Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · October 2026
A recent study from MIT FutureTech asks whether AI automation will arrive like a crashing wave or a rising tide. Drawing on more than 60,000 worker evaluations of AI responses to more than 6,000 text-based workplace tasks, the researchers found a rising tide: AI capability is improving broadly, and most text-based tasks could reach success rates of 88 to 97 percent by 2030 at a minimally sufficient quality level, in what they describe as an upper-bound scenario. As lead author Matthias Mertens put it, “The key point is that this gives us time.”
Success rates varied by type of work, ranging from below 50 percent for text-based legal tasks to more than 70 percent for text-based tasks associated with installation, maintenance and repair, the top of the range.
That is the world I worked in 28 years ago.
What 1998 achieved, with 1998 technology
In 1998, a defense maintenance operation was failing. Next-day delivery of parts, and with it the call close rate the contract was measured on, was running at 51 percent on a contract that required 90 percent.
Every task in that operation was being done correctly as it was measured. Service technicians were measured on call volume, so they batched their part orders to late afternoon. Buyers placed the orders they received. Finance followed its rules. Suppliers shipped. Evaluated one at a time, each of those tasks would have been rated a success. The operation was still failing.
We did not start with the tasks. We started with the failure, and traced back through the multiple and seemingly unrelated strands that produced it: the time of day orders arrived, the narrowing of sourcing and transportation options through the afternoon, the customs cutoff, and a part that would have cost $100 at 9 a.m. costing $900 by late afternoon. Only then did we build the system: self-learning algorithms, working as firm fences within a nascent AI framework, with people weighting what mattered and bringing in what no system could see. The technology came last.
Within three months, next-day delivery rose from 51 percent to 97.3 percent. It held there for seven years.
The difference between the two numbers
MIT’s 70 percent and our 97.3 percent do not measure the same thing, and that is exactly the point.
MIT measured how often an AI model, working alone and given everything it needed, produced an acceptable answer to a single written task. For text-based maintenance-related tasks, the study reports success rates of more than 70 percent.
In 1998, with 1998 technology, we measured whether an entire maintenance operation delivered the right part, to the right place, by the next morning. That is every task, every handoff, every supplier, courier and customs broker, working together under real conditions. It reached 97.3 percent.
The MIT researchers are careful to say what their study leaves out. In the study, the AI was given all the information it needed to complete each task; in practice, they note, that information may be incomplete, difficult to integrate or constrained by regulation. That caveat is where the outcome is decided. It is where the 1998 operation was failing, and it is what the 1998 methodology solved.
See it for yourself
In this six-minute video, I walk through the difference between how most organizations see their operation, as a sequence of tasks, and how the 1998 operation actually worked: as strands, where failures happen in the seams between tasks and ripple across them at the same time.
The takeaway
The tide may be rising, and MIT is right that it gives organizations time. The question is what they do with it. Better models will keep raising the score on individual tasks. They will not, on their own, tell an organization where its tasks are defeating one another.
That was true in 1998, and it is true now. The technology has changed. The methodology has not. Today, with ARA™ RAM 2025™ running on SLAP OS™, people and multiple AI agents trace those strands together, and over the past six months that approach has averaged 91 percent output accuracy, assessed against the evidence in our documented case labs.
Capability is one strand. The outcome is decided by how everything around it is orchestrated.
In 2026, AI Succeeds on 70% of Maintenance Tasks. In 1998, Firm Fences Took a Maintenance Operation to 97.3%.
Posted on October 19, 2026
0
Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · October 2026
A recent study from MIT FutureTech asks whether AI automation will arrive like a crashing wave or a rising tide. Drawing on more than 60,000 worker evaluations of AI responses to more than 6,000 text-based workplace tasks, the researchers found a rising tide: AI capability is improving broadly, and most text-based tasks could reach success rates of 88 to 97 percent by 2030 at a minimally sufficient quality level, in what they describe as an upper-bound scenario. As lead author Matthias Mertens put it, “The key point is that this gives us time.”
Success rates varied by type of work, ranging from below 50 percent for text-based legal tasks to more than 70 percent for text-based tasks associated with installation, maintenance and repair, the top of the range.
That is the world I worked in 28 years ago.
What 1998 achieved, with 1998 technology
In 1998, a defense maintenance operation was failing. Next-day delivery of parts, and with it the call close rate the contract was measured on, was running at 51 percent on a contract that required 90 percent.
Every task in that operation was being done correctly as it was measured. Service technicians were measured on call volume, so they batched their part orders to late afternoon. Buyers placed the orders they received. Finance followed its rules. Suppliers shipped. Evaluated one at a time, each of those tasks would have been rated a success. The operation was still failing.
We did not start with the tasks. We started with the failure, and traced back through the multiple and seemingly unrelated strands that produced it: the time of day orders arrived, the narrowing of sourcing and transportation options through the afternoon, the customs cutoff, and a part that would have cost $100 at 9 a.m. costing $900 by late afternoon. Only then did we build the system: self-learning algorithms, working as firm fences within a nascent AI framework, with people weighting what mattered and bringing in what no system could see. The technology came last.
Within three months, next-day delivery rose from 51 percent to 97.3 percent. It held there for seven years.
The difference between the two numbers
MIT’s 70 percent and our 97.3 percent do not measure the same thing, and that is exactly the point.
MIT measured how often an AI model, working alone and given everything it needed, produced an acceptable answer to a single written task. For text-based maintenance-related tasks, the study reports success rates of more than 70 percent.
In 1998, with 1998 technology, we measured whether an entire maintenance operation delivered the right part, to the right place, by the next morning. That is every task, every handoff, every supplier, courier and customs broker, working together under real conditions. It reached 97.3 percent.
The MIT researchers are careful to say what their study leaves out. In the study, the AI was given all the information it needed to complete each task; in practice, they note, that information may be incomplete, difficult to integrate or constrained by regulation. That caveat is where the outcome is decided. It is where the 1998 operation was failing, and it is what the 1998 methodology solved.
See it for yourself
In this six-minute video, I walk through the difference between how most organizations see their operation, as a sequence of tasks, and how the 1998 operation actually worked: as strands, where failures happen in the seams between tasks and ripple across them at the same time.
The takeaway
The tide may be rising, and MIT is right that it gives organizations time. The question is what they do with it. Better models will keep raising the score on individual tasks. They will not, on their own, tell an organization where its tasks are defeating one another.
That was true in 1998, and it is true now. The technology has changed. The methodology has not. Today, with ARA™ RAM 2025™ running on SLAP OS™, people and multiple AI agents trace those strands together, and over the past six months that approach has averaged 91 percent output accuracy, assessed against the evidence in our documented case labs.
Capability is one strand. The outcome is decided by how everything around it is orchestrated.
Sources and related reading:
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related