Processes & productivity
Video · 7 min read
The short answer
Watch: What does data first mean for AI in construction?
Watch the explanation, then explore the practical implications below.
Data first starts with information people can trust
A construction workflow often depends on several versions of the same story: a drawing, a specification, a scope, a rate and a conversation about a change. Before AI prepares an answer, the business needs a reliable way to identify which records apply and who can approve their use.
Becoming data first means giving information ownership, meaning and a clear role in the workflow. It does not mean replacing every existing system or moving every document into one database. Start with the records needed to complete one recurring process.
Recognise the information in the workflow
Structured information includes fields such as project numbers, quantities, costs and dates. Semi-structured information has identifiable fields but a variable shape, such as an update containing a changing list of activities. Unstructured material includes documents, recordings and photographs.
A quotation can depend on all three. The design needs to preserve the link between an extracted fact and the source that supports it. A figure found in a document should not automatically become an approved rate simply because the system can read it.
Layer 1: map sources and ownership
List where the workflow begins and which systems hold project, commercial and document records. For each source, record its owner, the information it supplies and the version or status that makes a record usable.
Where different teams disagree about the authoritative source, resolve that decision with the responsible people. An integration should not silently choose which commercial record is correct.
Layer 2: connect records with a shared meaning
A technical connection makes information available; it does not establish that two records refer to the same job. Map project identifiers and field meanings across systems. Define how to handle missing values, duplicates and superseded documents.
Make update failures visible and assign someone to respond. The workflow should distinguish information that is current from information whose connection has stopped updating. Decide when stale or incomplete information must block the next action.
Layer 3: give information an appropriate home
Choose storage around the job the information must do. The transcript distinguishes transactional records, managed documents and revisions, media storage and analytical reporting. A business may only need a small subset of these capabilities for its first workflow.
Existing systems can remain the authoritative home of their records. Where information is copied or transformed, document why, who can access it and how changes reach the copy. Evaluate cost, access, recovery and the ability to export alongside technical features.
Layer 4: assemble relevant evidence with permissions
For a quotation, gather the survey notes, customer requirements, applicable specification and approved rates. Attach the sources and check that the person using the workflow is entitled to see them.
Define how the system identifies the relevant project, revision and approval status. When evidence is missing or conflicts, make that clear to the reviewer. Adding unrelated documents increases the material to check without necessarily improving the decision.
Layer 5: give AI a bounded task
Use AI to help interpret supplied evidence or prepare a draft scope. Keep controlled calculations, pricing rules and approved rates under the agreed commercial process. Make the boundary between a draft and an authorised commitment explicit.
A reviewer needs enough evidence, time and authority to challenge the output. Test examples with ambiguous wording and incomplete requirements, so the system can ask for clarification rather than present uncertainty as a settled answer.
Layer 6: route controlled actions and record the result
The workflow can assemble a quotation and send it to the appropriate reviewer. Approval must happen before the external commitment where that is the agreed control. Record the action, the decision and the evidence used.
Design for failures between steps. Identify whether an action completed before retrying it, make exceptions visible and give an owner a route to resolve them. These details determine whether the process remains dependable in daily use.
Keep ownership active after launch
Rates change, permissions change and connections fail. Agree who monitors updates, approves changes and investigates exceptions. Include access removal, backup recovery and data export in the operating design.
Begin with one workflow where people repeatedly search, retype, wait or check. Establish a baseline and measure completed work, review effort and repeat errors. Expand when the information and controls support a result the business can depend on.
Questions for your leadership team
Which records are authoritative?
For a recent job, identify the approved scope, current drawing and applicable rates. Can the team find them without relying on one colleague?
Do systems refer to the same project consistently?
Check how job numbers, names and phases map across commercial, project and document records.
Who can use each source?
Define permissions for the workflow and how those permissions change when people move roles or leave.
What happens when information is stale or incomplete?
Decide when the system must stop, what the reviewer sees and who resolves the missing input.
Can you operate and recover the workflow?
Name the owners of monitoring, rate updates, backups, access removal and data export. Test the agreed recovery route.
Full video transcript
The full video transcript is included below for reference.
Video transcript
Most construction firms think their AI problem is the AI. It isn’t. It’s the data underneath it. Ask one question.
Can your team find the latest drawing, the agreed scope, and the correct rate without asking three different people? If not, no AI tool will fix that. Becoming data first will. And this is where CSM has a head start.
At CSM, we’ve built commercial intelligence systems into over 40 businesses connected with over 3,000 tools over the last five years. So connecting the systems your business already uses isn’t new territory for us. We’ve worked with and tested Supabase, Railway, Microsoft and Google And we weigh them against options like Neon, Snowflake and BigQuery on cost model, data residency and lock in. Here’s how we approach it.
First, know what kind of information you have.
Structured data lives in tidy fields, project numbers, quantities, costs and invoice dates. Semi structured data has shape but the shape shifts. An API update might carry a job number, a status and a changing list of activities.
Unstructured data is the messy part. Specifications, site photos and meeting recordings and one quotation can need all three. Now follow that information through six layers. Layer one, sources.
Procore holds project records. Dynamics holds commercial figures. SharePoint holds documents. Your setup will vary, so start by mapping what exists, who owns it, and which records you trust.
Layer two, integration. An API let systems talk, but a connection doesn’t make the data correct. The same job is project one zero four in finance, Riverside on-site, and phase two in a spreadsheet. Until they share one reference, those records can’t work together.
Then you check for missing fields, duplicates and outdated versions. Layer three, the foundation. Every kind of information gets the right home. Often, records stay where they are.
Reliable transactions belong in a relational database. Files and revisions go to document management. Photos and media go to object storage. Reporting runs from a warehouse and a lake house adds analytics on varied data.
You don’t need all of them. You need the right ones. Layer four, context. For a quotation, pull the survey notes, customer requirements, current specification and approved rates.
Check permissions. Keep every source attached. More information isn’t better. Relevant current information is. Layer five, models.
AI interprets the evidence and drafts the scope, but pricing stays on controlled rules and approved rates and a person checks the draft. Layer six, actions. The system assembles the quotation and routes it to a reviewer. A person approves it before it leaves.
Every action is recorded and failures are visible. That’s the connection between data and useful AI. Dependable information, relevant evidence, and controlled action. And it doesn’t stop at launch.
Who notices when a connection stops updating? What happens when rates change? Can you revoke access, restore a backup, and export your data?
We build those answers in from day one. So don’t start with a company wide transformation. Pick one recurring process where people search, retype, wait, or check. Map it.
Decide what you trust. Connect it properly. Then add AI and measure the difference. That’s how we approach the data layer at CSM.
Want help finding your starting point? Book a call with the CSM team.
Start with one process that matters
Bring a workflow where delays, repeated work or unclear decisions are limiting the business. We can explore the outcome you want, what needs to change and where AI could help.