Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
On 27 August, Cisco announced that MyAgent is rolling out to roughly 90,000 employees. Thimaya Subaiya, EVP Operations, described it as moving past chat-based assistance into “supervised autonomous execution” across the applications people already use. It runs on Circuit, Cisco’s governed multi-model platform, and it selects a model per task rather than routing everything to the most expensive one.
It is one of the largest enterprise AI deployments anyone has attempted.
It is also the second time I have watched Cisco build something I had already described.
December 2007
Nineteen years ago I wrote a post on this blog about Cisco’s manufacturing management system, Autotest. Working from Neil Shister’s World Trade Magazine piece, I noted that Autotest pulled real-time data from globally disparate manufacturers running disparate operating systems into a single view — and that it had, in Shister’s words, “intelligent agents built into the system,” capable of initiating corrective action remotely.
I said at the time that this looked like the elemental roots of a Metaprise™-based application.
And I anchored it to something I had written three years before that. From Acres of Diamonds, fall 2004:
It is my position that a true centralization of procurement objectives requires a decentralized architecture that is based on the real-world operating attributes of all transactional stakeholders starting at the local or regional level. In other words, your organization gains control of it’s spend environment by relinquishing centralized functional control in favor of operational efficiencies on the front lines. This is the cornerstone of agent-based modeling.
Centralize the objective. Decentralize the architecture.
Read the MyAgent announcement against that paragraph. The objective is held centrally — approved models, approved systems, enterprise-appropriate data pathways, governance and oversight designed in rather than bolted on. The capability is pushed to every front line, one agent per person, adapted to what that person actually does.
That is the 2004 formulation, running at 90,000 seats.
I am not claiming Cisco read the paper. Twice now they have arrived at this independently, which is a stronger fact than influence would be. When two parties working from different starting points reach the same architecture, that is evidence about the architecture, not about either party.
The agents were never only software
There is a detail in the 2004 language that matters more in 2026 than it did then.
All transactional stakeholders. Not all systems. Stakeholders — the technician releasing an order, the buyer holding a queue, the supplier, the courier, the customs authority, the approver.
I have never distinguished between human agents and AI agents. They are agents. What differs is placement and function, not kind: which decisions each one holds, which it hands back, and what happens when the expected result does not arrive.
That is not a position I adopted when agentic AI became fashionable. Yesterday I wrote about WO 01/65428 A2, a specification published through the international patent system on 7 September 2001, with priority filings in March 2000, naming me as inventor. Its business rules do not have one set for people and another for the system.
So MyAgent is not a new ecosystem arriving to be orchestrated. It is a new kind of agent joining a field that was already mixed — human and software, internal and external. MyAgent inherits that field, and every misalignment already in it.
What Cisco got right, and I want to be specific about it
In July I wrote that personal AI was becoming the new shadow spreadsheet: adopted bottom-up, for convenience, entirely outside anything the enterprise could see or stand behind. I argued you cannot ban it, because people route around a prohibition the moment the sanctioned path is slower than the unsanctioned one. And you cannot consolidate your way out — the roughly 367 applications Forrester put the average large enterprise at in 2022 are the monument to that belief.
That post did not stop at the constraint. It named the answer:
You do not govern the divide by forcing all AI up into one central platform the front lines will route around. You govern it by giving the front lines governed capability — provenance and human judgment built in — inside a human-led, agent-based Metaprise™ model. Centralize the objective. Decentralize the architecture.
MyAgent is that, executed. Governed capability at every front line; the objective — approved models, approved systems, oversight — held centrally.
Six weeks separate the two, which is the shortest interval in this sequence by a wide margin. And naming an answer is not the same achievement as shipping it to 90,000 people. The second is harder, and Cisco has done it.
The question rollout does not answer
Ninety thousand agents is a distribution achievement. It is not yet an authority architecture.
An apprentice can help every employee. It does not follow that every apprentice should be permitted to act at every decision point. The next question is not how to govern 90,000 agents. It is what each agent is allowed to do, on what evidence, and where human judgment remains decisive.
Cisco’s own framing is careful here — employees set intent, apply judgment, and remain accountable for outcomes. That is the right placement, and it is worth noticing that it is the same placement the 2000 specification describes: the system detects that an expected result has not arrived; a named human decides what to do about it. Twenty-six years of capability improvement, and the element that decides the outcome sits in the same place.
Here is the limit that no amount of capability moves.
That 2000 architecture could do something quite sophisticated: where it could not meet a delivery requirement, it derived an achievable alternate requirement and re-solved against that. A system permitted to change its own objective within bounds and keep going.
It still could not have found the four o’clock.
In 1998 a defence maintenance operation was delivering next-day parts 51% of the time against a 90% requirement. The determining condition was that service technicians were batching order releases to late afternoon, because they were measured on call volume and releasing orders as they arose interrupted service calls. Rational behaviour, one department away from the people being blamed, in nobody’s process document.
A system can renegotiate a variable it represents. It cannot discover one it does not. That boundary is not a function of model quality, and a better model does not move it.
Cisco has been on both sides of this
An example of why the 2000/2001 pattern of the patent document holds up so well in 2026 is that the same company supplies the counter-case.
In that 2007 post I recalled that Cisco’s adherence to its APS software recommendations in the late 1990s had it increasing production as the market turned into the dot-com collapse — mass layoffs, and a significant write-down of excess inventory. Same firm. Sophisticated technology, correctly followed.
What changed by 2007 was not the software. It was, in Jim Miller’s framing, getting people out of functional silos and thinking holistically. The attitude moved first and the system scaled the result.
And Cisco is saying something adjacent right now. EVP Liz Centoni has described the shift inside a large enterprise as “surgery without the drugs” and noted that adding AI to existing workflows did not, by itself, resolve the deeper issues underneath them.
That is a Cisco executive stating the readiness point in Cisco’s own voice. It is the most important sentence in the coverage and it has attracted the least attention.
In June I set the Cisco and Boeing cases side by side. In 2008 both were building what each called a complex adaptive network — the same architecture, named in the same paragraph. Eighteen years later Cisco sits in the top five of Gartner’s 2026 supply chain ranking and Boeing’s became a case study in failure. Boeing had the platform; a dated 2025 assessment rated it high on infrastructure and adaptive-network fit. The platform was present and the outcome failed anyway.
The variable was never the platform. It was whether something forced the organization to confront how it actually operates before it chose the tool.
Today’s takeaway
Cisco has removed the shadow by making the governed lane the convenient one, which is the correct response to a problem most organizations are still trying to ban their way out of.
What that rollout cannot settle is placement. Ninety thousand agents will amplify whatever operating reality they are dropped into — and if that reality has not been validated, a more capable agent widens the gap between what could have been delivered and what actually is. That gap is what the Hansen Deflator Formula™ exists to make visible, and its defining property is counterintuitive: it grows as the technology improves.
The architecture is right. It was right in 2004, and Cisco was building a version of it in 2007. The unfinished work is the same as it was then, and it is not technical.
Keep the human at the wheel. Everything else is just faster.
-30-
This analysis draws on the Procurement Insights archive — an independent record carrying zero vendor sponsorships, published openly since 2007 and consolidating documented client work, lectures, and articles reaching back to 1998; nearly three decades of contemporaneous observation, gathered in one place rather than created there. Every claim is held to the Provenance Ledger™: a verify-before-publish discipline that traces each assertion to a primary source and never quietly edits the record once posted. That record is the evidence base for two working lenses — Invariant Physics™, the constant that however far the technology advances the operating logic must be in place first, and Implementation Physics™, its per-engagement application. Phase 0™ is the discipline that validates operating conditions before the amplifier is switched on.
Getting it right rather than being right.
Jon W. Hansen, FCIPS — Procurement Insights | Hansen Models™
-30-
Related
Cisco Just Gave 90,000 People an Agent. The Architecture Is Not New — and the Hard Question Is Still Unanswered.
Posted on August 29, 2026
0
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
On 27 August, Cisco announced that MyAgent is rolling out to roughly 90,000 employees. Thimaya Subaiya, EVP Operations, described it as moving past chat-based assistance into “supervised autonomous execution” across the applications people already use. It runs on Circuit, Cisco’s governed multi-model platform, and it selects a model per task rather than routing everything to the most expensive one.
It is one of the largest enterprise AI deployments anyone has attempted.
It is also the second time I have watched Cisco build something I had already described.
December 2007
Nineteen years ago I wrote a post on this blog about Cisco’s manufacturing management system, Autotest. Working from Neil Shister’s World Trade Magazine piece, I noted that Autotest pulled real-time data from globally disparate manufacturers running disparate operating systems into a single view — and that it had, in Shister’s words, “intelligent agents built into the system,” capable of initiating corrective action remotely.
I said at the time that this looked like the elemental roots of a Metaprise™-based application.
And I anchored it to something I had written three years before that. From Acres of Diamonds, fall 2004:
Centralize the objective. Decentralize the architecture.
Read the MyAgent announcement against that paragraph. The objective is held centrally — approved models, approved systems, enterprise-appropriate data pathways, governance and oversight designed in rather than bolted on. The capability is pushed to every front line, one agent per person, adapted to what that person actually does.
That is the 2004 formulation, running at 90,000 seats.
I am not claiming Cisco read the paper. Twice now they have arrived at this independently, which is a stronger fact than influence would be. When two parties working from different starting points reach the same architecture, that is evidence about the architecture, not about either party.
The agents were never only software
There is a detail in the 2004 language that matters more in 2026 than it did then.
All transactional stakeholders. Not all systems. Stakeholders — the technician releasing an order, the buyer holding a queue, the supplier, the courier, the customs authority, the approver.
I have never distinguished between human agents and AI agents. They are agents. What differs is placement and function, not kind: which decisions each one holds, which it hands back, and what happens when the expected result does not arrive.
That is not a position I adopted when agentic AI became fashionable. Yesterday I wrote about WO 01/65428 A2, a specification published through the international patent system on 7 September 2001, with priority filings in March 2000, naming me as inventor. Its business rules do not have one set for people and another for the system.
So MyAgent is not a new ecosystem arriving to be orchestrated. It is a new kind of agent joining a field that was already mixed — human and software, internal and external. MyAgent inherits that field, and every misalignment already in it.
What Cisco got right, and I want to be specific about it
In July I wrote that personal AI was becoming the new shadow spreadsheet: adopted bottom-up, for convenience, entirely outside anything the enterprise could see or stand behind. I argued you cannot ban it, because people route around a prohibition the moment the sanctioned path is slower than the unsanctioned one. And you cannot consolidate your way out — the roughly 367 applications Forrester put the average large enterprise at in 2022 are the monument to that belief.
That post did not stop at the constraint. It named the answer:
MyAgent is that, executed. Governed capability at every front line; the objective — approved models, approved systems, oversight — held centrally.
Six weeks separate the two, which is the shortest interval in this sequence by a wide margin. And naming an answer is not the same achievement as shipping it to 90,000 people. The second is harder, and Cisco has done it.
The question rollout does not answer
Ninety thousand agents is a distribution achievement. It is not yet an authority architecture.
An apprentice can help every employee. It does not follow that every apprentice should be permitted to act at every decision point. The next question is not how to govern 90,000 agents. It is what each agent is allowed to do, on what evidence, and where human judgment remains decisive.
Cisco’s own framing is careful here — employees set intent, apply judgment, and remain accountable for outcomes. That is the right placement, and it is worth noticing that it is the same placement the 2000 specification describes: the system detects that an expected result has not arrived; a named human decides what to do about it. Twenty-six years of capability improvement, and the element that decides the outcome sits in the same place.
Here is the limit that no amount of capability moves.
That 2000 architecture could do something quite sophisticated: where it could not meet a delivery requirement, it derived an achievable alternate requirement and re-solved against that. A system permitted to change its own objective within bounds and keep going.
It still could not have found the four o’clock.
In 1998 a defence maintenance operation was delivering next-day parts 51% of the time against a 90% requirement. The determining condition was that service technicians were batching order releases to late afternoon, because they were measured on call volume and releasing orders as they arose interrupted service calls. Rational behaviour, one department away from the people being blamed, in nobody’s process document.
A system can renegotiate a variable it represents. It cannot discover one it does not. That boundary is not a function of model quality, and a better model does not move it.
Cisco has been on both sides of this
An example of why the 2000/2001 pattern of the patent document holds up so well in 2026 is that the same company supplies the counter-case.
In that 2007 post I recalled that Cisco’s adherence to its APS software recommendations in the late 1990s had it increasing production as the market turned into the dot-com collapse — mass layoffs, and a significant write-down of excess inventory. Same firm. Sophisticated technology, correctly followed.
What changed by 2007 was not the software. It was, in Jim Miller’s framing, getting people out of functional silos and thinking holistically. The attitude moved first and the system scaled the result.
And Cisco is saying something adjacent right now. EVP Liz Centoni has described the shift inside a large enterprise as “surgery without the drugs” and noted that adding AI to existing workflows did not, by itself, resolve the deeper issues underneath them.
That is a Cisco executive stating the readiness point in Cisco’s own voice. It is the most important sentence in the coverage and it has attracted the least attention.
In June I set the Cisco and Boeing cases side by side. In 2008 both were building what each called a complex adaptive network — the same architecture, named in the same paragraph. Eighteen years later Cisco sits in the top five of Gartner’s 2026 supply chain ranking and Boeing’s became a case study in failure. Boeing had the platform; a dated 2025 assessment rated it high on infrastructure and adaptive-network fit. The platform was present and the outcome failed anyway.
The variable was never the platform. It was whether something forced the organization to confront how it actually operates before it chose the tool.
Today’s takeaway
Cisco has removed the shadow by making the governed lane the convenient one, which is the correct response to a problem most organizations are still trying to ban their way out of.
What that rollout cannot settle is placement. Ninety thousand agents will amplify whatever operating reality they are dropped into — and if that reality has not been validated, a more capable agent widens the gap between what could have been delivered and what actually is. That gap is what the Hansen Deflator Formula™ exists to make visible, and its defining property is counterintuitive: it grows as the technology improves.
The architecture is right. It was right in 2004, and Cisco was building a version of it in 2007. The unfinished work is the same as it was then, and it is not technical.
Keep the human at the wheel. Everything else is just faster.
-30-
This analysis draws on the Procurement Insights archive — an independent record carrying zero vendor sponsorships, published openly since 2007 and consolidating documented client work, lectures, and articles reaching back to 1998; nearly three decades of contemporaneous observation, gathered in one place rather than created there. Every claim is held to the Provenance Ledger™: a verify-before-publish discipline that traces each assertion to a primary source and never quietly edits the record once posted. That record is the evidence base for two working lenses — Invariant Physics™, the constant that however far the technology advances the operating logic must be in place first, and Implementation Physics™, its per-engagement application. Phase 0™ is the discipline that validates operating conditions before the amplifier is switched on.
Getting it right rather than being right.
Jon W. Hansen, FCIPS — Procurement Insights | Hansen Models™
-30-
Share this:
Like this:
Related