# After You Fire the AI Agency: The 30-Day Migration Playbook (Part 2 of 2)

You ran the 11-question audit and decided to leave the agency. Here's the 30-day plan to extract your data, rebuild the workflows on Claude Pro for $20/month, and keep the lights on during the transition. Specific checklist, real cost math, no fluff.

Author: J.A. Watte
Published: May 14, 2026
Source: https://jwatte.com/blog/blog-post-agency-ai-migration-part-2/

---

You ran the [11-question pre-agency audit](/blog/blog-pre-agency-ai-self-audit-part-1/) on your current AI vendor. They hit 4 of 11. The price is hard to justify. You've decided to migrate off.

Now the harder question: how do you actually leave without breaking the workflows that depend on the agency's stack?

This is the 30-day plan. It assumes you have the agency on a month-to-month contract or you're inside a 30-day notice window. If you're locked into a longer contract, the plan is the same but you run it in parallel with the agency still operating, so the migration is finished before the contract ends.

## The 30-day plan, at a glance

| Days | What you're doing |
|---|---|
| 1-3 | Inventory + data export |
| 4-7 | Workflow priority ranking |
| 8-14 | Rebuild on Claude Pro: top-3 workflows |
| 15-21 | Rebuild on Claude Pro: next 3-5 workflows |
| 22-26 | Parallel run + sanity check |
| 27-28 | Cutover |
| 29-30 | Decommission + final invoice review |

The whole thing takes one focused person about 40-60 hours total. Spread over four weeks, that's 10-15 hours a week. If you're the owner, you'll feel that. If you have an operations person, they can run it as their primary focus for a month.

## Days 1-3. Inventory and data export

Before you can leave, you need to know what's actually running. Most agencies have built out more than the owner remembers. Sit down for two hours and map it.

For each workflow the agency runs, write down:

1. **What it does.** One sentence.
2. **What it connects to.** QuickBooks, Gmail, Stripe, Slack, etc.
3. **What data it touches.** Customer records, invoices, support tickets, etc.
4. **How often it runs.** Continuously, daily, on-event, manual trigger.
5. **What the output looks like.** Where does the result go.

You'll probably end up with somewhere between 4 and 12 workflows. If you have more than 12, you've been over-served and the migration is going to take longer than 30 days.

Then export every piece of data the agency has of yours. Email them with a specific list. The good agencies will export to standard formats (CSV, JSON) within a few business days. The reluctant ones will need a firmer follow-up; if they drag past two weeks, escalate in writing and remind them of the contract's data-portability clause.

Specifically, export:

- Customer/contact records the agency has stored
- Any "memory" or fine-tuned context (the agency's accumulated learning on your business)
- Templates, prompts, and skills the agency built for you (in plain text, not just locked behind their UI)
- Logs of past actions (sent emails, applied tags, etc.)
- Any custom configuration files

Save everything to a single folder on your machine. Back it up to a cloud-storage account you control (not the agency's).

## Days 4-7. Rank the workflows by impact

Not every workflow gets rebuilt. Some you decommission entirely because they were never that useful. Some you replace with cheaper purpose-built tools. Some you rebuild on Claude. The ranking decides what goes where.

For each workflow, score:

- **Owner-time saved per week.** Honest number.
- **Revenue impact.** Does this directly affect collections, sales, customer retention?
- **Risk if it stops.** What happens if it doesn't run for two weeks?

The 3-5 highest-impact workflows go into Claude. Anything below the cut line gets retired or replaced with a single-purpose tool (a Zapier zap, a calendar app, etc.) that's cheaper than continuing to pay for it inside a $3,500/month bundle.

A typical SMB outcome is something like:

- Invoice chase → Claude (high impact, daily)
- Books reconciliation → Claude (medium impact, weekly)
- New customer onboarding → Claude (medium impact, on-event)
- Calendar triage → Zapier or built-in calendar rules (low complexity)
- Slack notifications routing → Slack workflow builder (free, built-in)
- "AI marketing campaign" → retire entirely (was never that useful)

The retire-it list is usually the most painful conversation. You paid for these. Sunk cost. The right move is still to retire them.

## Days 8-14. Rebuild your top 3 workflows on Claude Pro

Sign up for [Claude Pro](https://claude.ai) at $20/month if you don't have it. Install [Claude for Small Business](/blog/blog-claude-for-small-business-walkthrough/). Connect the three connectors your top-3 workflows need.

For each of the top 3 workflows:

1. **Use the closest matching skill from the 15 ship-with-it skills.** Eight of those 15 will cover most of what the agency was doing. Start with the default skill text.
2. **Customize the skill text** based on the agency's prompt/template export. The agency learned things about your business (acceptable tone, specific edge cases, customer naming conventions). Don't lose that. Port the lessons into the skill instructions.
3. **Run the [Claude Skill Linter](/tools/claude-skill-linter/)** on every skill before turning it on.
4. **Set the daily cost cap and alerts.** $20/day cap, alert at $10. (See [the cost-controls post](/blog/blog-ai-agent-cost-controls-smb/) for the full safeguard stack.)
5. **Run the skill on a small test batch.** One overdue invoice, not all 40. Read the draft. Adjust until it sounds right.
6. **Promote to draft-only auto-run.** The skill drafts, you approve.

By end of day 14, three workflows are running drafted-but-not-sent on Claude Pro. The agency is still running its versions. Both are emitting the same drafts. You're comparing.

## Days 15-21. Rebuild the next batch

Same process for the next 3-5 workflows. By day 21, your full top-5-to-top-8 workflow set is shadow-running on Claude alongside the agency.

A few things you may need to build yourself if the ship-with-it skills don't cover them:

- **Custom report generation.** The agency may have been emailing you a weekly business-health report. Build this as a Claude skill that pulls QuickBooks data and emails a templated summary to you.
- **Customer segmentation logic.** If the agency's "AI customer scoring" was important, replicate it as a Claude skill with the scoring rules made explicit. You'll discover the rules were probably simpler than the agency made them sound.
- **Multi-step approval workflows.** Slack's workflow builder + Claude can usually cover what the agency was charging extra for.

## Days 22-26. Parallel run and sanity check

Both systems run in parallel for a week. Every workflow that the agency runs, Claude also runs. You compare the drafts side by side.

Look for:

- **Output quality.** Are the drafts comparable? Where does Claude's draft fall short? (Usually you can fix this with one more line of skill instructions.)
- **Coverage.** Does Claude catch the same edge cases the agency was catching? (If not, you've found something the agency was actually earning their fee on, which is a useful data point.)
- **Speed.** Is Claude faster or slower? For most SMB workloads, Claude through Cowork is fast enough; the agency may have been adding latency through their own backend.

If Claude is consistently as good or better, you're ready to cut over. If Claude is meaningfully behind on 2+ workflows, decide: invest another week tuning the skill, accept the gap and rebuild only the rest, or in rare cases keep the agency for the specific workflows where they're earning their fee.

## Days 27-28. Cutover

This is the actual transition. Two days.

**Day 27 morning.** Promote all Claude skills from draft-only to auto-send (or whatever the live mode is). The agency's versions also still run.

**Day 27 afternoon.** Watch both. Spot-check 5-10 actions across the workflows. Confirm Claude is doing the right thing.

**Day 27 evening.** Pause the agency's automations (most agency platforms have a "pause" button on each workflow). Don't cancel yet, just pause.

**Day 28.** Spend the day with Claude running solo. Monitor more carefully than usual. End of day, if everything looks right, send the agency the formal termination notice (in writing, as your contract specifies).

## Days 29-30. Decommission and final invoice review

Disconnect the agency's connectors from your QuickBooks, Gmail, Stripe, etc. Revoke their OAuth permissions in each of those platforms (don't trust the agency to revoke their own; do it from your side).

Confirm the agency has returned all your data (the bulk export from days 1-3 plus any incremental data they've added since).

Review their final invoice. Most agencies will pro-rate the final month. Some will try to charge a full month even if you cut over on day 14. Push back on this; the contract usually allows pro-rating.

Save the data export and the prompts/skills to long-term storage. You may want them as historical reference.

## Cost comparison, real numbers

For a 30-person service business with the workflow set described above:

**Before:**
- AI agency: $3,500/month
- Annual: $42,000

**After:**
- Claude Pro: $20/month
- 1 extra Google Workspace mailbox for Claude's outbound (optional): $7/month
- A couple of Zapier zaps for the retired-workflow replacements: ~$20/month
- Annual: ~$564

**Savings:** ~$41,400 per year. Net of the 40-60 hours of migration work.

That's not a hypothetical. For typical SMBs with typical AI agency contracts in 2026, those are the actual numbers. The math is rarely close.

## Where the agency was actually earning their fee

To be fair: not every agency is a wrapper. Some of what they were doing is genuinely worth paying for. Be honest with yourself about what stays useful even after the cutover.

Common things that genuinely earn the fee:

- **Custom integrations that bridge platforms with no native connector.** If the agency built a Salesforce-to-Claude bidirectional sync that doesn't ship out of the box, that's real engineering and worth replacing carefully or paying for.
- **Compliance work in regulated industries.** Healthcare, financial, legal SMBs often have audit and retention requirements that an agency can demonstrate compliance with. Don't drop this layer without a replacement plan.
- **Ongoing iteration based on watching the workflows in production.** If the agency was tuning prompts every two weeks based on what was working and what wasn't, you'll need to take that on yourself or hire a part-time prompt-engineer / ops person to do the equivalent.

For most 25-50 person SMBs in non-regulated industries, none of the above applies, and the migration is straightforward. For some, it isn't. Be honest before you walk.

## The audit tools that support the migration

- **[AI Vendor Cost Reverse-Calculator](/tools/ai-vendor-cost-reverse-calculator/).** Run on the agency's quote before you decide to leave. Confirms the math.
- **[Claude Skill Linter](/tools/claude-skill-linter/).** Run on every skill you write for the migration.
- **[Invoice Generator](/tools/invoice-generator/).** If the agency was generating invoice templates, this tool replaces that piece for free.
- **[DNS Email Audit](/tools/dns-email-audit/).** Confirm your sending domain is configured cleanly before Claude takes over invoice-chase. Common gotcha: agency had their own sending domain; you need to switch to yours.
- **[FBI Fraud Reflex Card for SMBs](/tools/fbi-fraud-reflex-card/).** Useful during the cutover for spotting common social-engineering attempts (fake "you need to renew your AI subscription" emails are surprisingly common after a vendor switch).

## What happens if it doesn't work

Real talk: a small percentage of migrations don't go cleanly. You discover the agency was doing more than you thought. The replacement skills don't quite match. Customer drafts come out too robotic.

Options if you find yourself there:

1. **Hire one good prompt engineer for two weeks.** $3,000-$5,000, way cheaper than a year of agency fees, and you own the result.
2. **Switch to a transparent wrapper.** A vendor that says "built on Claude" on the homepage and charges $200-$500/month for the workflow layer. Usually a better fit than the original agency.
3. **Re-hire the agency for just the workflows that don't work.** Lower contract value, narrower scope. They may take it, they may not. If not, see option 1.

You're not locked into "leave or stay." There's a middle path that gets you 80% of the savings with 90% of the risk mitigation. Take it if you need it.

## The deeper version

The complete framework for evaluating, leaving, and replacing vendor relationships is in [The $20 Dollar Agency](https://www.amazon.com/dp/B0FB17VG3D) (Digital Empire series, $9.99 on Kindle). This migration is one chapter. The book covers the broader pattern of when to keep a vendor, when to leave, and when to bring it in-house.

## Related reading

- [Part 1: The 30-minute pre-agency self-audit](/blog/blog-pre-agency-ai-self-audit-part-1/), what to do before you sign in the first place.
- [Claude for Small Business walkthrough](/blog/blog-claude-for-small-business-walkthrough/), the destination platform.
- [Connector permission cheat sheet](/blog/blog-connector-permissions-claude-smb/), the security baseline on the way over.
- [Invoice-chase skill deep dive](/blog/blog-invoice-chase-skill-deep-dive/), the most-rebuilt workflow in any migration.
- [Before you pay an agency $3,500/month for proprietary AI](/blog/blog-spot-ai-vendor-markup/), the original markup analysis.

## Fact-check notes and sources

- Claude Pro pricing ($20/month) per [anthropic.com](https://www.anthropic.com).
- Google Workspace Business Starter ($7/user/month annual) per [workspace.google.com/pricing](https://workspace.google.com/pricing).
- The 15 skills + 12 connectors in Claude for Small Business per Inc.com's [coverage of the May 13, 2026 announcement](https://www.inc.com/ben-sherry/anthropics-newest-claude-feature-is-here-to-help-small-business-owners-with-their-pain-points/91343926).
- Data portability obligations on AI vendors aren't yet codified in U.S. federal law; rely on your specific contract's data-portability clause, with backstop from state laws like [CCPA](https://oag.ca.gov/privacy/ccpa) (California) and emerging equivalents.
- The 40-60 hour migration estimate is from informal SMB owner interviews; YMMV widely depending on workflow count and integration complexity.

*This post is informational, not legal, financial, or contract-termination advice. Mentions of Anthropic, OpenAI, Zapier, and other third-party services are nominative fair use. Before terminating a vendor contract, review your specific contract language with counsel. No affiliation with any named platform is implied.*


---

Canonical HTML: https://jwatte.com/blog/blog-post-agency-ai-migration-part-2/
RSS: https://jwatte.com/feed.xml
JSON Feed: https://jwatte.com/feed.json
Hero image: https://jwatte.com/images/blog-post-agency-ai-migration-part-2.webp