Engineering and consulting firms are adopting AI faster than ever. According to the 2025 Deltek Clarity A&E Study, 53% of AEC firms now use AI tools, up from 38% just a year ago. The top use cases? Proposal development and business development.
But most firms trying AI for proposals are running into the same problem: the results aren't that good.
The resumes sound generic. The project descriptions miss key details. The "right people" the tool recommends aren't actually right. And proposal managers end up spending almost as much time fixing the AI output as they would have spent doing it manually.
The issue isn't the AI, especially as the models improve. But no matter how good the model gets, the issue will always be the data behind it.
The document problem
Most professional services firms store their proposal data in documents. Word files, PDFs, PowerPoint decks, SharePoint folders, DAMs. Resumes live in one location, project descriptions in another, credentials somewhere in an HR system.
This is how the industry has worked for decades. And for human-driven proposal processes, it's functional. Messy, slow, and frustrating, but functional.
For AI, it's a disaster.
When an AI tool tries to work with a Word document or PDF, it has to figure out what the document is, what each section means, where one piece of information ends and another begins, and how different pieces relate to each other. A resume in a Word file doesn't have a field labeled "Years of Experience" or "Key Certifications." That information is buried in paragraphs, formatted differently across every document, and often outdated.
Structured data is different. In a structured system, every piece of information has a clear label, a defined relationship to other data, and a consistent format. "Project Name," "Client," "Start Date," "Role," "Certifications." The AI doesn't have to guess, it is able to find it quickly and most importantly, clearly. For example, what happens when AI is asked to fill out a form, but it is missing content? Without structured data and controls, it hallucinates.
The research is clear
This isn't theoretical. Multiple studies have now shown that AI performs dramatically better with structured data than with documents.
The J&J experiment
One of the most compelling examples comes from XBRL US, the organization behind the global structured data standard for financial reporting. They ran a simple test. They gave Claude, Anthropic's AI model, access to Johnson & Johnson's 2024 financials and asked for a revenue breakdown by business segment. Then they ran the same prompt twice: once without structured data, and once with structured XBRL data connected through an MCP server.
Without structured data, Claude returned the headline numbers. Total 2024 sales of $88.8 billion, split across two broad segments: Innovative Medicine at $57 billion and MedTech at $31.9 billion. It named some subsegments but could only find partial quarterly figures for a few of them, and the geographic view stopped at "US" and "International." To fill the gaps, it pulled from third-party sources like Motley Fool, Yahoo Finance, and Wikipedia.
With structured data, the same model returned a completely different level of detail. It reported revenue for all six subsegments inside Innovative Medicine, then went a layer deeper and gave full-year figures for individual drugs, including every product in the Oncology line. It did the same for all four MedTech subsegments down to the product category. The geographic breakdown expanded from two regions to four: US, Europe, Asia Pacific/Africa, and Western Hemisphere.
Same AI model. Same company. The difference in depth came entirely from the data underneath it.
The most important difference was the source. With XBRL, every figure came straight from J&J's own 10-K filing. Without it, the AI stitched together numbers from outside websites that may be normalized or restated, and can differ from what the company actually reported. Structured data did not just make the answer more detailed. It made it traceable to the most reliable source there is.
The PDF confusion problem
XBRL International ran a separate set of experiments comparing AI analysis of structured data versus PDF documents. When working with PDFs, the AI consistently pulled incorrect figures, got confused by nearby content on the page, and incorporated irrelevant information from adjacent sections.
With structured data, those errors disappeared. The AI could locate exactly the right information because every data point was clearly labeled and contextualized.
The 72% vs. 17% gap
A research team from Columbia University and Georgia Tech built a benchmark called FinTagging to test how well AI models handle structured versus unstructured financial information. They tested ten state-of-the-art AI models and found that while the models could locate and extract information from documents about 72% of the time, they could only correctly classify and interpret that information 17% of the time without structured data guiding them.
That's a massive gap. And it explains why AI tools that rely on parsing through your existing documents produce mediocre results.
What this means for proposal teams
Now, to bring this back to our beloved proposals.
Your proposal data, the resumes, project descriptions, credentials, and expertise that win bids, almost certainly lives in documents. SharePoint folders full of Word files. Old proposal PDFs. Maybe a DAM that stores images but doesn't structure the underlying data. Maybe an internal database that hasn't been updated in years.
When an AI tool tries to find the right people for a bid, it's scanning through thousands of these files trying to piece together who has the right experience. When it tries to tailor a resume, it's parsing unstructured text and guessing at what's relevant. When it tries to match your team's credentials to an RFP's requirements, it's doing pattern matching across inconsistent (and often outdated) documents.
This is why the AI output feels generic. It's not because the AI is bad, it's because the data is unstructured, inconsistent, and hard to interpret, even for a very capable model.
Autodesk's Head of Generative AI for AEC, Racel Amour, put it well: "Organizing your data with clear standards, such as consistent naming conventions and structured formats, will make it easier for AI to deliver value."
Will AI get better at reading documents?
Yes. AI models are improving at parsing PDFs, understanding table layouts, and handling longer documents. Vision models can now read scanned pages. Retrieval systems are getting smarter at finding relevant sections.
But the fundamental advantage of structured data isn't going away.
When data is already labeled and organized, you eliminate entire categories of potential error. Often fatal errors if you’re submitting made up credentials in a proposal 😳. The AI doesn't have to figure out what a number means, where it came from, or whether the surrounding text is relevant. It just retrieves what it needs.
As XBRL International wrote in a recent analysis: the idea that AI can just read documents and produce perfect insights is "dangerously wrong."
The real trend is that AI will increasingly be used to turn unstructured data into structured data as a preprocessing step. The best enterprise AI systems already work this way. They extract, structure, and label information first, then analyze it. The structuring step is what makes everything downstream work. Brief aside - it’s why we built Flowcase’s Data AI Assistant.
Structured data wins before AI even enters the picture
Here's the thing that often gets lost in the AI conversation: structured data doesn't just help AI. It helps your proposal team right now, with or without AI tools.
When your resumes, project descriptions, and credentials are structured in a central system, your proposal managers can search for the right expertise across the firm in seconds. They can tailor a resume for a specific RFP without creating a duplicate document. They can pull together a project team based on actual certifications and project history rather than tribal knowledge about who's worked on what.
The 2025 Deltek Clarity Study found that AEC firms are submitting fewer proposals but winning more of them. Proposal volume dropped 38% while the value of awarded work grew 52%, and win rates hit 50%. Firms are getting more selective and strategic about what they pursue. That kind of discipline requires knowing exactly what expertise you have and being able to deploy it quickly.
What Flowcase does about this
This is the problem Flowcase was built to solve.
Flowcase takes the people and project experience that lives in your firm and structures it. Every employee has a structured profile with their skills, certifications, project history, and credentials in labeled, searchable fields. Every project has a structured description with defined roles, services, industries, and outcomes.
That data lives in one place, stays current because employees update their own profiles, and can be instantly searched, filtered, and tailored for any proposal.
When your team needs to find every structural engineer who's worked on healthcare projects in the Southeast, that's a search that takes seconds. When you need to tailor a resume for a specific RFP, you adjust the structured record without breaking the master version. When you're ready to connect AI tools to your proposal workflow, the data is already clean, labeled, and ready.
Firms like Sweco, Ramboll, and WSB Engineers already use Flowcase to manage their expertise this way. As one Tender Specialist at Sweco put it: "Flowcase saves us a lot of time in our tender process, both for the bid managers and tender coordinators and for the employees who need to maintain their CVs and references."
The bottom line
AI is getting better every day. But the firms that will get the most value from AI in proposals aren't the ones buying the fanciest writing tools. They're the ones structuring their data now.
Clean, structured people and project data is the foundation everything else builds on. Better proposals today. Better AI results tomorrow.
If you want to see what structured proposal data looks like in practice, check out Flowcase's interactive demo.
