Proof
Results from delivered work.
Three accounts of work we have done, including the practice we run ourselves. Client names are withheld by agreement. The numbers are real.
A major Australian transport operator
3,200 documents sorted in two days. The estimate by hand was three to four weeks.
- The situation
- Before a major review of how the organisation manages its assets, more than 3,200 documents had to be found, read and matched to the questions being asked. Done by hand, that work delays everything behind it.
- What we built
- A pipeline that read each document, worked out what it covered and sorted it against the structure of the review.
- What stayed with people
- Every judgement about quality. The system found and sorted the evidence. Our assessors read it and decided what it meant.
- What it changed
- The assessment timeline was reduced by seven weeks, and senior people spent their time assessing instead of searching.
- to sort the documents, against an estimate of 3 to 4 weeks
- 2 days
- accuracy on checked samples
- 94%
- fewer consultant hours on document review
- ~60%
Source: SAS-AM project record, client anonymised.
The same pattern applies to supplier contracts, tender responses and policy libraries.
An asset intensive operator
Thousands of free text job records, turned into data a manager can act on.
- The situation
- Years of maintenance job records had been typed in free text. Nobody could say which problems came up most, because no two people described a fault the same way.
- What we built
- A classifier that matched each record to a known cause of failure. Around it we built a confidence framework, which turned out to matter more than the model.
- What stayed with people
- Anything the system was unsure of went to a review queue. An engineer made the call.
- What it changed
- Years of notes became something that could be counted and compared. The system also surfaced causes of failure that the existing analysis had never listed.
- job records processed
- 3,126
- classified by the system, the remainder went to a person
- 99.2%
- causes of failure the existing analysis had missed
- 337
Source: SAS-AM project record, client anonymised.
The same pattern applies to service tickets, customer complaints and warranty claims.
Our own operation
We run this practice on the systems we build for clients.
SAS Asset Management runs a governed fleet of AI agents for its own research, drafting, estimating and project work. It is in use every working day.
Every output carries a review level. Nothing is published, sent or delivered until a person has approved it at the level the work demands. The system drafts. A person decides.
We treat the agents as untrusted by design. Each one can reach only what its task needs, and anything it produces is checked before it counts.
This is the strongest evidence we can offer. The advice on this site comes from running these systems ourselves, including the parts that went wrong.
Why it matters to you. The policies we recommend are the ones we follow ourselves. They are set out on the page about how we use AI.
What this proof does not show
We have not yet published a result from a smaller business.
These results come from large, asset intensive organisations. That is where SAS Asset Management has worked for years, and where our AI work began.
The method carries over: measure the task, score the idea, build beside the team, keep a person at each checkpoint. What we cannot yet show you is a published figure from a business of your size.
That is why our first engagement has a fixed scope you can read before you commit, and why references are available in conversation.
Next
Ask us about the work behind the numbers.
Client names are withheld by agreement. References are available in conversation.