all research

Construction Management · 8 min read

The first draft of every RFI: document agents on live construction projects

Site teams lose more hours to structured paperwork than to any technical problem. That work has a shape, and the shape is automatable.

11.4 hrs

Weekly admin recovered per PM

2.1 days

RFI turnaround, from 6.8

412 pp

Tender reviewed in one pass

Ask a project manager on an Irish or UK site where their week goes and the answer is rarely construction. It is the request for information that needs writing up properly, the submittal that needs checking against the specification, the daily log nobody filled in, and the tender addendum that changed something in a document five hundred pages long.

This is highly structured work. It has a template, a required set of references, and a defined output. Those properties are exactly what makes it suitable for agents, and they are also why construction has adopted this faster than more glamorous applications.

Where the hours actually go

Project manager time by activity, typical week on a live commercial project
hours per week
RFI drafting and chasing6.4
Submittal review5.1
Daily logs and reporting4.2
Tender and variation review3.8
Correspondence and records3.3
Site walking and coordination12.5
Meetings8.7

The first five bars total 22.8 hours. That is more than half a working week on document handling, and it is the half that a competent PM most resents, because it is the half that does not require them.

What first pass means

The framing that works commercially is first pass, not replacement. An agent that drafts the RFI, cites the drawing and specification clause it relates to, identifies who needs to answer it and by when, and puts it in front of the PM for approval is doing about eighty percent of the work and none of the deciding.

The RFI loop, agent-drafted and human-approved
  1. 1

    Detect the trigger

    A site query, a drawing discrepancy, a specification ambiguity raised in the daily log.

  2. 2

    Locate the references

    Pull the relevant drawing revision, specification section and contract clause.

  3. 3

    Draft

    Write the RFI in the project's own template with references cited and a proposed position.

  4. 4

    Route

    Identify the responsible discipline and the contractual response window.

  5. 5

    Track

    Chase the response, escalate on the contractual deadline, log the outcome against the programme.

The tracking step is quietly the most valuable one. RFI turnaround is a programme risk that most projects manage through somebody's memory, and the cost of a late response is usually borne long after anybody remembers who was waiting on what.

Nobody on site wants an AI that makes decisions. They want the paperwork to have already started when they get back to the cabin.

Tender and bid analysis

The second application is less frequent and higher value. A tender pack runs to hundreds of pages across employer's requirements, specification, preliminaries and drawings, and estimating teams read it under time pressure with commercial exposure attached to anything missed.

An agent reading the full pack in one pass produces a different kind of output to a human reader: a complete list of every obligation, every deviation from the standard contract form, every performance requirement, and every place two documents contradict each other. It will not price the job. It will make sure the person pricing the job knows what is in it.

Tender review, 412 page pack
hours
First read and familiarisation90.594%
Obligation extraction12192%
Contract deviation check70.889%
Internal contradiction check60.493%
Estimator review of findings43.513%
conventional agent-assisted

What to be careful about

Two failure modes come up repeatedly. The first is deploying against document sets that are not under version control, where the agent confidently cites a superseded drawing revision. The second is allowing the agent to issue rather than draft, which converts a helpful tool into a contractual liability the first time it gets something wrong.

Both are solvable, and both need solving before the first live project rather than after.

Work with me

Run this model against your own project

I am Kanishk Kapoor, Technical Accounts Manager at AI Institute in Dublin. I build agentic AI systems with built-environment teams across Ireland and the UK. If any figure here looks wrong for your business, that is the useful conversation. Send me your assumptions and I will re-run it.

Continue reading