Vendor pages for AI phone agents are written entirely in the affirmative. Ours included, mostly. So here is the other list — what Osokoro Phone will not do, why, and which of those are decisions rather than gaps.
The distinction matters more than the list. A gap gets filled next quarter. A decision does not, and if one of them rules us out, you should find that out now rather than three weeks into a beta.
It answers. It does not dial.
The agent is inbound only. Its instructions literally begin "You are the inbound support and sales representative for" your business, and there is no outbound path anywhere in the product: no dialer, no campaign mode, no callback queue, no "the agent will ring them back when the part arrives".
This is the limit that disqualifies the most people, so it is first. If you are costing out a sales team making calls, or a clinic confirming tomorrow's appointments, none of the rest applies.
Calls end at five minutes
The default maximum call length is five minutes. The hard ceiling — the most any configuration can be set to — is ten. A timer issues a real hangup; the call does not trail off or degrade, it ends.
For a booking line, an opening-hours line, a "do you have this in stock" line, five minutes is generous. Most such calls are ninety seconds. For anything where the caller has to describe a problem before you can help them, five minutes is a transfer, not a resolution, and you should plan the line around that.
The cap exists for two reasons that happen to agree. A realtime speech model bills by the minute, so an unbounded call is an unbounded cost. And a conversation that has run ten minutes without either resolving or transferring is not going to.
There are no transcripts
Nothing a caller says is stored. No recording, no transcript, no caller number.
What survives a call is a content-free journal row: whether it connected, how long it lasted, how many turns, how many searches, which pricing version applied, and the estimated cost. Those rows are deleted after thirty days. Calls that never complete are marked failed after twenty minutes.
Two consequences, and they point in opposite directions.
You cannot go back and read what somebody said last Tuesday. There is no quality-assurance review, no after-call summary in your inbox, no searchable history of what customers keep asking about. If mining call content is central to how you run the line, this is the wrong product, and it is the wrong product by design rather than by schedule.
The other direction: there is nothing to leak, nothing to subpoena, and nothing to include in a breach notification. For a lot of small businesses that trade is worth making. It should be a decision you make, though, not one you discover.
It cannot choose who to transfer to
The agent can hand a caller to a person. It cannot decide which person. The transfer destination is a single number in server configuration, set by an owner or admin, and the model is told to use it — never to pick one.
That reads like a missing feature and it is a security property. A model that can choose a forwarding number is a model that can be talked into choosing a different forwarding number by a caller who wants it to. There is a well-known shape of attack here: persuade the system to route a call somewhere, then answer that call yourself. Removing the choice removes the attack.
So there is one destination. If you need different transfers for sales and support, you need two numbers and two agents, not one clever one.
It searches, up to a point
The agent has exactly three tools: search over documents you gave it, web search when you explicitly enable it, and transfer to a person. That is the whole surface.
Search is capped at twenty completed searches in a single call, after which the agent hangs up. The ceiling is a cost control and a circuit breaker at the same time — a model on its twentieth search in one phone call is not converging on an answer, and letting it continue costs money to produce a worse call.
The integrations do not exist
None of these are built:
- CRM, calendar, Shopify, Stripe, Zendesk, or any other system of record
- Looking up a customer's account, order, booking, or balance
- Identity verification, one-time passcodes, or any authorisation step
- Taking a payment
- A shared inbox where your team sees calls the agent handled
So the agent knows what you wrote down for it and what it can find on the web. It does not know who is calling, what they bought, or when their appointment is. An agent that says "let me look up your order" would be lying, and it is not configured to.
That bounds the useful cases fairly sharply: questions with the same answer for everyone. Hours, location, directions, prices, what you do and do not sell, whether you serve a particular area, and routing the rest to a person.
It is not safe for regulated data
Business knowledge lives in collections hosted by xAI, whose model runs the call. Zero data retention is arranged per xAI team rather than per API key.
We are not going to tell you that is fine for protected health information, and you should be suspicious of any vendor at this stage who does. If you are in healthcare, finance, or anywhere with a statutory duty over what callers tell you, the answer today is no.
Recording disclosure is your call, and your obligation
There is a toggle for a recording disclosure, configured per business, and the agent speaks whatever you write in it.
Consent law for recorded calls varies by state and several states require every party to consent. We give you the toggle and the field. We do not give you legal advice, and the disclosure being configurable means getting it right is yours. Ask someone qualified what your state requires before the line goes live.
What it does do
Having spent nine sections on limits, the positive case is narrow and real: an inbound line that answers on the first ring, at any hour, with the same answers every time, on a number you already own, at a per-minute cost you can compute in advance. It speaks English, Spanish and Portuguese naturally, and you can hint at a language or let it detect one.
For a business whose phone rings with the same eight questions and where the ninth needs a person, that is most of the job. For anything that needs to know who is calling, it is not close.
The status, plainly
Osokoro Phone is a private beta. Access is approved by a person, one business at a time. Number provisioning is done by an operator rather than by you in a dashboard, the published rate is proposed rather than a billing commitment, and production metering is still gated.
If the limits above are survivable for your line, that is worth a conversation. If one of them is not, I would rather you closed this tab than found out in week three.