We use a three-layer validation system, targeting 99%+ production accuracy:

• Grounding: The AI only extracts data from your documents—it is incapable of generating or guessing information from the internet.
• Confidence Scoring: If the AI scores <90% confident on any item, it flags the record for human review by YOUR staff through a simple approval queue.
• ERP Cross-Checks: We validate extracted data against your existing vendor master lists, project codes, and approval thresholds for hard-coded correctness.

When rare errors occur (usually edge cases like damaged scans), they route to your team within seconds for quick correction, maintaining high automation efficiency.

The Protocol is designed for low friction. The total time commitment is ~15 hours spread over 8–10 weeks:

• Phase 1 (Feasibility): 30 minutes to send us 5-10 sample PDFs. Zero staff time after that.
• Phase 2 (Roadmap): 2 hours from Leadership (kick-off/review) and 3 hours from SMEs (Finance/Project Mgmt) over 3 weeks.
• Phase 3 (Compliance): 1 hour for IT Admin kick-off and 2 hours for policy review/approval.
• Phase 4 (Deployment): 4 hours for User Training and 2 hours for UAT testing with 2-3 end users.

We provide all agendas in advance and work around your project deadlines.

The solution is built on Microsoft Power Platform. If you have on-premise file storage, we use Microsoft's Power Platform Data Gateway—a secure, Microsoft-certified connector that:

• Runs inside your firewall (no inbound ports need to be opened).
• Only sends encrypted data to your M365 tenant for processing.
• Is maintained by Microsoft, not Curam AI.

Limitations: On-premise setups are slightly slower (file retrieval adds 2-5 seconds). During Phase 2, we will assess the infrastructure and recommend the optimal integration architecture.

You own the system. We provide full source code, documentation, and video walkthroughs of maintenance tasks.

Maintenance: Routine tasks (adding vendors, adjusting low-code rules, managing permissions) can be handled by any IT admin familiar with M365. No coding is required.

Enhancements:

• DIY: Your team can build additional workflows using the low-code foundation we deliver.
• Re-engage us: New, complex workflows are priced as a fixed Phase 4 sprint (typically starting from ~$15k).

Tier 2 Support (Optional): We offer an optional Tier 2 retainer (covering AI model retraining, complex troubleshooting) for $500/month (up to 2 hours support) with a 24-hour response SLA.

Our validation layers are designed to flag errors before they reach your ERP. If an error does bypass the system (e.g., a critical field is misread):

• Workflow: The human review process ensures the data is corrected and re-entered quickly.
• Feedback Loop: The incorrect document/data immediately enters a feedback loop to retrain the AI model for that specific edge case.
• Professional Indemnity: We carry Professional Indemnity insurance which covers errors made by the AI implementation. Details of our policy are provided during the contract phase.

Phase 1 is $1,500 and comes with a 100% money-back guarantee based on objective performance:

• Failure Criteria: If we cannot extract your target data with >90% accuracy from your sample set (minimum 10 documents).
• Measurement: You verify the extracted data against your ground truth. Even one field wrong on a document means that document is a failure.
• Refund: If the 90% accuracy is not met, you receive a full $1,500 refund within 5 business days.

We pride ourselves on transparency: 43 of our last 47 Feasibility Sprints passed. The 4 failures were due to poor source material (e.g., 100% handwritten forms) and all received a full refund.

Yes. Our models are pre-trained on engineering-specific formats, not just generic invoices. We can handle documents like:

• Concrete Test Reports (Extracting lot numbers, slump, and compressive strength)
• Beam Schedules (Extracting dimensions, steel grade, and placement notes)
• Drawing Registers (Auto-extracting revision numbers, issue dates, and file paths)

During a live demonstration, we can use redacted samples from previous deployments to show how our models handle these specific documents compared to off-the-shelf solutions.