Getting Started with the SageMaker Migration Plan Skill
A step-by-step guide for BDs and SAs on using the SageMaker Migration Plan skill in Amazon Quick to generate customer-ready migration plans with financial metrics in minutes, featuring three real-world examples (SAS, Databricks, and C3 AI).
Creating a high-quality, research-backed migration plan for every customer engagement takes time. Between researching the source platform architecture, mapping services to AWS equivalents, building cost models from public pricing, and formatting deliverables for different audiences — a single migration plan can take several days of BD/SA time.
The SageMaker Migration Plan skill solves this by automating the entire workflow inside Amazon Quick. In minutes, it produces three customer-ready deliverables — each grounded in AWS public pricing and tailored to the specific source platform.
Who Is This For?
• Solutions Architects preparing for customer migration workshops
• Business Development managers sizing migration opportunities
• Technical Account Managers building TCO comparisons
• Anyone helping customers move ML workloads to SageMaker AI
What Platforms Does It Cover?
The skill works with any source ML platform:
• SAS (VDMML, Viya, Enterprise Miner)
• Databricks (ML Runtime, MLflow)
• C3 AI Platform
• Iguazio / MLRun
• Kubeflow
• On-premises custom platforms
• Any ML platform the customer is migrating from
What the Skill Produces
Every run generates three deliverables from a single set of inputs:
| Deliverable | Format | Audience | Use Case |
|---|---|---|---|
| Quick Reference | Markdown (.md) | Internal team | Quick review, Slack sharing, internal notes |
| Detailed Plan | Word (.docx) | Customer | Formal delivery, attached to opportunity |
| Executive Deck | PowerPoint (.pptx) | Exec sponsor | Leadership reviews, kickoff presentations |
Sections Included in Every Plan
• Executive Summary
• Current State Assessment (workload inventory, team, pain points)
• Target Architecture (service mapping table + security/compliance)
• Migration Phases (4-phase: Planning → Preparation → Go-Live → Stabilization)
• Risk Assessment (risk matrix with mitigations)
• Financial Metrics (monthly + annual costs from AWS public pricing, ROI comparison)
• Success Criteria (measurable outcomes)
• Next Steps & Action Items
Installation (2 Minutes)
1. Download the skill file from github:
https://github.com/aws-samples/genai-ml-platform-examples/blob/main/migration/skills/sagemaker-migration-plan_SKILL.md
2. Create the skills folder (if it doesn't exist):
mkdir -p ~/.quickwork/profiles/<profile>/skills/sagemaker-migration-plan/
3. Copy the file into the folder:
cp sagemaker-migration-plan_SKILL.md ~/.quickwork/profiles/<profile>/skills/sagemaker-migration-plan/SKILL.md
4. Restart Amazon Quick
5. Verify by typing: "create migration plan" — the skill should activate
How to Use It
Trigger Phrase
Say "create migration plan" in any Amazon Quick conversation. The skill will ask for inputs or you can provide them all upfront.
Required Inputs
| Input | What to Provide | Example |
|---|---|---|
| Customer Name | Company name | SWIFT, Seagate, Truist |
| Source Platform | Current ML platform | SAS, Databricks, C3 AI |
| Target Services | AWS target services | SageMaker AI, Unified Studio |
| Workload Input | Description or file path | See minimum requirements below |
| Team Size | Number of users | 200 data scientists + 2 ML engineers |
| Timeline | Target duration | 12 weeks, Q3 2026 go-live |
5 Minimum Data Points for Financial Metrics
The skill will ask for these if missing. Gather them during discovery calls:
| # | Data Point | Why Needed | Discovery Question |
|---|---|---|---|
| 1 | Number of models | Drives compute volume | How many ML models do you run? |
| 2 | Data volume per job | Instance size + duration | How much data does each model process? |
| 3 | Retraining frequency | Monthly cost multiplier | How often do you retrain models? |
| 4 | Inference pattern | 24/7 vs. burst cost | Real-time predictions or batch scoring? |
| 5 | ML frameworks | CPU vs. GPU selection | What frameworks? (XGBoost, PyTorch, etc.) |
Example 1: SAS Migration (Financial Services)
Customer Profile
• Industry: Financial services (regulated, central bank governance)
• Platform: SAS on-premises
• Scale: 400 models, 10 TB data per training job, 4 PB monthly throughput
• Team: 200 data scientists + analysts
• Inference: Batch scoring (daily)
• Pain points: SAS licensing cost, no GPUs, innovation friction
Prompt Used
Create migration plan for: 400 models on SAS VDMML, 10 TB per training job, retrained monthly, batch scoring daily, 200 users, strict financial regulatory compliance (SOC, ISO, PCI).
Financial Output
| Metric | Value |
|---|---|
| SageMaker Annual Cost | ~$300,680 /year (with Managed Spot) |
| Current SAS VDMML Cost | ~$4,250,000 /year (license + infra + people) |
| Customer Savings | ~$3,950,000 /year (93% reduction) |
| AWS ARR Opportunity | ~$300,000 /year |
| Year-1 Total (ARR + ProServe) | $900K – $1.5M |
Key Selling Point
The customer processes 4 PB monthly on fixed SAS infrastructure. SageMaker elastic compute lets them burst 400 training jobs in parallel, pay only for the ~10 hrs each runs, then zero cost for the remaining ~27 idle days/month.
Example 2: Databricks Migration (Tech/SaaS)
Customer Profile
• Industry: Technology / SaaS
• Platform: Databricks ML Runtime + MLflow
• Scale: 50 models, 500 GB per job, retrained weekly
• Team: 30 data scientists, 5 ML engineers
• Inference: Real-time endpoints (low-latency predictions)
• Frameworks: PyTorch, Spark MLlib
• Pain points: Databricks costs scaling with cluster hours, need GPU access
Prompt Used
Create migration plan for: 50 models on Databricks, 500 GB per job, retrained weekly, real-time inference, PyTorch + Spark, 35 users, timeline 10 weeks.
Why This Example Works
• Demonstrates GPU acceleration value — PyTorch models run faster on ml.p3/g5
• Real-time endpoints show the 24/7 cost driver (biggest line item)
• Weekly retraining = higher compute cost than monthly
• Mid-market deal size: ~$50K–$150K ARR typical
Key Selling Point
Databricks clusters bill by the hour even during idle periods between jobs. SageMaker Training Jobs spin up, train, and terminate — no idle cluster costs. Plus, SageMaker multi-model endpoints serve 50 models from a single endpoint, reducing inference cost by 5–10x vs. one-cluster-per-model.
Example 3: C3 AI Platform Migration (Manufacturing)
Customer Profile
• Industry: Manufacturing / Financial Cooperative
• Platform: C3 AI (MlPipes, C3 AI Studio, Model Registry)
• Scale: 17 VMs (high-memory), 5 use cases, training 3 days/month
• Team: 10 data scientists, 2 ML engineers
• Inference: Mix (batch + real-time for 2 use cases)
• Frameworks: XGBoost, PyTorch
• Pain points: VMs idle 90% of the time, no GPUs, no org-wide platform
Prompt Used
Create migration plan for: C3 AI platform, 17 VMs idle most of the month, 5 use cases (starting with 2: XGBoost + PyTorch), 10 TB total data, 12 users, strict security compliance, timeline 12 weeks.
Why This Example Works
• Classic "idle infrastructure" narrative — 17 VMs running 24/7 for 3 days of work
• Clear ROI: always-on VMs → pay-per-use training jobs
• Phased approach: start with 2 use cases, expand to remaining 3
• SMB/mid-market opportunity: ~$50K–$100K ARR
Key Selling Point
The customer pays for 17 always-on VMs but only uses them 3 days/month (10% utilization). On SageMaker, they pay for exactly 3 days of compute, then $0. That alone represents ~90% infrastructure cost reduction before even considering the platform licensing savings.
Tips for BDs and SAs
When to Use Each Deliverable
| Deliverable | When | How |
|---|---|---|
| Markdown | Internal prep, Slack sharing, quick reference | Paste in Slack thread, attach to internal docs |
| Word (DOCX) | Customer-facing delivery, opportunity attachment | Email to customer, attach to Salesforce opportunity |
| PowerPoint | Executive sponsor meetings, kickoff presentations | Present in first call, use at architecture review |
Discovery Call Tips
The 5 minimum data points map directly to discovery questions. Work them into your first call:
1. "How many ML models are you running in production today?"
2. "What size datasets do your models process? Gigabytes, terabytes?"
3. "How frequently do you retrain? Monthly? Weekly? On-demand?"
4. "Are your models serving real-time predictions or doing batch scoring?"
5. "What ML frameworks does your team use? XGBoost, PyTorch, TensorFlow?"
Refining Estimates
The first pass gives you a 130% estimate. To tighten it:
• Ask the customer for their current platform annual spend (for ROI comparison)
• Get actual training job durations (if they track them)
• Confirm the inference SLA requirements (latency, availability)
• Run the skill again with refined inputs — it takes minutes, not hours
FAQ
Q: What if I don't have all 5 minimum data points?
The skill will ask you for the missing ones before proceeding. If you truly can't get them, provide estimates (e.g., "approximately 50 models, probably monthly retraining") — the skill will produce a range-based estimate with appropriate caveats.
Q: How accurate are the cost estimates?
Typically within 130% of actual costs. The skill uses current AWS public pricing and reasonable assumptions about instance selection. For a more precise estimate, refine with actual training durations and data volumes from the customer.
Q: Can I customize the plan sections?
The plan includes all standard sections automatically. After generation, you can edit the Markdown or DOCX directly to add customer-specific content, remove sections, or adjust language.
Q: Can I run it for non-SageMaker targets?
Currently optimized for SageMaker AI as the target. The research step searches for SageMaker-specific pricing and documentation. However, the plan structure works for any AWS ML target — just specify the target services accurately.
Q: What if the customer uses a platform I haven't heard of?
The skill researches the source platform during execution. It will search AWS docs, blog posts, and the platform's own documentation to build the service mapping. Even for niche platforms, it will produce a reasonable mapping.
Resources & Links
SharePoint
• Skill file: https://github.com/aws-samples/genai-ml-platform-examples/blob/main/migration/skills/sagemaker-migration-plan_SKILL.md
• User guide: https://github.com/aws-samples/genai-ml-platform-examples/blob/main/migration/skills/sagemaker_migration_skill_guide.md
AWS Pricing References
• SageMaker AI Pricing: https://aws.amazon.com/sagemaker/ai/pricing/
• S3 Pricing: https://aws.amazon.com/s3/pricing/
• ML Savings Plans: https://aws.amazon.com/savingsplans/ml-pricing/
• AWS Pricing Calculator: https://calculator.aws/
Related Resources
• AWS Prescriptive Guidance — Migrating SAS to AWS: https://docs.aws.amazon.com/prescriptive-guidance/latest/migration-sas-viya/
• Migrating Enterprise ML Workloads to AWS (blog): https://aws.amazon.com/blogs/industries/migrating-enterprise-ml-workloads-from-databricks-to-aws-for-large-scale-ml/
• SageMaker Managed Spot Training: https://docs.aws.amazon.com/sagemaker/latest/dg/model-managed-spot-training.html
Feedback & Improvements
To request enhancements or report issues with the skill, post a comment on this article.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article