What should I collect during the first working week?

The following timings are planning suggestions, contingent on staff availability, permissions and the complexity of the existing systems. A simple internal review may be quick; an enterprise integration or unresolved data permission may take considerably longer.

Days 1–2: write the problem and acceptance brief

• Name one workflow, such as schedule-to-reviewed-quantity-list or approved-estimate-to-proposal.
• State which work is excluded: design, engineering sign-off, automatic bid submission, or any package outside the pilot boundary.
• Identify the accountable estimator, commercial approver, system owner and finance contact.
• Record the actual bottleneck and what a useful result looks like in the receiving tool.
• Define critical omissions or errors that require a stop, regardless of average accuracy.
• Record known document restrictions and who can authorise processing.

Deliverable: a one-page brief with inputs, outputs, owner, exclusions, test evidence and stop conditions. Avoid “make estimating faster” as the entire requirement; state which part and how you will count the full effort.

Days 3–5: build the evidence pack

• Select completed projects representing the intended work, including an awkward revision and a poor scan where relevant.
• Preserve original drawings, schedules, specifications, addenda and supplier quotes with revision identifiers.
• Have an experienced estimator mark the accepted items and known ambiguities; resolve disagreements openly.
• Record preparation, takeoff/extraction, checking, correction and downstream handoff time separately.
• List the live systems, versions, integrations, export formats and owners.
• Obtain written authority for any external processing before uploading real project documents.

Proposed initial sample: 8–12 completed projects for workflow discovery, followed by a separate held-out set for acceptance. This is an editorial starting point, not a statistically sufficient sample or a guarantee of generalisation; increase or stratify it if document types and project risks vary widely.

Document extraction itself has limitations: AWS's technical guidance flags table structure and document quality as practical considerations. Include the conditions that make your work difficult rather than selecting only clean schedules for the demonstration.