AI Integration Planning Workbook: Map Systems, Data Flows, and Integration Points Before You Build
Most AI projects fail not at the model level but at the integration level. The AI performs well in isolation and falls apart when it has to connect to your CRM, your aging database, or the five-year-old platform that has no documented API. This workbook gives you a structured process to audit every system, data source, and workflow before a project starts. Work through it before you talk to vendors. When you finish, you will know exactly what you are connecting to, where the risks live, and what to scope and price before the contract is signed.
How to Use This Workbook
This workbook is designed to be completed by the internal project owner — typically an operations lead, IT director, or senior department head — before the AI vendor or integrator is engaged. It takes 3–5 hours to complete thoroughly for most businesses. Work through each section in order; later sections reference answers from earlier ones.
Print it or paste it into a shared doc and fill it out collaboratively with the people who actually operate each system daily. The person who built your CRM configuration and the person who runs reports out of it both have information you need. Department heads who have never opened the back end of their own systems regularly underestimate how messy their data is.
This workbook does not require technical knowledge to fill out most sections. The sections flagged [Technical] require input from an IT lead, developer, or your AI consultant.
Section 1: Problem Definition and Scope
Before mapping systems, write out exactly what you are asking the AI to do. Vague scope produces vague integration requirements.
1a. In one sentence, what decision or task will the AI perform? (Example: 'The AI will draft a first-pass response to incoming customer support emails and route unresolved ones to the right agent.')
1b. What triggers the AI? Is it a user action (a button click, an upload), a scheduled event (every night at midnight), a real-time stream (every new record, every incoming message), or an API call from another system?
1c. What does the AI output? A structured record, a text response, a routed workflow, a prediction score, a document, an API call to another system?
1d. Who or what consumes the output? A human in a dashboard, an automated downstream workflow, a customer-facing product, a report?
1e. What defines success? Set at least two measurable targets (e.g., 'reduces email response time from 4 hours to under 30 minutes; routes correctly 90% of the time').
Document these answers before proceeding. If you cannot write a clear one-sentence answer to 1a, stop here. Unclear problem definition is the single most common root cause of failed AI projects.
Section 2: System Inventory
List every system the AI will need to read from or write to. For each system, fill in this row:
| System Name | Vendor | Purpose | Hosted Where | Year Last Updated | Integration Method | Data Owner |
Common systems to include: CRM (Salesforce, HubSpot, Zoho), ERP (QuickBooks, NetSuite, SAP), HRIS (BambooHR, Rippling, Workday), marketing automation (Marketo, Klaviyo), customer support platform (Zendesk, Freshdesk, Intercom), document storage (SharePoint, Google Drive, Box), databases (MySQL, Postgres, SQL Server, legacy flat files), communication tools (Slack, Teams, email), ecommerce platforms (Shopify, WooCommerce), scheduling systems, industry-specific platforms.
For hosted where: 'Cloud SaaS', 'On-premise server', 'Hosted server we manage', or 'Unknown' (unknown is a flag — find out before proceeding).
For integration method: 'REST API', 'Webhook', 'SFTP/file export', 'Native connector', 'Manual export only', 'None/unknown'. 'None' or 'Manual export only' means this system is a risk item for Section 5.
Rate each system's overall data health 1–5 (1 = known mess, 5 = clean and well-governed). Be honest; this number affects the integration cost estimate.
Section 3: Data Source Mapping
For each data source the AI will use as input, document:
3a. Data source name and which system it lives in.
3b. Data type: structured (rows and columns), semi-structured (JSON, XML, email), or unstructured (free text, PDFs, images, audio).
3c. Volume: current records/size and monthly growth rate.
3d. Freshness requirement: does the AI need real-time data, hourly batch, daily batch, or is historical data sufficient?
3e. Completeness: what percentage of required fields are actually populated? (Pull a sample of 50–100 records and count null or blank fields for the fields the AI needs.)
3f. Consistency: are values formatted consistently? (Date formats, phone number formats, name capitalization, category labels.) If you have 'New York', 'NY', 'New York, NY', and 'ny' representing the same value in the same field, document that.
3g. Sensitivity: does this data contain PII, PHI, financial records, or anything governed by GDPR, CCPA, HIPAA, PCI-DSS, or industry regulations? Annotate each sensitive field.
3h. Historical depth: how many years of usable historical data exist? Many ML models need 2–3 years of history to detect patterns.
Red flags: any data source with completeness below 70%, consistency issues, or sensitivity classification that hasn't been reviewed by legal/compliance goes into Section 5 as a risk item.
Section 4: Data Flow Diagram
Draw or describe the complete data flow the AI will participate in. Include every system, the direction data moves, the frequency, and the transformation that happens at each step.
A simple format if you are not using a diagram tool:
STEP → SYSTEM → TRANSFORMATION → DESTINATION
Example:
1. Inbound email arrives → Email server → AI reads email body and metadata → AI model
2. AI model → generates response draft and extracts intent category → CRM ticket queue
3. CRM ticket queue → agent reviews and approves/edits draft → sent to customer
4. Sent response → CRM closes ticket → reporting database
Document the expected latency at each step. A trigger that needs to process within 5 seconds of a customer action has very different infrastructure requirements than a nightly batch job.
Flag any step where data crosses organizational boundaries: third-party vendor systems, subcontractors, cloud services that may be in different regulatory jurisdictions. These have compliance implications.
[Technical] For each integration point, document: authentication method (OAuth2, API key, service account), rate limits on the source system, and whether you have a sandbox or test environment. Missing test environments add 20–40% to integration timelines.
Section 5: Integration Risk Register
For each risk identified in Sections 2–4, create a row:
| Risk | Affected System | Severity (H/M/L) | Mitigation | Owner | Due Date |
Common risks to check against:
No API available: requires custom extraction, SFTP file drops, screen scraping, or a middleware build. Score severity High — this will extend timelines and cost.
Real-time requirement vs. batch-only source: if the AI needs live data from a system that only exports nightly, architectural workarounds are required.
Data quality failure: fields below 70% completeness or high inconsistency require a data cleaning phase before the AI can produce reliable output.
PII/PHI crossing system boundaries: legal/compliance review required. May require data masking, tokenization, or on-premise model deployment.
Legacy system without documentation: reverse-engineering integrations for undocumented systems is expensive and risky. Budget a discovery phase.
Vendor API instability: check the vendor's API changelog and status page history. APIs with frequent breaking changes or downtime require additional error handling and monitoring investment.
Single point of failure: if the AI depends on a system with no redundancy or SLA, the AI's uptime is capped by that system's uptime.
Review this register with your AI consultant before scoping the project. Every High severity item is a scope and cost risk.
Section 6: API and Connectivity Assessment [Technical]
For each system with an available API:
6a. Document the API type (REST, GraphQL, SOAP, proprietary SDK) and version.
6b. Retrieve and save the API documentation. Note: is it current and complete, or outdated and partial?
6c. Request a sandbox or developer environment from the vendor. If they do not offer one, your team will need to test against production — document this as a risk.
6d. Check the API's rate limits. If the AI will call the API frequently (every few seconds in a high-volume system), confirm the rate limit supports that volume or negotiate an enterprise tier.
6e. Identify authentication requirements: OAuth2 scopes needed, API key rotation policy, IP allowlisting requirements, webhook signature verification.
6f. Verify write permissions: many SaaS systems offer read access via API at all tiers and require enterprise tier for write access. Confirm before scoping.
6g. Confirm data format: what field names, date formats, and encoding does the API use? Mapping mismatches are the most common source of integration bugs.
For systems without APIs: document the available alternatives — file export format and frequency, direct database access (if self-hosted), middleware (Zapier, Make, Workato, MuleSoft) compatibility, or vendor-provided data sync tools. Identify which alternative your integrator will use and what the tradeoffs are.
Section 7: Compliance and Security Checklist
Answer each question. Any 'No' or 'Unknown' that touches regulated data requires resolution before production deployment.
Data Residency
- Do you know where each data source is hosted geographically?
- If operating under GDPR, does all EU personal data remain within approved regions?
- Does your AI vendor/model provider store your data, and if so, where?
Access Controls
- Will AI system credentials be stored securely (secrets manager, not hard-coded)?
- Is access to AI system logs and outputs limited to authorized personnel?
- Is there an audit trail for AI decisions affecting customers or employees?
Data Minimization
- Does the AI actually need PII to perform its function, or can it work with anonymized or tokenized data?
- Have you identified the minimum necessary data fields and removed the rest from the AI's input?
Vendor Agreements
- Does your AI vendor's Terms of Service address data ownership and model training? (Many consumer LLM products use uploaded data for training by default unless you opt out.)
- Is there a DPA (Data Processing Agreement) in place with every vendor that processes personal data on your behalf?
Incident Response
- Do you have a procedure for what happens when the AI produces incorrect or harmful output?
- Is there a human-in-the-loop review step for high-stakes decisions?
- Do you have a way to turn off the AI quickly if a problem is detected?
Section 8: Integration Readiness Score and Go/No-Go Checklist
Rate your readiness in each area from 1–5 (1 = not ready, 5 = fully prepared):
Problem Definition: [ ] The AI task is written in one sentence and has measurable success criteria.
System Inventory: [ ] All systems the AI touches are documented with integration method and data owner identified.
Data Quality: [ ] All input data sources are above 70% completeness; inconsistencies are documented and a cleaning plan exists.
API Availability: [ ] All required systems have accessible APIs or a documented alternative integration path.
Compliance: [ ] All sensitive data flows have been reviewed by legal/compliance; DPAs are in place.
Security: [ ] Credential management, access controls, and audit logging plan are documented.
Test Environment: [ ] A non-production environment exists for integration testing.
Rollback Plan: [ ] You can disable the AI and revert to the prior workflow if a problem is discovered post-launch.
Go/No-Go: Score 5 in all 8 areas before signing a vendor contract. Any area scored 3 or below is a pre-project work item, not a post-launch cleanup item. Teams that defer low scores to 'we'll fix it during the project' consistently miss timelines and budgets by 50–100%.
Bring this completed workbook to your first conversation with an AI consultant or integrator. It reduces the discovery phase from weeks to hours and produces a more accurate project scope on day one.
Frequently Asked Questions
Frequently Asked Questions
Most businesses take 3–5 hours to complete all eight sections thoroughly. Schedule a 2-hour working session with your IT lead and at least one department head who owns the key data source. Expect a follow-up session to chase down answers about systems that no one fully understands. Rushing the workbook is the fastest way to produce an inaccurate project scope.
Sections 1–3 and 5 can be completed by a business-side project owner with no technical background. Sections 4, 6, and parts of 7 require input from an IT lead, developer, or someone who has administrator access to your systems. If you have no technical staff, an AI consultant can complete Sections 4 and 6 during a paid discovery engagement — factor that cost into your project budget.
It is not a project stopper, but it is a cost and timeline item. Your integrator will need to evaluate alternatives: file-based exports (SFTP, scheduled CSV), direct database access if the system is self-hosted, a middleware platform like Make or MuleSoft, or in rare cases, a screen-scraping approach. Each alternative has tradeoffs in reliability, cost, and maintenance burden. Surface this early so it can be scoped and priced accurately rather than discovered mid-project when it is expensive to course-correct.
Review it annually or whenever a major system changes — a CRM migration, a new ERP rollout, a vendor changing their API. AI systems that connect to production data drift over time: APIs change, data schemas shift, volumes grow past original estimates. A stale integration map is one of the most common reasons that an AI system that worked reliably at launch starts producing errors 18 months later with no obvious cause.