1000+
Lines in a typical price estimation spreadsheet
Enormous, formula-filled sheets are the industry norm, not an edge case.
Case Study
Applying agentic AI automation in document management for the construction industry
An undisclosed construction documentation platform asked Vstorm to automate cost and time estimation with an agentic workflow that works on the Excel files the industry actually runs on.
The outcome
Thousand-line estimates, updated from the documents that land
Cost estimation is the costliest problem in construction. A single mistake can run to hundreds of thousands of dollars, and estimation time is the main bottleneck on winning new work. Before this engagement, the platform extracted the information but a person still had to type it into the spreadsheet.
Figures describe typical documents in Canadian construction, as supplied by the client.
600+
Pages in a typical Canadian documentation set
Drawings, tables and charts included, deeply interconnected.
Seconds
To fill the estimation spreadsheet with the agent
The work that used to be manual entry across thousands of rows.
About the client
A construction documentation management platform, not named here. It supports construction companies in staying compliant and preparing offers for clients — more accurate price estimations, more sophisticated proposals, and tighter project budgets.
The product focuses on document intelligence, easing the burden of preparing documentation for clients and vendors so construction companies can focus on what they are best at: construction.
Vstorm's impact
Vstorm's impact, the TL;DR
- A spreadsheet parser designed to fit the existing, agent-native system, so Excel becomes something agents can read, edit and hand back
- Sheets modified from the documents uploaded to a project, without losing content, formulas or formatting
- Llamaparse reserved for documents that need it, cutting extraction cost on the plain-text majority
- Benchmarking environment reviewed and tested to verify the changes actually improved results
The challenge
The skyscraper of documents
Construction documentation is extremely complicated, because it has to connect a legal framework and building reality to the architect’s vision. The documents themselves are dense with drawings, tables, charts and other visual elements.
They are also deeply interconnected. Some are quotes from contractors, some prices of components, some time and material estimates. The standard workflow takes all of that and connects it into an overall estimate of time to deliver and cost to build. This is usually done in enormous, formula-filled Excel sheets, very often running past a thousand lines.
Providing a tool that let those files be filled automatically was the core of our engagement.
How we delivered
TriStorm, applied to one bottleneck
The client already had an agent ecosystem. Our scope was the part of it that could not yet touch a spreadsheet.
Strategic alignment and planning
Our transformation consultancy team worked with the founders to identify the technologies best suited to the problems and constraints the product faced.
- Founder workshops
- Technology fit
- Scoped bottleneck
Proof of Value
Having settled the approach, the team moved forward and tested the best-performing technologies against the actual challenge rather than a simplified version of it.
- Parser prototype
- Benchmark environment
- Verified results
Process augmentation
The process was integrated into the existing tool ecosystem, unclogging one of the biggest bottlenecks in the app and improving the overall user experience.
- Agent toolset in production
- Formulas preserved
- Lower extraction cost
Our work
Reducing the cost of data extraction
Initially the company used Llamaparse for content extraction, handing the extracted knowledge to the user so they could copy and paste new values into the cost estimation document. That raised two problems.
Llamaparse costs stack up, particularly on complex documents full of images and tables, which is exactly what construction documentation is. And running every incoming document through it was unnecessary spend, because the majority were plain text and needed nothing like that degree of sophistication.
Automated estimation update
Before working with Vstorm, the tool extracted all the information but the user still updated the spreadsheet by hand. On high-rise or highway projects, where a single sheet can carry several thousand entries, that is a serious amount of labour.
Our engineering team built an agent toolset able to modify those cost estimation spreadsheets on the fly, without loss, confusion or hallucination at any step.
The flow was woven into the company’s environment. The existing agent ecosystem worked on HTML-transformed documents, so the system converts Excel files into HTML-based objects, modifies them, then turns them back into whichever format is needed: XLSX, XML, JSON or Markdown, depending on the task.
Formulas and formatting survive the round trip. When a user uploads a new document, an offer from a contractor for instance, the cost and time estimate updates itself.
Document uploaded Contractor offer, spec, drawing set
↓
Extraction, tiered by need Llamaparse only where complexity warrants it
↓
Excel to agent-native object Converted to HTML the ecosystem already speaks
↓
Agent edits the estimate Formulas and formatting preserved
↓
Estimate updated in place Back to spreadsheet, ready to send
Vstorm × undisclosed construction documentation platform — spreadsheet toolset inside an existing agent ecosystem
Results
Vstorm’s impact
After implementation, a significant part of the manual and repetitive work the end user had to do was cut. The cost and time estimation document now updates itself according to the supporting documents uploaded.
The agent can then be asked to adjust the document further, changing contractors depending on scope of work, price or any other variable.
Alongside the new capability, maintenance costs fell, through a smarter approach to when Llamaparse is worth calling.
Summary
Cheaper to run, and easier to live with
The system is now more convenient for end users and less costly to maintain and run. The project was a good opportunity to apply what we know about existing agents and to balance cost against performance. It was also a substantial step in turning a largely manual, labour-based tool into a modern agentic system able to take on what comes next.
The system runs on Pydantic AI as the agentic framework. Depending on the task, the company uses Claude, Gemini or OpenAI models to process the workflows for the best result.
Work with us
Have a document workflow that will not automate?
Meet directly with our founders and PhD AI engineers. We will walk through real implementations from 30+ agentic projects and the practical steps to integrate them into your workflows.