The Three-Lane Router: a local vs cloud AI checklist for every job

This is the working sheet behind the video: the three lanes, the five questions that put a job in a lane, the rules to write down and the log to keep. It takes an hour to fill in for a typical office, and it does not require you to own any hardware yet.

The three lanes

1. Lane one: plain rules. If a rule like "invoice older than thirty days" does the job, you do not need a model at all.
2. Lane two: the local model. Routine work that repeats every night, like flagging stalled repair orders or writing the morning briefing, on hardware you control. Local is not secure by itself: every agent still gets limited access and a log.
3. Lane three: the cloud model. The rare, messy job, like a strange customer email, goes to a more capable model, and a person approves anything that commits you.

The lanes run in that order. A job only moves to the next lane when the one before it cannot do the work.

This is the pattern Anthropic calls routing in its guide to building agents: classify an input and direct it to a specialized follow-up task, sending easy, common questions to smaller, cost-efficient models and hard or unusual ones to more capable models. The sheet below turns that into a decision you make once per job, on paper.

Who this is for

Owners and office managers at service businesses who are being told to "run AI locally" or "just use the cloud" and want a reason for each job. If you have not decided which jobs a local model can handle at all, run the Fifty Example Test on one job first. For the background on what a hybrid setup is, see our hybrid AI system guide.

Step 1, list the jobs and not the models

Write down every AI job you want, in one sentence each, grouped by when it happens:

• Every night: check open repair orders or jobs for anything that has not moved; compare completed jobs to invoices.
• Every morning: a one-page briefing for the owner.
• Every time something odd shows up: a customer email with three problems in it, an estimate that fits no template.

Step 2, answer the five questions for each job

Copy this table. Answer each question for each job, then count which column wins.

Question	Leans local (lane two)	Leans cloud (lane three)
1. How often does it run?	Many times a day, same shape	Rarely, different every time
2. How hard is the reasoning?	Compare, match, flag, summarize a known format	Read something messy and make a judgment
3. How sensitive is the data?	Customer names, notes and invoices you would rather not send out every hour	Little or no customer detail needed
4. What happens if it is wrong?	A person reviews every output anyway	A person reviews, and a better model saves rework
5. What if the internet drops?	It must still run	It can wait

Before you use the table, ask the lane one question: could a plain rule do this? "Repair order with no status change in two days" is a rule. "Invoice older than thirty days" is a rule. Rules are cheaper than either model and easier to check, so they go first.

Step 3, write the routing rules down

A vague intention is impossible to check. A written rule is easy to audit. Use this format, one line per job:

Job	Lane	Data it may read	Who approves the output	Logged?
Stalled repair order check	1 (rule)	Open ROs, status dates	Service manager reviews the list	Yes
Morning briefing	2 (local)	Yesterday's ROs, payments, estimates	Owner reads it	Yes
Odd customer email reply	3 (cloud)	That email and the customer's job history	Office manager approves before sending	Yes

The rows are an example. Then add the house rules that apply to every lane, for example:

• Customer messages never go to a cloud model without a person's approval.
• Payment details never go to any model.
• No agent sends, discounts or spends money on its own.

Step 4, lock down lane two like any other system

Running a model on your own machine keeps the prompts there. The Ollama FAQ says that when you run locally, it does not see your prompts or data, and it documents a setting to turn its cloud features off. That is the data benefit.

It is not a security program. A box with a shared password and no logs is worse than a well set up cloud service. For each local agent, write down which systems it reads, give it its own credentials, and log every run. Cortex agents each get defined access to defined systems, with sensitive actions requiring approval and activity recorded.

Step 5, keep a log of every run

Three columns are enough: which lane and model ran, on what input, and who approved the output. If a job ran in the wrong lane, the log shows it. If you cannot see it in the log, treat it as if it should not have happened.

Step 6, review the lanes every quarter

Models change. A job that needed the cloud last year may run fine locally now, and the reverse. NIST's AI Risk Management Framework, which is intended for voluntary use and organized around govern, map, measure and manage, works as a quarterly agenda: who owns each job, which lane it is in, whether it ran that way, and what you are changing.

A real example

Detail Driven, an automotive group in Jackson Hole, runs four kinds of shop on Tekmetric, CCC ONE, GoHighLevel and QuickBooks. Cortex installed dedicated local AI hardware with hybrid local and cloud inference: routine watches and the daily briefing run locally, and harder reasoning goes to a stronger cloud model when a job needs it. Tekmetric and CCC ONE stay the systems of record. We have not published hours or dollar figures for it, and this sheet does not claim any.

What to watch out for

• Buying hardware before you know the jobs. Start in the cloud if you have nothing local, and move jobs to lane two when volume or sensitivity justifies it.
• Skipping lane one. A model doing a rule's job costs more and is harder to check.
• Everything in lane three out of habit. You send customer detail out every hour for work a smaller model could do.
• Unwritten rules. If the routing lives in one person's head, nobody can audit it.

How to tell it worked

Every AI job has a lane, a written rule, named data, a named approver and a log. You can open the log and see yesterday's runs by lane. Our article on local versus cloud AI for service businesses covers the trade-offs in more depth, and sorting your own jobs is part of the free AI Operations Audit.