Who maintains the AI system after it goes live?

Three parties share the work after an AI system goes live, and the split does not move: we keep the platform and the models under it running, you own the client and every change they ask for, and the client owns the facts about their own business. Nobody hands the thing off and walks away. A voice agent is not a website that sits there for four years. Prices change, staff change, and the model your agent runs on has a retirement date published by the vendor.

This is the part of the sale most agency owners have not thought through when they quote a monthly. So here is what month two onward actually looks like.

Who owns what once the system is live

The client owns their own truth. Hours, service list, price bands, who gets the leads, what the holiday schedule is. When any of that changes and nobody tells you, the agent confidently says the old thing. That is the single most common way these systems go wrong, and it is not a technical failure.

You own the relationship. We never contact anyone who opens your demo or uses your client's system, and we never will, which is what makes the white label real. It also means every change request lands on you first. If you are selling this as a hands-off recurring revenue stream where nobody ever calls you, adjust your expectations now. See what your prospects actually see when you sell white-label AI for where the seams are.

We own the platform. The voice and chat infrastructure, the agent runtime, the model migrations, the demo site, the client portal. There is no API key for you to rotate and no phone number for you to buy, so there is also nothing on that layer for you to maintain.

What actually changes in the first ninety days

Four things, roughly in order of how often they happen.

The client's own information drifts. A new service gets added, a price goes up, the Saturday hours change, one of the two people who used to get the leads quits. Assume this starts happening sooner than the client will think to tell you.

The script needs tuning against real calls. Not because it was built wrong, but because real callers ask things nobody predicted during intake. The first few dozen live calls teach you more about the prompt than the whole onboarding form did.

The model underneath gets deprecated. More on that below, because it is the piece that surprises people.

The messaging rails need attention. Registration, consent language, and carrier policy all have maintenance attached.

Your agent's model has a retirement date, and it is published

This is the part that separates running a system from building one.

OpenAI publishes a deprecations page with a stated notice policy: at least six months before retiring a generally available model, at least three months for specialized variants, and for anything with "preview" in the name, possibly as little as two weeks. Their own guidance is blunt about that last tier, saying they do not recommend preview models for business-critical production work unless you can migrate on short notice.

Those are not hypothetical dates. On July 20, 2026, OpenAI announced the retirement of its legacy audio and realtime models, including gpt-realtime, gpt-audio, gpt-4o-realtime and gpt-4o-mini-realtime, all shutting down January 20, 2027, with gpt-realtime mapping to gpt-realtime-2.1. A month later, on August 26, 2026, the transcription models went the same way, with whisper-1 and the gpt-4o-transcribe family scheduled to shut down February 26, 2027.

Anthropic runs a similar calendar with a shorter floor. Its model deprecations page commits to at least 60 days of notice before retiring a publicly released model, and the recent history matches: Claude Opus 4.1 was deprecated on June 5, 2026 and retired on August 5, 2026. Requests to a retired model fail. They do not degrade gracefully.

Read that as an operator, not as a developer. If you assembled your own stack, every one of those dates is a migration you schedule, test, and eat the cost of, on somebody else's calendar, for every client you have. Miss one and a plumber's phone stops answering on a Tuesday. That is the maintenance bill nobody includes when they price a DIY build, and it is a large part of why we ran the numbers the way we did in build your own AI stack, or use a white-label one.

On our side, the migration is ours. Your client's agent keeps answering and you find out about it in a changelog rather than a support ticket.

The texting side needs upkeep you cannot see

Missed-call text-back is usually the feature that sells the deal, and it is also the one with paperwork behind it. US carriers route application-to-person SMS through a registration system called A2P 10DLC, and Twilio's documentation describes what that means in practice: a brand registration identifying the business, a campaign registration describing how people opt in, opt out and get help, and carrier fees for sending unregistered traffic if you skip it.

The registration is not a one-time form you forget. A few things from Twilio's business information requirements that bite small local businesses in particular:

The registry checks the business website. A URL that is parked, times out, or sits behind a login gets rejected. So does a privacy policy that is not publicly reachable or that fails to state that mobile numbers are not shared.

Newly issued EINs take 30 to 90 days to propagate across the validation databases, so a business that incorporated last month may simply fail registration until it does not.

Volume is capped by brand type. A sole proprietor brand gets one campaign and roughly 1,000 SMS segments a day to T-Mobile, which is fine for a two-van plumbing company and not fine for a 40-seat call operation.

And a suspended campaign does not just pause. Twilio's campaign troubleshooting guide says that after 30 days of suspension, the campaign is deleted and its phone numbers are released.

None of this is dramatic. It is also not the kind of thing your client will ever think about, which is exactly why it belongs in the price you charge.

What a change request actually looks like

Say the client adds a new service in October and wants it quoted correctly on the phone.

You get the detail, in their words, with the price band. You update it the same way the information was collected at signature, through the onboarding path that is already attached to that build. Then you call the agent yourself and ask about the new service the way a customer would. Then you text the owner that it is live.

That last step is not decoration. It is the entire retention mechanism. A client who hears from you twice a quarter with something small and specific does not shop the line item. A client who only hears from you when the card declines does.

For how the same handoff works before go-live, what happens after your client signs an AI buildout walks the intake side of it.

What we do not promise

We publish no support SLA, no guaranteed turnaround for change requests, and no 24-hour emergency line. I would rather say that plainly than have you promise a two-hour response to a client and then be the one holding it.

So do not sell what you cannot control. Tell your client that changes happen within a few business days and that they should send them to you, not to a portal they expect to be instant. If they need same-day script changes at 9pm on a Sunday, they need an in-house person, not an agency retainer at your price point.

Two more honest limits. The usage allowances on the $49 plan are sized for your demo site, not for a client's production phone line, and the pricing page states them in full so you can do that arithmetic before you quote. And if a client wants the agent wired deep into an unusual internal system that nobody else in their trade runs, this is not the product for that. Say so early. A clean no in month one costs you a deal. A messy yes costs you a reputation in a town where every contractor knows every other contractor.

What keeps a client past month three

Retention on these systems is mostly about whether anyone ever showed the owner what the thing did. Calls answered after hours. Leads that came in at 11pm and got a reply. The Saturday the agent booked four jobs while the shop was closed.

Pull those numbers once a month and send four sentences. Not a dashboard link, because they will not open it. Four sentences and one number. That habit is worth more than any feature you could add.

FAQ

Who fixes it when the agent says something wrong? You catch it or the client reports it to you, and the change goes through the same path the original build did. Wrong answers are almost always stale business information or a gap in the script rather than a broken model, so the fix is usually content, not engineering.

Do I have to migrate anything when OpenAI retires a model? Not on our platform. We handle model changes on the infrastructure we run. If you built your own stack, those dates are yours to manage, and OpenAI's deprecations page is where you would track them.

How often do clients actually ask for changes? We do not publish numbers on this, so treat what follows as planning advice rather than data. The early requests are almost always hours, pricing, or a new service, and they tend to thin out once the first round of corrections is in. Budget your time around a short call once a month rather than a constant stream.

Can my client edit their own information? They have a branded client portal for status and assets. Treat script and pricing changes as something that comes through you, because a client editing their own prompt at midnight is how a good system starts giving bad answers.

What happens to a client's system if I cancel my Lumina subscription? Published demos unpublish and your work stays as drafts, per the FAQ. Do not put yourself in a position where cancelling mid-contract leaves a client without the thing they are paying for.

Your next step

Take your oldest live client and check three things this week: their current hours, their current price bands, and whether the person listed as the lead recipient still works there. If any of the three is wrong, you just found the reason their results dipped.

If you are not running one yet, the $49 plan gets you the demo site, the client portal, fulfillment access and onboarding help, and the site says that price rises on October 1. Start here.