If someone wiped the board and said “build AI ops from zero, you have 30 days,” you would not start with a platform bake-off. Start with one leak, then the smallest stack that can close it without inventing a company inside a company.
The stack you want is boring on purpose. Boring ships. Boring can be operated by a tired person on a Monday. Boring is allowed to get fancy after it makes money or saves hours you can count.
This is what you would actually assemble today, knowing what you know, and knowing how many “AI operating systems” Teams often find that gather dust.
First, pick the job, not the logo
My default job, because Keep seeing it unpaid: after-hours capture to a human callback that can book. You might pick ticket triage or internal admin. Fine. Pick one.
Then refuse a second channel until the first one writes into a system of record. Chat, voice, SMS, web widget: one front door. The rest is how demos multiply.
If you cannot name the leak in a sentence, you are shopping. Shopping feels like work. It is not.
The actual pile of tools
Aim for five pieces that talk to each other:
- A front door. Phone number or chat, with a provider that can stream audio or messages and give me transcripts.
- An LLM API I can swap. Don’t marry a wrapper so tight that changing models is a rewrite.
- A place truth lives. CRM, even if it’s an unglamorous one, or a database the team already opens.
- Logs of every tool call and model call, with redaction.
- A morning queue. Could be the CRM inbox. Could be a tiny portal. Must be something humans already will open.
That’s the stack. Not twenty SaaS logos in a slide.
Around that Add the adult clothes: secret store, staging vs production keys, a status check, backups if you own the data. None of that is AI. All of that is why the AI is allowed to exist.
Models: start with one strong, one cheap
you would not build a custom router in week one. you would pick a strong model for anything customer-visible or decision-shaped, and a cheap model for extraction and drafts. Hard-code the split in config.
Local models can wait unless you have a data constraint that actually forbids the API. Local is great. Local is also a second job. From zero, you want a vendor I can bill and a contract I can read.
Evals start ugly: a folder of ten real transcripts and a checklist. Did it capture the number. Did it escalate crisis. Did it invent insurance. Run that before you change a prompt. You can buy eval software later when the folder hurts.
Knowledge: one source, short files
you would not boil the ocean into a vector store on day one. you would write three documents:
- what we do and don’t do
- the intake script and forbidden claims
- location or product facts in a table
If the docs are too long to review, they are too long to trust. Retrieval is a later problem. Most early agents fail because the facts were never written down, not because cosine similarity wasn’t tuned.
Voice and chat should share those docs so they don’t contradict. One brain. We’ve covered that. Still true from zero.
Integration beats originality
you will use the calendar the team already has, the phone system they already pay for if it can forward, and the CRM they already lie to themselves in. Replacing those in month one is how AI projects become IT projects that forget the caller.
A thin tool layer in my own code is the one original piece you want. Validate arguments. Inject secrets. Write the result back. The model never holds the keys. The CRM remains the record.
If the CRM cannot take a note and an owner, you fix that before I hire a more poetic model.
What to would skip on purpose
- a custom vector database until the docs don’t fit in a prompt
- a multi-agent org chart
- a public chatbot on the marketing site as the first win
- training a model
- building my own telephony stack
- an internal copilot that answers “what’s our vacation policy” while the phones still bleed
Skip is a strategy. From zero, every extra component is a thing that can fail at 2am when you thought you were done.
The 30-day shape
Week 1: map the leak, write the fence, get a number forwarding.
Week 2: smallest agent, human review, notes in the CRM.
Week 3: morning queue discipline, measure time-to-callback, fix the two worst prompt failures.
Week 4: one booking path or a scheduled callback the caller agreed to. Kill anything that doesn’t serve that.
If that month works, then you earned a second location overlay, a second channel, a real eval set, maybe a portal that isn’t ugly. If it doesn’t work, you learned on a small pile, not on a platform purchase.
Bottom line
From zero, AI ops is not a cathedral. It’s a phone, a brain with a fence, a log, and a human who owns the next step. Start there. The rest of the market will still be there when you have a reason to care.
