Use caseConstruction project software
Contractors want to ask Brickline's project files a question and get the page back. Two prototypes later, the parts list keeps growing.
Answers cite the document and the pageThe assistant reads only what the user may readThe model changes with a setting, not a release
Stairwell doors need a 90-minute fire rating with self-closing hardware. Door 7-S2 also needs panic hardware.
source · spec 08 71 00 p. 12, door schedule rev. D p. 3Found in the documents · each statement links to its page
4,812 documents sit in one Harrow Build project alone: specifications, drawings, change orders, requests for information. Brickline sells the project management software that holds them, and its customers want an assistant that answers from them. On a roadmap that looks small. Behind it sit search by meaning, a model to choose and host, one contractor's files kept from another, and a record of every answer. That is an infrastructure project, and it is not Brickline's product.
Brickline calls those parts instead of building them. Project documents go into the repository by API with the rights Brickline already manages. The assistant answers through an API, reads only what the asking user may read, and cites its pages. Brickline builds the panel its users see; the pipeline behind it is already there.
Three moves carry this page: the documents arrive with their rights, the task is written once in the console, and every answer in Brickline's panel comes back with its pages. Priya makes the first two in a week; Luis meets the third on level 7.
Priya Raman, staff engineer, has a week before the internal demo and three things to connect.
Brickline's back end sends each project document to the repository through its REST API: the file, its project, its discipline and the groups allowed to read it. Content and metadata are indexed on write.
In the console, Priya picks the model for each task. Providers are grouped by hosting and flagged by capability. Harrow Build's contracts ask for US-only hosting, so its tasks run on a sovereign model hosted in the US.
Priya and Daniel Osei, VP of product, write three tasks in the prompt library: answer from the project documents, draft a reply to a request for information, and summarize the change orders. The instructions ask for a citation on every statement, and for a plain sentence when the documents do not hold the answer.
The panel is in Brickline's project page. Luis Ortega, project manager at Harrow Build, takes it onto the site.
Brickline's back end calls the assistant API with the question, the project and the user. Permissions apply before the call: the assistant reads only the documents that user is entitled to see, and searches them by meaning.
Luis is on level 7 with the door installer. He asks a follow-up in the panel on his tablet and taps the first citation. The specification opens in the viewer on page 12, inside Brickline.
A request for information arrives from the drywall subcontractor about the wall rating around stairwell B. Luis clicks Draft an RFI reply. The draft highlights the values it took from the documents. He changes the partition type, adds the site meeting date and sends it under his name.
The demo goes well. Brickline opens the assistant to customers, then keeps an eye on what it costs and what it said.
At Harrow Build, Keisha Brooks manages Brickline for the whole company. She keeps specifications, drawings and requests for information open to project managers, and closes bids, subcontract prices and claims to every group but management.
Every Monday Daniel reads the console dashboard: requests, conversations and tokens across every surface, the model mix, and consumption by task. Brickline uses it to size its plans and to see which task its users actually rely on.
In March, Harrow Build asks why an answer from November quoted door schedule revision D, when revision E now exists. Priya opens the exchange: the question, the answer, the provider, the model, the date and the cited source are all kept. In November, revision D was the current one. The check takes minutes.
Every question to the assistant ends in one of three places. Each one tells the user what to do next.
Each statement links to the document, the page and the revision it comes from.
What follows: the user opens the page and checks it before acting.The task's instructions ask for a plain sentence, with no citation and no guess.
What follows: the user asks the architect or files a request for information.Permissions apply before the call, so the document is never read for this user.
What follows: the answer is built from what the user may see, or not at all.It sends nothing to a subcontractor. It grants nobody access. It changes no document. Across the nine steps on this page, every message that leaves Brickline is sent by a named person, who read the draft first.
The assistant searches, answers with sources and drafts. The rights are Brickline's and its customers', and the AI works inside them.
Three counts from Brickline's first customer, Harrow Build. A product with other documents and other customers will tally its own.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
A document assistant ships inside a product only when the product lead, the engineers and each customer admin are on board. Their reasons differ:
The assistant on your roadmap keeps turning into an infrastructure project.
Search, models, rights and logs: four systems to build and run, none of them your product.
You are asked to trust an AI with every folder your company owns.
With one task and one customer. Take the question your users ask support most often, and answer it from one project's documents in a test environment.
Then put the panel in front of an internal team before any customer sees it, and read the exchanges together.
Next step
Bring a sample of one customer's documents and the questions their users ask. We will get your team to a first cited answer in your own screen.