An incoming claim from the client: missed deadlines, a penalty charged, ten business days to file a reasoned response. The contractor does have arguments – the delay started with the client's own late delivery, and this was put in writing at the time. The problem is that this "in writing" is scattered across a year and a half of correspondence, across acceptance certificates, work logs and meeting minutes. Two engineers and a lawyer drop everything for a week just to pull it all together.
We have studied the document flow of construction companies in Kazakhstan – general contractors and technical clients. The pain is always the same: not the absence of documents, but the inability to quickly find the right one and prove your position with it. Let's look at what AI agents do here.
How many documents one project generates
Over the course of construction, a mid-sized project accumulates three to six thousand incoming and outgoing letters, hundreds of hidden-works certificates and completed-works certificates, dozens of revisions of the design documentation sections, work logs, technical-supervision orders, minutes of weekly meetings. It lives in email, in folders on a server, in the document management system, and partly on the chief project engineer's flash drive.
A separate layer is messengers. Some decisions are made in the site managers' chats and stay right there: the decision existed, the document does not. In our reviews this is the most common gap in the evidence base, and it closes simply – a chat export is connected to the archive the same way a mailbox is.
Searching such an archive means interviewing the project's old-timers. "Remember, we approved that junction detail? Who was responsible – the designer or the client?" If that person has left, the navigation of the archive leaves with them. The cost of the question surfaces at the most expensive moments: a claim, a lawsuit, an inspection, handover of the project.
How does an AI agent search a project archive?
The agent indexes the entire body of material – letters with attachments, scanned certificates, PDF volumes, minutes – and searches by meaning, not by file name. The question is posed the way an engineer would ask it: "who approved replacing the facade insulation, when, and what confirms it?"
In one of the prototypes, the agent assembled the chain for such a question in a minute: the designer's letter with the justification for the replacement, the technical council minutes with the decision, the revision of the architectural (AR) section with the change entered, and the client's letter granting approval. Each item comes with a link to the document and page. The engineer does not take the agent's word for it: they open the documents by the links and check. Previously the same chain took two days to assemble – and not necessarily in full. We show how this search works on the page for the archive search solution.
A crucial detail: the agent answers only from the archive documents and always with links. If there is no confirmation, it says "not found" rather than inventing a plausible version. For claim and litigation work, this is condition number one.
Comparing versions of the design documentation
Revision 4 of a section has replaced revision 3. What changed? The "rev. 2" stamp on the title page does not answer that question. The agent compares the versions and produces a list: changed sheets, affected junctions and specifications, discrepancies with adjacent sections. Separately, it flags what drags scope and money along with it: a material substitution, added works, changed elevations.
In our reviews we regularly see one expensive habit: extra works get done "on a phone call", because no one tracked and formalized the change in the documentation in time. Claiming payment for such scope afterwards is almost impossible. Version comparison with an automatic list of changes closes the root cause itself: every change to the design becomes a visible event that a supplementary agreement can be attached to.
Every difference found gets a status: accounted for in the estimate, requires a supplementary agreement, awaiting the designer's response. This turns the question "has anyone seen rev. 4?" into a managed process with deadlines and owners.
Claims and letters: a draft with a timeline in an hour
Let's return to the claim from the start of the article. With the platform the process looks like this: the lawyer states the task – "assemble the timeline on the deadlines for block B and prepare a draft response". The agent goes through the correspondence, certificates and work log and builds the timeline: notice of the client's delayed delivery of the steel structures, the letter about suspension, the certificates for additional works, the downtime entry in the log. Along with the timeline comes a draft response with links to every document.
The lawyer edits rather than digs. A week of two engineers and a lawyer turns into a day for one lawyer. Quality rises for the same reason: the response includes everything that is in the archive, not everything that was remembered.
The stakes here run into tens of millions: the penalty under a standard works contract is 0.1% of the stage cost for each day of delay. A response assembled in two days instead of ten is a negotiating position, not just saved hours.
Incoming letters from contractors and the client are handled by the agent the same way: it classifies them, determines the response deadline, prepares a draft resolution – whom to assign it to and what to attach. Events keep the deadlines under control: two days left before a response is due and no outgoing letter yet – the owner gets a reminder, then escalation kicks in. For volumes of tender documentation there is a separate pipeline – extracting requirements into a checklist, which we have built into the documentation parsing solution.
Certificates and completeness of as-built documentation
Before a stage is handed over, the as-built documentation turns into a quest: are all hidden-works certificates signed, does each have the material certificates and passports attached, do the certificate dates line up with the work log. The agent checks the set against a list and produces the discrepancies: what is missing, where dates contradict each other, which certificates reference a cancelled revision. The list appears in hours, not on the night before the commission.
The same mechanism works on the way in: subcontractors' completed-works certificates are checked against the log and the contract before signing – scope, rates, references to the documentation. Disputed lines are highlighted, the grounds attached.
What manual searching costs
Cost of searching per month = number of search tasks × hours per task × specialist rate
An estimate from one of the reviews: about 120 search tasks per month per project – from "find the certificate" to "assemble the history of a junction", on average an hour and a half per task, with the rate of a technical-department engineer or a lawyer around 6,000 tenge per hour with taxes. That comes to about a million tenge a month – just for people searching through documents. Claims and downtime are counted separately and are usually an order of magnitude more expensive.
What it takes to make this work
Less than is commonly assumed. There is no need to "tidy up the archive" in advance – the agent works with whatever mess is there: the project's mailboxes, folders on the server, an export from the document management system. Scans are read along with stamps and handwritten notes. For more on why cleaning up data before you start is unnecessary, see the article on data for a prototype.
This runs inside the organization's perimeter: commercial correspondence and estimate information do not go outside, access is segmented by project and role, and every action of the agent is recorded in a log. The platform connects to the document management system and accounting systems via API, and where there is no API – through the interface.
We set up a prototype on one of your projects in about a week, free of charge: we take the archive, index it, and your own engineers ask it their usual questions. A production rollout takes from eight weeks, followed by a subscription. You can discuss your project through the contacts.
Frequently asked questions
The documents are all over the place: email, server, document management system. Is that a problem?
No, that is the starting state of almost every project we have seen. The agent builds an index from several sources at once; the only thing we need from you is access to those sources and the boundaries: which projects and folders we include. After that the index fills itself: the event "a document arrived" triggers processing.
Are drawings recognized too?
Partly. The agent works confidently with stamps, specifications, schedules and the text part of drawings, including scans. It does not check geometry or clashes – that is a CAD task, and we draw that line honestly at the demo.
How secure is this for commercial information?
The archive stays inside the organization's perimeter, nothing is passed outside. The agent's access mirrors employees' permissions: one project's subcontractor correspondence is not visible to another project's team. The action log shows who searched for what and which documents the agent read.