turning the firm’s accumulated experience into something queryable. What is our position on annuities? What did Mrs. Lee’s last plan say about Roth conversions? Ask-the-firm as a real internal product. These five share clean inputs, reviewable outputs, and a recognizable “good enough” bar. Workflows that don’t fall cleanly into one of these usually aren’t good first skills. What AI workflows work for financial advisors—and what fails Skills that work have clean structured inputs, reviewable outputs, and a review loop where corrections fold back into the skill file. Skills that fail depend on judgment the model can’t make, have no review step, or pull from stale data. Language matters more than it looks. A tax skill should not say “Recommend a Roth conversion.” It should say “Identify facts that may warrant advisor review, summarize assumptions, list open questions, and draft a client-facing explanation for professional review.” A portfolio skill should not say “Execute a rebalance.” It should say “Identify drift, tax considerations, restrictions, cash needs, and questions for the advisor; then prepare a review packet.” That distinction is the line between intelligence and execution. A day in the life Imagine an advisor at a five-person RIA, two months into running a basic Claude stack. Morning: The daily briefing skill has already run. Three meetings today—for each, client context, recent emails, last quarter’s review notes, current portfolio state, two planning observations, a draft agenda. The advisor scans, edits, notes the second client just had a liquidity event. The agenda updates. Late morning: The meeting runs. Afterward, the meeting-notes skill turns rough notes into action items, a draft follow-up email, and three flagged items for the next planning review. The advisor reviews, removes one misread item, sends the email. Afternoon: A Roth-conversion question. The tax-review skill produces facts, assumptions, open questions, and drafts a client-facing explanation that’s marked for professional review. The advisor refines and routes to the firm’s tax specialist before anything goes to the client. Nothing here is unfamiliar advisor work. What’s different is cycle time, consistency across advisors, and the firm’s own knowledge—encoded in skill files—doing the heavy lifting on the repeatable parts. The maturity curve Stage What it looks like 1 Single-use One advisor, one workflow. Learning what good output looks like. 2 Team Same skill, deployed across all advisors. House style lives in a file. 3 Connected Skills call other skills. Output of one becomes input to the next. 4 Agentic Multi-step automation with human review at each gate. 5 Firm-as-product Skill library as part of the firm’s value prop. Most firms should plan to spend the first six months in Stages 1 and 2. Trying to build Stage 4 before Stage 1 produces fragile automation nobody trusts. Two prerequisites: data and governance The plugins don’t enforce either. They’re your work. Data. A skill is only as good as the data it can see. A meeting-prep skill pulling from a half-empty CRM produces half-empty briefs. The work of getting your data into a state skills can rely on is bigger than the work of building the skills. CRM hygiene, document organization, authoritative sources—unglamorous but prerequisite. The good news: This work has value even if the AI strategy never materializes. Governance. A skill file can tell the model to escalate. It can’t make that happen at runtime. Approval gates, audit logs, separation of duties, and reviewer sign-off have to be built around the model, not inside the prompt. A firm deploying these without checkpoints is taking on real liability. The compliance officer’s job doesn’t get easier, but it does evolve. Skills are intellectual property A SKILL.md file is a process document with teeth. Read end to end, a firm’s skill library is a fingerprint of how that firm operates. There are two main implications. Skill libraries grow—each correction is permanent; improvements depend on firm discipline rather than vendor roadmap. And skill libraries become acquisition assets: A mature library means a portable operating model. Faster ramp for new advisors. Cleaner integration of acquired firms. Higher capacity per advisor. It changes the firm’s value as an operating business. A minimum viable advisor-AI stack Component Purpose Claude financial-services plugins Baseline finance and workflow skills Wealth-management skills Reviews, plans, reports, proposals, rebalancing, TLH Microsoft 365 integration or Google Workspace Workflows inside Excel, Word, Outlook, PowerPoint or the Google equivalent CRM connector Client records, tasks, notes, service model Email / calendar connector Meeting context and follow-up workflow Document connector Tax docs, statements, plans, agreements, prior notes Portfolio / custodial data feed Holdings, balances, tax lots, account models Firm-specific skills Investment philosophy, tax process, tone, templates Approval workflow Human review before client-facing or regulated action Audit log Source data, prompts, outputs, edits, approvals It’s not an afternoon project, but it’s also no longer a multi-year build. The takeaway The advisor who treats AI as a tooling decision will get tooling: useful, fungible, increasingly cheap. The advisor who treats AI as a practice-design exercise will get an operating system: slower to build, harder to copy, integrated with how the firm actually serves clients. The first article argued the intelligence layer is becoming a foundation-model utility. This one is about how to build inside that reality. More to come.

An analysis:

In the first piece in this series, I argued that AI middleware is being commoditized while durable value lives one layer down, in regulated execution. The natural next question is how to actually build on that frame.

This piece is for the firms ready to answer it. Not a how-to-install. A thinking guide: Which workflows to encode, what good output looks like, where the failure modes hide, and how the work compounds as a strategic asset rather than a tooling expense. Less code, more practice design.

AI for financial advisors is a workflow design problem

Stop thinking like a buyer. Most advisors evaluate AI as a buying decision: Which platform, what features, what price. That’s the right frame for a CRM, but the wrong one for an AI operating layer.

The right frame is practice design. What work in your firm is repeatable, document-heavy, pattern-driven, knowledge-dependent? What do you do every week with the same shape but different inputs? What do you delegate poorly because there’s no one to delegate to?

Skills are not features. They are encoded versions of the firm’s existing processes.

The advisors who get the most out of this aren’t the ones who buy the most plugins. They’re the ones who can describe their workflows clearly enough to encode them.

Five categories of advisor work that map to this

Category	What it means
Synthesis	Pulling information from many sources into one coherent picture. Pre-meeting briefs, quarterly review packets, onboarding fact-finders. Anywhere the answer exists across systems but no one has time to assemble it.
Drafting	First-pass written work that a human refines. Client emails, plan summaries, investment proposals, meeting follow-ups. The skill produces version one; the advisor produces version two.
Pattern recognition	Surfacing things that need attention without telling the advisor what to do. Portfolio drift, tax-loss candidates, plan gaps, clients who haven’t been contacted in N days. The skill flags; the advisor decides.
Routing and triage	Daily prioritization. Which clients need a call. Which tasks should escalate. Which prospects deserve attention.
Memory	Turning the firm’s accumulated experience into something queryable. What is our position on annuities? What did Mrs. Lee’s last plan say about Roth conversions? Ask-the-firm as a real internal product.

These five share clean inputs, reviewable outputs, and a recognizable “good enough” bar. Workflows that don’t fall cleanly into one of these usually aren’t good first skills.

What AI workflows work for financial advisors—and what fails

Skills that work have clean structured inputs, reviewable outputs, and a review loop where corrections fold back into the skill file. Skills that fail depend on judgment the model can’t make, have no review step, or pull from stale data.

Language matters more than it looks.

A tax skill should not say “Recommend a Roth conversion.” It should say “Identify facts that may warrant advisor review, summarize assumptions, list open questions, and draft a client-facing explanation for professional review.” A portfolio skill should not say “Execute a rebalance.” It should say “Identify drift, tax considerations, restrictions, cash needs, and questions for the advisor; then prepare a review packet.”

That distinction is the line between intelligence and execution.

A day in the life

Imagine an advisor at a five-person RIA, two months into running a basic Claude stack.

Morning: The daily briefing skill has already run. Three meetings today—for each, client context, recent emails, last quarter’s review notes, current portfolio state, two planning observations, a draft agenda. The advisor scans, edits, notes the second client just had a liquidity event. The agenda updates.

Late morning: The meeting runs. Afterward, the meeting-notes skill turns rough notes into action items, a draft follow-up email, and three flagged items for the next planning review. The advisor reviews, removes one misread item, sends the email.

Afternoon: A Roth-conversion question. The tax-review skill produces facts, assumptions, open questions, and drafts a client-facing explanation that’s marked for professional review. The advisor refines and routes to the firm’s tax specialist before anything goes to the client.

Nothing here is unfamiliar advisor work. What’s different is cycle time, consistency across advisors, and the firm’s own knowledge—encoded in skill files—doing the heavy lifting on the repeatable parts.

The maturity curve

Stage	What it looks like
1 Single-use	One advisor, one workflow. Learning what good output looks like.
2 Team	Same skill, deployed across all advisors. House style lives in a file.
3 Connected	Skills call other skills. Output of one becomes input to the next.
4 Agentic	Multi-step automation with human review at each gate.
5 Firm-as-product	Skill library as part of the firm’s value prop.

Most firms should plan to spend the first six months in Stages 1 and 2. Trying to build Stage 4 before Stage 1 produces fragile automation nobody trusts.

Two prerequisites: data and governance

The plugins don’t enforce either. They’re your work.

Data. A skill is only as good as the data it can see. A meeting-prep skill pulling from a half-empty CRM produces half-empty briefs. The work of getting your data into a state skills can rely on is bigger than the work of building the skills. CRM hygiene, document organization, authoritative sources—unglamorous but prerequisite. The good news: This work has value even if the AI strategy never materializes.

Governance. A skill file can tell the model to escalate. It can’t make that happen at runtime. Approval gates, audit logs, separation of duties, and reviewer sign-off have to be built around the model, not inside the prompt. A firm deploying these without checkpoints is taking on real liability. The compliance officer’s job doesn’t get easier, but it does evolve.

Skills are intellectual property

A SKILL.md file is a process document with teeth. Read end to end, a firm’s skill library is a fingerprint of how that firm operates.

There are two main implications. Skill libraries grow—each correction is permanent; improvements depend on firm discipline rather than vendor roadmap. And skill libraries become acquisition assets: A mature library means a portable operating model. Faster ramp for new advisors. Cleaner integration of acquired firms. Higher capacity per advisor. It changes the firm’s value as an operating business.

A minimum viable advisor-AI stack

Component	Purpose
Claude financial-services plugins	Baseline finance and workflow skills
Wealth-management skills	Reviews, plans, reports, proposals, rebalancing, TLH
Microsoft 365 integration or Google Workspace	Workflows inside Excel, Word, Outlook, PowerPoint or the Google equivalent
CRM connector	Client records, tasks, notes, service model
Email / calendar connector	Meeting context and follow-up workflow
Document connector	Tax docs, statements, plans, agreements, prior notes
Portfolio / custodial data feed	Holdings, balances, tax lots, account models
Firm-specific skills	Investment philosophy, tax process, tone, templates
Approval workflow	Human review before client-facing or regulated action
Audit log	Source data, prompts, outputs, edits, approvals

It’s not an afternoon project, but it’s also no longer a multi-year build.

The takeaway

The advisor who treats AI as a tooling decision will get tooling: useful, fungible, increasingly cheap.

The advisor who treats AI as a practice-design exercise will get an operating system: slower to build, harder to copy, integrated with how the firm actually serves clients.

The first article argued the intelligence layer is becoming a foundation-model utility. This one is about how to build inside that reality. More to come.