Imagine a cost estimating team asking an AI agent to gather unit rates from past projects for a preliminary cost plan. The table looks complete and the totals add up, but some rates come from a different type of building. If the team reviews only the total, it may miss the question that matters for the estimate. Are those rates suitable for this project?

The Royal Institution of Chartered Surveyors (RICS) uses a case like this to explain the Judgement Layer. It is a way of organising work so professionals have the evidence and opportunity to exercise judgement when it matters, from defining the task and reviewing evidence along the way to deciding whether the AI output should inform advice to a client.

Checking how the answer was produced

When a chatbot summarises a document, people typically exercise judgement when they frame the task and when they assess the answer. An AI agent may search for documents, select evidence, use tools and adjust its approach as it goes. If it picks the wrong drawing revision or uses rates from projects that are unsuitable for comparison, the error can carry through to the final report.

Where do people exercise judgement?

Chatbot

  1. Frame the task
    Define what the chatbot should answer.
  2. The chatbot responds
  3. Assess the answer
    Decide which parts to rely on.
  4. Professional advice to the client

AI agent

  1. Design the workflow and set the goal
    Specify permitted tools, evidence and checkpoints.
  2. The agent works in cycles
    Plan → act → observe → adjust.
  3. Decide at a checkpoint
    Continue or redirect ↶ back to the agent.
    Keep paused → wait for a decision.
  4. Assess the completed work
    Decide whether the output can inform the advice.
  5. Professional advice to the client
Figure 1. The workflow can pause whenever a condition set by the team is met. Adapted from Figure 1 of the RICS paper.

A team might require the agent to pause when sources disagree, a significant assumption changes or the evidence falls outside the agreed scope. The person responsible can then ask for more evidence or redirect the work before the result feeds into the next step.

The three parts of the Judgement Layer

Connect makes relevant evidence available to the agent and gives people a way to trace the sources it used. For a BIM team, that might start with identifying where approved drawings and models live and how to distinguish them from drafts.

Codify means writing down the criteria behind recurring decisions. A team might record which project types provide comparable rates, which assumptions need further explanation and when to ask a senior colleague to review the work.

Govern sets permissions, checkpoints and responsibilities. The team decides what the agent may do, when it must pause, who can stop it and who signs off before the work reaches the client. It then tests whether the agent pauses under the agreed conditions.

When the agent finds a rate from a different building type

  1. Connect
    Give the agent access to the firm's rate library and project classifications.
  2. Codify
    Apply the team's criteria to flag a rate from a different project type.
  3. Govern
    Pause the workflow and send the flagged rate and its source to the cost estimator.
  4. The cost estimator decides
    Keep the rate and record why, choose another rate or request more evidence.
Figure 2. How the three parts work together in a cost plan. Adapted from the example in section 6.4 of the RICS paper.

Record what the AI did and why the professional decided

RICS describes two records. The system trail captures sources, tool calls, assumptions and intermediate results. The professional decision record explains what evidence the reviewer examined, where they found limitations and why they accepted, changed or rejected the output.

For a cost plan, the system trail might show which project supplied a unit rate, while the estimator’s record explains why they chose a rate from another project. A reviewer can then trace both the source of the figure and the reason for using it. RICS guidance on record keeping

Start with one recurring task

A construction or BIM team can start with a task it already understands, such as gathering unit rates for a preliminary cost plan. Identify the evidence the team normally uses, write down one or two recurring decisions, and test the workflow on routine cases and cases where evidence conflicts. Check whether the agent uses the agreed sources and pauses when expected. Use the findings to refine the data, criteria or checkpoints.

The trial also gives junior colleagues a chance to frame the task, inspect sources and see how experienced colleagues question assumptions. As AI takes on more of the work of gathering evidence, involving junior staff in reviews and asking them to explain their reasoning before sign-off gives them practice in exercising professional judgement. RICS guidance on getting started and developing professional judgement

Adapted from How to build a judgement layer for AI use in surveying by Anil Sawhney, Chris de Gruben, James Garner and David Wong, published on 19 October 2026. The example on controlling drawing and model revisions applies the framework to BIM practice. RICS publication page

← All notes