← Speaking & podcasts

Human Driven AI

Who is actually driving your network's AI?

The SME is the scarce resource. Not the model.

25–30 minute general session + Q&A All levels — engineers, leads, architects, and budget owners AI AdoptionNetwork AutomationLeadershipAgentic SystemsHuman-in-the-LoopAI GovernanceNetDevOps

"Why haven't we seen full adoption of network automation yet?" The honest answer isn't technical — it's human. The tools work, the business cases are proven, the frameworks exist. What's missing is a clear answer to the question every disrupted engineer is quietly asking: where do I fit now that AI is here? This talk argues that the SME — not the model — is the scarce resource, and that adopting AI is a leadership job, not an engineering upgrade. Through side-by-side workflow examples — incident triage, config generation, and source of truth — it contrasts the "AI-first, human-absent" anti-pattern with a "human-driven" architecture where people set the frame before AI enters. You leave knowing how to put your engineers at the helm — and exactly where to place the ICs that AI disrupts.

What you leave with

  • The bottleneck to AI adoption is organizational, not technical — and AI amplifies that pattern instead of fixing it.
  • "Human at the helm" beats "human in the loop" — set intent and guardrails before AI enters, don't just approve its output.
  • A repeatable pattern for human-driven workflows — the human owns intent and validation, AI proposes with evidence, execution is deterministic, and autonomy is earned up a trust ladder.
  • Where disrupted ICs actually go — three concrete, higher-value SME roles: knowledge curator, intent/workflow architect, and data steward.
  • How to lead the change — name the fear, write the intent down, demystify escalation, and measure value, not motion.

Why this topic

The network automation community keeps asking one question: why haven’t we seen full adoption yet? The tools work. The business cases are proven. The frameworks exist. And still the projects stall.

The honest answer isn’t technical. It’s human — and the data says so plainly. Roughly 70% of projects fail outright, and 80–90% of those failures are organizational rather than technical. The Wharton and MIT numbers on AI initiatives disagree wildly on the headline (75% succeed vs. 95% see zero return) but agree completely on the cause: organizational readiness, not technology, is the limiter. The London Stock Exchange burned £400M on Taurus before someone on the team said the quiet part — “it was always a political and a business problem.”

AI does not rewrite that pattern. It amplifies it. Networks demand something close to 100% accuracy, and a non-deterministic model pointed at a deterministic-critical system is a liability rather than a shortcut. So the interesting question stops being “how autonomous can we make this?” and becomes “where does the human’s judgment sit, and where do we place our people so that judgment compounds?”

The argument

The SME is the scarce resource, not the model. Models are becoming a commodity; context and judgment are not. No one knows your network, your team, and your business workflows the way your engineers do. That makes adopting AI a leadership job rather than an engineering upgrade — a different job, not engineer++.

The practical form of that argument is an architecture. The human owns the ends of the workflow: intent and guardrails at the front, validation at the gate. Automation runs the middle. AI proposes with evidence you can actually check, deterministic execution does what the SME defined, and feedback tightens the loop. Autonomy is earned by climbing a trust ladder — human-in-the-loop, then human-on-the-loop, then, only once it’s earned, autonomous.

The same shape works whether you’re unblocking one engineer, scaling a team out of toil, or compressing time-to-answer during a live incident. What changes across those three is blast radius — and as blast radius widens, the grip on the helm tightens rather than loosens.

Why it matters to the people in the room

The anxiety is real and worth naming out loud: the fear that the industry won’t upskill most of us, that it will just replace us with agents. This work is the other answer, and it isn’t wishful — Lumen retrained tier 1/2/3 engineers into automation engineers, codified their judgment into 350+ workflows, reached roughly 80% machine-to-machine operations, and moved people up rather than off the org chart.

The resistance that has to be led through is identity, not laziness. People aren’t avoiding Python; they’re avoiding a threat to who they’ve been for twenty years. That is a leadership problem with leadership answers: make it safe, cross the river first, tell better stories, and name where each person lands before they have to ask.

The failures aren't technical

70%

of automation projects fail outright

Eyvonne Sharp, "Beyond the Code"

80–90%

of those failures are organizational, not technical

Eyvonne Sharp, "Beyond the Code"

75 / 95

Wharton finds 75% of AI efforts succeed; MIT finds 95% see zero return. Both land on the same cause — organizational readiness, not technology.

cited by Greg Freeman, Lumen

£400M

The London Stock Exchange's Taurus project was abandoned after £400M. A team member's verdict: "it was always a political and a business problem."

cautionary case study

Networks demand ~100% accuracy

99.9% still means false assumptions, failed pipelines, and career-limiting incidents. A non-deterministic model pointed at a deterministic-critical system is a liability, not a shortcut.

So the question is architecture

Not "how autonomous can we make this?" but "how do we put a human's judgment at the center — and place our people where that judgment compounds?"

“The industry won't upskill 70% of us. It's going to replace them with agents. The 30% build the agents.”

— the stark framing that landed hard in the room (paraphrasing John Capobianco)

This work is the other answer — redeployment, not replacement.

The SME is the scarce resource. Not the model.

No one knows your network, your team, and your business workflows like you do.

“The first word is still network. Your protocol knowledge is the asset.”

— Dinesh Dutt

“Leading is a different job — it's not engineer++.”

— Eyvonne Sharp

“Human at the helm — not just in the loop. Set the frame before AI enters.”

— Barbara Roos

Two ways to build the very same workflow

AI-first, human-absent

  • ·Point the model at the network with broad tool access
  • ·"Find it and fix it" — opaque and non-deterministic
  • ·No model of THIS network's normal or its known quirks
  • ·No audit trail; one hallucinated action equals an outage
  • ·Nobody trusts it, so nobody uses it

Human-driven

  • ·Human owns the intent and the guardrails
  • ·AI proposes — with evidence you can check
  • ·A human gate before anything touches the network
  • ·Deterministic execution the SME defined
  • ·Trust compounds; oversight relaxes as it is earned

The human owns the ends. Automation runs the middle.

Intent & guardrails
AI proposes (with evidence)
Human gate / validation
Deterministic execution
Feedback tightens the loop
Trust ladder Human-in-the-loop Human-on-the-loop Autonomous (earn it)

"Non-deterministic AI triggers deterministic, human-made workflows." — Greg Freeman

"Start read-only. Earn the trust." — John Capobianco

The front of the loop is the helm — set the frame before AI enters. Real oversight means tracing the reasoning, not just the recommendation.

Maps to the NAF Network Automation Framework — the human owns Intent & Presentation; automation runs Executor, Collector & Orchestrator.

AI enhances the human. Pick your on-ramp.

The blast radius widens at every step — so your grip on the helm tightens, it doesn't loosen. Walk them in order; each one earns the next.

1

Enhanced Knowledge

Unblock yourself

Scope: you Trust: low

The moment: You hit your own skill wall. The task needs Python, an API, or a protocol you don't know — so the project stalls, or you quietly avoid it.

What works — human at the helm

  • ·You bring the judgment; AI supplies the syntax and scaffolding
  • ·Ground it in YOUR world — your configs, your docs, your intent
  • ·Start read-only — documentation, diffs, "what changed?"
  • ·You stay the author — you must be able to read and defend it

Real proof

  • ·Andy Lapteff — 20 years avoiding automation, then one netmiko script became "fireworks," and a restarted career.
  • ·John Capobianco's Packet Buddy — talk to a pcap in natural language, 90% of the way there without memorizing Wireshark filters.

AI-first gets this wrong

Letting the model write what you can't read. You ship code you can't debug or defend — and you learn nothing. The skill gap becomes a dependency.

Where it leaves the human Blocked engineer / CLI lifer Self-sufficient author who ships
2

Enhanced Productivity

Scale the boring

Scope: your team Trust: medium

The moment: Your week is toil. The same low-hanging tasks, the same tickets, the same reports — and you never reach the work that matters.

What works — human at the helm

  • ·Pick the low-hanging, high-impact workflow first
  • ·Write the intent down BEFORE the code
  • ·Codify SME judgment into deterministic, reusable workflows
  • ·AI triggers and composes them — it doesn't improvise them
  • ·Your data is the ceiling; the source of truth is a human asset

Real proof

  • ·Lumen — 350+ human-built workflows running ~60M tasks a month, triggered by non-deterministic AI.
  • ·Freeman's toil curve — automation that isn't reusable just becomes another job.

AI-first gets this wrong

"Vibing" a one-off script per problem. No intent doc, no reuse, no owner — you've traded toil for maintenance debt.

Where it leaves the human Config / ops engineer Intent & workflow architect
3

Enhanced Capability

Solve it live

Scope: your network Trust: high

The moment: 3am. It's down. You're groggy, staring at a blank CLI, and the clock is the enemy.

What works — human at the helm

  • ·Engineers author the triage playbooks and runbooks
  • ·A deterministic workflow pre-stages the evidence before you wake
  • ·AI correlates and ranks root cause — grounded in YOUR docs
  • ·It shows its reasoning, not just its answer (auditability)
  • ·A human approves remediation — in-the-loop before on-the-loop

Real proof

  • ·Indiana University Global NOC "Network Troubleshooter" (Joshua McNamer) — engineer-written playbooks plus a RAG "Knockbot" that recommends while humans approve.
  • ·Lumen's fiber-cut agent pinpointed the GPS coordinate — and called the contractor who'd cut it.

AI-first gets this wrong

Handing an agent write access and saying "find it and fix it." A hallucinated remediation at 3am doesn't prevent the outage — it IS the outage.

Where it leaves the human NOC / tier-1 engineer Knowledge curator & agent supervisor

Same workflow, two architectures

Incident triage & troubleshooting

An alarm fires overnight. Get to root cause fast — without waking a groggy engineer into a blank CLI.

✕ AI-first, human-absent

  • ·An agent with write access auto-diagnoses and remediates
  • ·No sense of your network's normal, or its known quirks
  • ·A hallucinated "fix" at 3am becomes the outage
  • ·Engineers can't see why it did what it did

✓ Human-driven

  • ·Engineers author the playbooks and runbooks
  • ·A deterministic workflow pre-stages triage before you wake
  • ·RAG grounded in your real vendor docs recommends root cause
  • ·A human approves any remediation — in-the-loop first

Real proof: Indiana University Global NOC "Network Troubleshooter" (Joshua McNamer) — engineer-authored playbooks plus a RAG "Knockbot" that recommends, with humans approving remediation.

Where the disrupted IC lands NOC / tier-1 engineer → playbook & runbook author, knowledge curator

Config & change generation

Generate and deploy a service or config change — an EVPN service, a golden config.

✕ AI-first, human-absent

  • ·"Generate the config for X" then push it
  • ·No schema, no dry-run, no digital twin
  • ·99.9% right isn't good enough on a change window
  • ·Drift and surprise become someone's 2am

✓ Human-driven

  • ·SME owns intent, golden templates, and the validation suite
  • ·Executor is multi-protocol, dry-runnable, idempotent
  • ·AI renders and self-checks against the SME's model
  • ·A review gate before execute; scope of autonomy is earned

Real proof: Lumen — 350+ deterministic, human-built workflows that non-deterministic AI now triggers (Greg Freeman). NVIDIA — schema-driven pipelines with structured outputs and validation loops (Senad Palislamovic).

Where the disrupted IC lands Senior config engineer → intent / workflow architect

Source of truth & data quality

Keep an accurate model of the network and reconcile reality against intent.

✕ AI-first, human-absent

  • ·Let AI "discover" state by scraping devices
  • ·Treat the model's output as truth
  • ·Garbage in, confident garbage out
  • ·No timestamp, no way to ask "what changed?"

✓ Human-driven

  • ·SME defines what's authoritative and the data model
  • ·Observability is timestamped and queryable — operational truth
  • ·AI reconciles discovered vs. intended, flags drift and "what changed"
  • ·Human adjudicates; docs stay living, not stale

Real proof: Capobianco's living documentation — diff this hour's JSON against last hour's to answer "what changed?" And Freeman's rule, "you're only as good as your data."

Where the disrupted IC lands CLI expert / doc owner → data steward, source-of-truth owner

Lumen inverted the pyramid — and moved people up, not out

1

Retrained tier 1/2/3 into automation engineers, meeting people where they were, with psychological safety

2

Codified SME judgment into 350+ workflows that non-deterministic AI now triggers

3

Reached ~80% machine-to-machine operations, with humans moved to higher-value work — not off the org chart

“That's the codified wisdom of human experience. Go back and be the wisdom that drives the future.”

— Greg Freeman, Lumen

Where the disrupted ICs go — the new org chart

This is redeployment, not replacement. These are higher-value SME roles, not consolation prizes.

Blocked engineer / CLI lifer Self-sufficient author who ships

Path 1 — their reach now clears their old skill wall

Config / ops engineer Intent & workflow architect

Path 2 — they own "what good looks like" and the guardrails

NOC / tier-1 engineer Knowledge curator & agent supervisor

Path 3 — their pattern memory becomes the system's context

The resistance is identity — not laziness

  • ·Negativity bias weights the bad roughly 10:1
  • ·"Coding is bad, avoid bad" — a 20-year-old story
  • ·We avoid the threat, not the Python
  • ·Until you tell yourself a different story, you can't learn

Andy Lapteff spent 20 years anti-automation, then wrote one netmiko script and called it "fireworks." The wildebeest cross the crocodile river only after the first one goes. Be an early crosser. Make it safe. Tell better stories — and don't wait for the layoff.

How you actually lead human-driven AI

1

Say the obvious thing out loud

Name what the AI is and isn't doing — no mystique.

2

Write the intent down

Your canonical doc IS your guardrails and your source of truth.

3

Demystify "escalation"

That's just human-in-the-loop with a nicer name.

4

Be the calming force

When you walk in, the team should exhale — curiosity over confidence.

5

Measure value, not motion

"Worth the effort and expense" is the only success test.

You don't buy AI adoption. You lead it.

  1. 1

    Pick ONE read-only workflow — documentation or triage. Start where the blast radius is zero.

  2. 2

    Write the intent and guardrails down before any model touches anything.

  3. 3

    Put a human gate in. Measure whether the value was worth the effort.

  4. 4

    Tighten the loop; move up the trust ladder only as it's earned.

  5. 5

    Name where each disrupted IC lands — before they have to ask.

It's not whether AI runs your network. It's who's driving.

"Be the wisdom that drives the future." — Greg Freeman"Do it afraid." — Eyvonne Sharp

Push back on this.

Ask anything about Human Driven AI — the evidence, the architecture, where it breaks down, or how it applies to your team. Answers come from the full research behind the talk.

Whose work this builds on

The synthesis and framing here are Ryan Booth's. The research, numbers, phrases and case studies belong to the practitioners below — cited so you can go read the originals.

Eyvonne Sharp

"Beyond the Code"

Leadership as a different job; the organizational failure data; "worth the effort and expense" as the success test.

Greg Freeman · Lumen

"The NetDevOps Journey"

The inverted pyramid, deterministic workflows triggered by AI, the toil curve, and "be the wisdom that drives the future."

John Capobianco

"From CLI to GPT" / Packet Buddy

The read-only-first trust progression, living documentation via JSON diffs, and the fear framing this work answers.

Andy Lapteff

"Confessions of a CLI Lifer"

The psychology of resistance and the wildebeest river-crossing.

Joshua McNamer · Indiana University Global NOC

Network Troubleshooter

A shipping, human-driven triage system — engineer-authored playbooks with a RAG assistant that recommends while humans approve.

Dinesh Dutt

NAF Network Automation Framework

The six building blocks; "the first word is network"; automation isn't pixie dust.

Barbara Roos

"When Human-in-the-Loop Isn't Enough"

The "human at the helm" vs. "human in the loop" distinction and the auditability test.

Senad Palislamovic · NVIDIA

Schema-driven pipelines with structured outputs and validation loops.