Advanced Prompting: Ask What the AI Actually Read
A practical AI prompt for construction teams, with a way to check the answer before acting on it.
Before I ask AI to analyze a spreadsheet, I ask one boring question: What did you actually read?
A polished summary can hide a missing worksheet, misunderstood PDF table, or a cost field confused with revenue.
I’ve been thinking about how construction folks can get more out of the information they already have. I don’t want AI to replace the people who know the work. I want it to help us organize the records, ask better questions, and spend less time rebuilding context.
Start with one real question
The best starting point isn’t every file the company owns. I’d pick one specific decision or process and a set of records the team understands. We can expand once we know what works.
Here’s what I’d bring to the first review: A small approved set of files with known row counts, dates, and column definitions. Confirm the workspace can read them and verify calculations.
I’d also make sure the team has approved the AI system for this information. Bids, costs, contracts, and internal procedures can be sensitive. It matters where those records go and who can access them.
The prompt I’d use
This is the prompt I’d start with in a workspace that can read the approved documents and perform any calculations the task requires. It’s a starting point, not a promise that every chat system can read every file or verify every number.
Before analyzing these files, inventory what you can access: files, sheets or sections, record counts, date ranges, key fields, and unreadable or excluded content. Check for duplicate rows, missing IDs, inconsistent units, and unclear definitions. Show sample rows with source locations. Do not infer unseen content or treat blanks as zero.
State whether this workspace can run and check calculations, and what method it will use. Ask questions that could materially change the analysis, then wait. Do not produce business conclusions until the input inventory is verified.
What a useful answer should look like
First, I want an inventory. What did the AI actually read? Which dates, versions, and fields did it use? If something was unreadable or missing, I want that called out rather than quietly ignored.
Second, I want the working behind any conclusion. If the answer depends on a percentage, show the numerator, denominator, formula, and source rows. If it depends on a clause or a note, show the document and section. A confident sentence without evidence is not a finding.
Third, I want a short list of questions for the people who know the work. I’d rather bring three well-supported questions into a meeting than twenty recommendations that nobody can explain.
A simple example
Imagine a workbook with Original Estimate and Revised Budget tabs. If AI reads only the second, it could give a convincing answer to the wrong question.
This is an illustrative example, not an AB-09 customer result. I’d use it to teach the distinction between something the records establish and something we still need to investigate.
Bring the answer back to the team
I would compare its inventory to the originals and stop if record counts or field meanings do not reconcile.
I’d also ask somebody to check a few source records. Did the AI understand the definitions? Did it count the same item twice? Was there a more recent version? That little bit of verification can keep a plausible-looking answer from becoming a bad business decision.
Once the team agrees on the findings, pick one practical next step. Maybe that’s a clearer handoff, a better checklist, or a question to ask before the next bid. The goal isn’t a flashy report. It’s a change that helps people do the work.
Why private AI matters for this
This is one reason we built the AB-09 Private Intelligence Appliance at AutomateBetter. It’s a physical AI appliance installed in your office. Your team works with approved company information through a browser, and the AI processing happens on your own network rather than sending prompts and files to an outside AI model.
We work with the company on access, approved file formats, and specialized workflows. We coordinate with the existing IT provider and provide installation, training, and agreed ongoing support. We’re based right here in Dover, NH, serving construction companies across New England.
Local processing doesn’t magically make an answer correct or replace security controls. People still need to verify the records and make the decisions. But it gives construction folks a way to work with company information that they may not want to send to an outside provider.
If you’d like to see how this could work for your team, take a look at the AB-09 and book a demo. Tell us what problem you’d like to solve. There’s no need to send confidential files with your request.