AI-Ready Network Engineer · Verify

Fundamentals with an AI lens: why LLMs invent commands, trust-but-verify, AI config audit

Why an AI can write a config that looks perfect and is wrong, the verify-first habit, how to audit an AI answer line by line, how to read show output as evidence, and six hands-on AI Co-pilot labs where the AI makes real mistakes that you must find, fix and explain.

55 min read7 chapters6 labs15 quiz7 scenarios15 interview Q&A

This first module is free: read the lesson and take the quiz. Labs need an account: sign up free to run up to 3 hands-on labs.

Log inStart free trial
This reading material cannot be printed. Read it on config.nwkings.com.
01

Why AI-written configs fail: trust, but verify

What you will learn

  • How a language model produces text, and why that is different from knowing your network.
  • Why networking is an especially hard case for AI: vendor and version differences, many small numbers that must match, and configs that fail silently.
  • The one habit this whole module builds: trust, but verify.
  • What the six hands-on labs in this module will teach you.

Prerequisites: CCNA-level routing and switching (OSPF, VLANs, ACLs, NAT basics), comfort with the IOS command line and show commands. The earlier module arn-start is recommended, because it explains what AI is good and bad at in general. Here we go deeper into the bad part.

Think of a fast, fluent junior engineer

Imagine a new junior who has read thousands of configs and manuals. They type fast, they never get tired, and they always sound sure of themselves. But they have never logged into your devices, they do not know your topology, and they will never say "I am not sure". You would still use that junior, because the speed is real. You would also read every line before it went near production. An AI model is that junior. It is a fast, fluent, unreliable junior who must be checked.

How a language model produces text

Build the idea in four small steps.

  1. It learned patterns from a huge amount of text. That text included manuals, forum posts, blogs and example configs from many vendors.
  2. It predicts the most likely next words. When you ask for an OSPF config, it writes the words that usually follow such a request. It does not run a calculation and it does not open a lab.
  3. It does not see your devices. It cannot run show ip ospf neighbor. It has no live link to R1 or R2.
  4. It has no memory of your topology unless you provide it. If you do not say that the ISP is AS 64500, it will fill the gap with something that looks normal.

Two consequences matter. First, fluent is not correct: smooth, well-formatted text is what the model is built to produce, whether or not it is right. Second, confidence in tone is not evidence. A wrong answer and a right answer are often written in the same calm voice. Only the device, the output and the documentation can be evidence.

Why networking is a hard case

  • Vendor and version differences. IOS, NX-OS, Junos and others use different syntax for the same idea. A command that is valid on one platform or version may be invalid on another.
  • Many small numbers that must match. An OSPF area, a BGP AS number, a VLAN id, a wildcard mask, a subnet. One digit off on one side and the feature does not work, yet every line looks tidy.
  • Silent failures. The router accepts the config without any error and then does the wrong thing. Valid syntax is not proof of a correct design.

The one habit: trust, but verify

Use the AI for speed. Do not use it as the final authority. Before a config reaches the network, check it against your own spec, and after it is applied, check the live state with show commands. Your job moves from typing the config to proving the config.

Yourprompt Modelpredictstext Config thatlooksright YOUverify Network Nothing in the middle checks your real network. The verify step is yours.

The AI step produces text. Only your verification step connects that text to the real network.

Worked example. You ask for OSPF on the link between R1 and R2. The AI returns a tidy config. On R1 the link is in area 0. On R2 the link is written as network 10.0.12.0 0.0.0.3 area 1. The router accepts it with no error. The text looks complete and professional. But the two routers disagree on the area, so they never become neighbours. Reading the config line by line against your spec catches it. So does one command: show ip ospf neighbor shows an empty table. The config looked right; the evidence said otherwise.

Map of the six labs

In each lab you work on a host called NOC. You run ai task to read the request and ai generate to get the AI answer. Then ai apply all pastes it into the devices, ai faults shows how many mistakes remain, ai hint F1 gives a logged nudge, and ai report F1 <text> files your finding. The answers are fixed, so you can repeat each lab freely.

LabWhat the AI gets wrongWhat it teaches you
arn-ai-ospfWrong area on one side; wildcard 0.0.0.0 so the LAN is never advertisedCheck both ends of a link and read the network statement
arn-ai-vlanTrunk allows only VLAN 10; wrong dot1Q id on a subinterfaceVerify trunks and VLAN ids against the spec
arn-ai-aclPermit above the deny; ACL applied in the wrong directionOrder and direction matter
arn-ai-bgpWrong remote AS; mask that does not match the connected prefixCheck the session state and the exact prefix
arn-ai-natInside and outside swapped; ACL for the wrong subnetRead NAT roles and match the real LAN
arn-ai-inventedAn invented command, Junos syntax on IOS, and a silent holeReject, translate, and test what is really open

Common mistake. Treating "no errors shown" as success. When you paste a config and the router prints nothing, it only means the syntax was accepted. It says nothing about whether the design is correct.

Exam trap. If a question asks how to confirm an AI-generated config works, the answer is never "the AI said so" or "the device accepted it". The answer is a verification step on the live device, such as a show command that proves the intended behaviour.

The config that looked perfect

A junior engineer asked an AI to bring up OSPF between two branch routers. The answer was well formatted and had comments on every block. The engineer pasted it during a maintenance window and saw no errors, so the change was closed. Next morning the branch had no route to head office. A senior engineer ran show ip ospf neighbor on both routers and found no neighbours. One side had the link in area 1, the other in area 0. The fix was a one-line change, but the lost hours came from believing a config because it looked finished.

Lesson: accepted syntax and a confident tone are not proof. Verify the live state before you close the change.

"How do you use AI safely when you configure network devices?"

Strong answer: I use it as a fast drafter, not as the authority. I give it the exact topology and requirements, I read every line against my spec, I check commands against the vendor reference, and after applying I verify with show commands. I treat it like a fluent junior whose work I must review before it reaches production.

Key takeaways

  • A language model predicts likely text from patterns; it does not run your network or see your devices.
  • Fluent is not correct, and a confident tone is not evidence.
  • Networking is hard for AI because of vendor and version differences, small numbers that must match, and silent failures.
  • Accepted syntax is not proof of a correct design; only live verification is.
  • The habit of this module is trust, but verify: use AI for speed, and prove the result yourself.

Upgrade to keep reading

The free plan lets you read chapter 1 of this module. Upgrade to open the remaining chapters.

  1. Five ways AI gets networks wrong
  2. The verify loop: six steps for every AI answer
  3. Audit an AI config line by line
  4. Show output as evidence: accepted is not correct
  5. Prompting to reduce errors
  6. Summary and exam checklist

Join the community