AI at work / In development
AI may help read documents, organize information and prepare a first draft. The useful question is whether it improves the work after someone has checked it.
We are developing this approach. The examples below describe ideas to test, not finished products or client results.
How we would test itHow we would start
Choose a specific piece of work and agree on what a useful result would look like. Compare the new method with the way the task is done today.
Follow the task from the first document to the final decision. Find out where people repeat work, wait for information or correct mistakes. Check whether the source material is reliable enough to use.
What we would write downThe task, its current cost and the improvement we would test for.
Use a limited set of documents and give one person responsibility for checking the output. Decide which information the tool can use, what it may do and when it must stop for a person to review it.
What we would write downThe test, who checks it and the conditions for stopping.
Compare accuracy, time and cost with the original method. Include setup, software, checking, corrections and upkeep. If the result falls short, change the test or stop.
What we would write downWhat worked, what failed and whether to continue.
Ideas we are exploring
These examples show where we would begin testing. They do not describe software that is ready to buy.
A possible starting point is to gather facts from deal documents, show where each fact came from and flag gaps for review. That could help someone prepare an initial analysis or a committee paper.
A person would still need to check the sources, make the assumptions and decide whether the investment makes sense.
Another task to test is comparing property reports, finding figures that do not agree and listing questions for the asset manager. The test would need to show whether it catches useful issues without creating too many false alarms.
Every flagged figure would need a source that the reviewer can check.
Find the right documents, identify the current versions and agree on what the numbers mean. Decide who may see each file. These steps matter even if no AI tool is used.
Someone must be responsible for the answer. The test should make it possible to trace important facts to their sources and should require a person’s approval before the tool takes actions beyond its agreed limits.
Testing our own work would be an internal check. An independent review would require a separate reviewer.
Count software, setup, staff time, checking, corrections and upkeep. Compare those costs with a measurable improvement in the work. A draft produced in seconds may still need a long review.
Start with a task
Tell us what the task involves, who does it and where it gets stuck. We can start by discussing whether it is worth testing a different way.
Tell us about the taskThis opens your email application. Please describe the task without sending confidential documents.