The problem

A San Francisco Bay Area construction firm came to us with a familiar complaint: leadership could not see the business in real time. The numbers existed, but they lived across a handful of cloud systems, and assembling them into a picture a partner could act on meant a person pulling exports and stitching a spreadsheet together. By the time the spreadsheet was done, the question that prompted it had moved on.

The second problem was on the legal side. Every new project brings NDAs and subcontractor contracts, and a small team was reading every one by hand. The reading was not the hard part. The hard part was remembering which clauses the firm had pushed back on before, and catching the one redline in a forty-page document that actually mattered.

Both problems share a root: the firm already had the data and the prior decisions. Nobody could retrieve them fast enough to act.

What we built

Two things, built against the firm’s own cloud-hosted data, not a model’s memory.

Real-time executive dashboards. Leadership opens a view that reads live from the firm’s systems and answers the questions partners actually ask, with every figure tracing back to the record it came from.

• A leader asks where a project stands and gets the current numbers, pulled live from the firm’s records with a link to each source.
• A partner asks which jobs are trending over budget and sees them ranked, each line citing the underlying entries.
• Someone asks a follow-up in plain language and the dashboard answers from retrieved firm data, not a guess.

AI-assisted NDA and contract review. The reviewer uploads a document and the assistant surfaces the clauses worth a human’s attention and drafts suggested redlines, citing the source text every time.

• A reviewer drops in an NDA and the assistant flags the clauses that differ from the firm’s prior positions, pointing at the exact paragraph.
• A reviewer asks how a payment-terms clause compares to last time and gets an answer cited to the earlier contract.
• The assistant drafts a redline the reviewer can accept, edit, or reject, and nothing is sent or signed without a person.

How it’s grounded

This is a RAG build, not a freestanding agent. Every time the system speaks, it speaks from the firm’s own data, retrieved at request time: live records behind the dashboards, the firm’s stored contracts and NDAs behind the review assistant. We do not ask the model to remember facts about the firm. We hand it the facts and ask it to reason over them.

Retrieval before generation. A dashboard answer is assembled from records pulled at the moment of the question. A contract flag is assembled from the clause in front of the reviewer plus the firm’s prior agreements. No retrieved source, no answer.

Citations are not optional. Every figure on a dashboard links to the record behind it. Every clause the assistant flags points at the paragraph it read. The reviewer or the partner can always click through to the source and check our work, which is how a skeptical user starts to trust the output.

A human decides, every time. The dashboards inform a decision; they do not make one. The contract assistant drafts and surfaces; it does not negotiate, send, or execute. A person reviews and signs. We built it as decision-support, with the human in the loop by construction, not as a setting we can switch off.

What we were careful about

Data security and access control came first. This is the firm’s financial and legal data. Who can see which dashboard, and who can read which contract, follows the firm’s own access rules. The system does not widen access; it surfaces what a given person is already entitled to see.

No fabricated terms, no fabricated financials. The fastest way to lose a legal or finance team is to make up a clause or a number. The assistant only quotes and reasons over text that exists in the firm’s documents. If a contract is silent on a point, the answer is that it is silent, not an invented term that sounds plausible.

Grounding is enforced, not encouraged. An answer without a retrieved source does not ship to the user. We would rather the system say it cannot find the basis for an answer than hand a partner a confident figure with nothing behind it.

What we’d carry to your build

1. Citations are what earn trust from finance and legal teams. The people who read contracts and sign off on budgets do not want a black box. They want to see the source and decide for themselves. Building the click-through to the underlying record was not a polish step; it was the feature that got the tool used.

2. The data has to be clean enough to retrieve over. Both builds work because the firm’s underlying systems hold real, current records. If your data is scattered, stale, or contradictory, the honest move is to fix that before you put AI on top of it. Retrieval over messy data just gives you confident wrong answers faster.

3. Decision-support beats automation for high-stakes work. For dashboards a partner acts on and contracts a firm signs, the right design keeps a person in the loop. We draft, flag, and surface; the human decides. That boundary is not a limitation we apologize for. It is the reason the firm trusts the output enough to use it.

What this means for you

If your leadership cannot see the business in real time, or your team is reading every contract by hand and re-deriving decisions you already made, the shape of this build generalizes. The use case rarely transfers cleanly; the architecture does. We would start by understanding which questions your leaders ask most often and whether your data is clean enough to answer them.

The fastest read on whether your org is ready is the AI Readiness Assessment. The data-readiness dimension is the one a grounded build like this lives or dies on.

Or book a 30-minute discovery call and bring the dashboard you wish you had, or the contract pile nobody wants to read. We will tell you whether it is a RAG system, workflow automation, or the kind of project that needs data cleanup first.

— The AIX Automation Lab team