I Tried Onboarding Instinct, Then Asked About Phone Calls
I asked Instinct whether it could make phone calls for me. Its answer was direct:
“No, I can't place phone calls for you on this account.”
That reply came on October 1, 2026, after we had already considered phone work as a possible useful role for it. We got ahead of the evidence. We should have checked the actual account before suggesting the job.
This is my account of one onboarding attempt and one account capability inquiry. We did not place a test call, run a controlled model benchmark or establish what every Instinct customer can do.
One report received. Reading limits disclosed. Some claims contradicted by current public pages.
My assistant said outbound calling was unavailable on my account.
No call placed. Call quality, completion and cost remain untested.
What I wanted Instinct to do
If you own a successful service business and use several AI subscriptions, you can end up doing a new kind of management work. You carry the context between assistants. You explain the business again. You reconcile their answers. Then you copy the result into the place where the team actually works.
I want models to work together from durable shared knowledge and finish useful work with less owner coordination. Another assistant needs to earn a role by helping with that goal.
My assignment to Instinct was based on our real operation: read the published operating guide, identify the sources actually read, reconcile the roles, distinguish tested tools from untested ones and propose a bounded useful role.
The starting sources were our public agent onboarding page and How the Agents Run the Company. These are instructions a worker can inspect. They explain ownership, handoffs and what evidence is needed before calling a job finished.
We were checking whether this assistant could understand the setup and find a useful place in it. We had not established an installed operating role, a recurring job or a completed handoff.
Instinct disclosed what it had not read
Instinct's report said it fully read Start Here, the guide root and eight named guide pages. It said it skimmed other routes and the workshop companion PDF. It also disclosed that it could not access private documents and older PDFs.
That transparency deserves credit. A link is not proof of a successful read. A partial read should be described as partial, and private material should stay behind its access boundary. The onboarding assignment was not permission to breach inaccessible files.
The report also said the underlying model and reasoning settings were unknown. We should preserve that limit rather than invent a model name to explain the result.
These are its reported reading receipts, not an independently recorded trace of every page fetch. They help us understand the report's scope. They do not establish that it received the current revision of every source.
Some conclusions did not match the current pages
The report said Start Here lacked a version stamp. It also described an older ownership arrangement involving Claude rather than the current Happy and dot responsibilities.
A fresh public check found Start Here marked Documentation v1.0.1, checked 1 October 2026. It names Happy as fleet coordinator and dot as the documentation and QA owner for the onboarding page and company guide. The guide's public teaching edition is also v1.0.1, dated October 1, and names the same roles.
Those current pages contradict the report's description. The evidence points to stale fetched content or a source-verification problem. We do not know enough to identify the precise fetching failure, and we have no evidence of deliberate deception.
The useful correction is to identify which revision the worker actually received and compare its conclusion with that text. Repeating a source URL doesn't establish that the current page was read.
Instinct also confused an earlier client session with the Office Hours session we intended. The intended Office Hours took place later on October 1. We do not need to publish a private recording or identify the client to explain the problem: a worker needs to distinguish the source requested from a different session it found.
That error matters because the next job may depend on the meeting's decisions. It still does not justify turning one onboarding mistake into a verdict on the entire product.
Approval and a completed handoff can coexist
Instinct treated the requirement for approval before sending as me as a conflict with our instruction that I should not be the paste layer.
Those requirements can coexist. An agent can prepare the complete result, identify the intended destination, obtain a reserved approval when required, deliver through an authorized channel and check that the result arrived.
Making me copy a report into another tool is a separate problem. It transfers routine delivery work to the owner. The goal is to reduce that work while keeping consequential decisions with the person authorized to make them.
Instinct proposed an independent reviewer role, but its proposal still escalated source conflicts to me. Some questions really do require an owner decision. A mismatch between a report and the current public instructions should first be investigated against the sources and routed to the responsible documentation owner where a supported handoff exists.
We had a proposed role. We did not have evidence of a demonstrated handoff or a useful completed reviewer outcome.
Another part of the report proposed corrections based partly on page dates. Age can be a reason to inspect an instruction. It does not establish that the instruction is defective. The proposed correction needs a content conflict, a failed result or another checkable reason.
Why we checked phone capability next
Phone work looked like a possible reason to consider Instinct separately from the assistants already in our setup. That possibility needed an account check.
Instinct's public product page describes an assistant users can text or call and says it can proactively call users. Those statements describe interaction with the user; they do not establish every outbound business-calling workflow.
The company's September 28 funding announcement separately describes Instinct Concierge handling tasks that include phone calls. The announcement reports a $1 billion Series C from Sequoia, Benchmark and Coatue at a $10 billion valuation. The amount raised and the valuation are different numbers. Neither proves that a capability is enabled on my account or completes the work I need.
Before giving Instinct a phone job, we asked whether it could place calls for me. The reply was:
“No, I can't place phone calls for you on this account.”
It also said its available tools and product notes lacked outbound calling. That is an account-specific self-report from the assistant. We have not independently inspected its internal tool configuration or obtained a team confirmation of the account's entitlements.
The practical result was enough to stop us assigning it a phone task on the assumption that it could perform one. It is not enough to say outbound calling is unavailable to everyone or that the company's announced Concierge capability does not exist.
What remains unverified
We did not test an incoming call, callback, forwarding, caller identity, call records or call allowances. Exact fees remain unverified. We also did not establish whether a particular Concierge task is performed entirely by AI or with human assistance.
Instinct's public terms of service contemplate paid features. They do not supply a verified price for the phone job we were considering. I cannot give you a cost comparison from this experience.
And because no call was placed, we have no call-quality score, success rate, completion time, recording or business outcome to report. The assistant's answer tells us what it said it could do on this account. A functional trial would be a different piece of evidence.
The decision we changed
Our first thought was to give Instinct a bounded reviewer trial. We tightened that recommendation after examining the report. General onboarding should be parked until a distinctive useful capability gives us a reason to continue.
We then considered phone work too quickly. The account inquiry corrected that assumption. It would be misleading to leave the original recommendation standing after the assistant said it could not place calls for me.
Our A-Players Only project-lead standard judges useful outcomes and the supervision needed to get them. One mistake is not automatic disqualification. But a polished role proposal is not a completed outcome, and a worker that keeps returning routine coordination to the owner has not solved the management problem.
For now, Instinct has not earned an additional operating role in our setup. That is a decision about this attempt, this account and the work we need. It is not a claim of broad corporate untrustworthiness, and it does not mean we cancelled or removed the product.
I would reconsider when a needed capability works on the actual account, the economics make sense and the result can be checked without making me coordinate every step. We have not established those conditions here.
How this helps you choose your next assistant
Before adding another AI subscription, name the work you're hiring it to do. Use the operating instructions your business actually follows. Check which revision the assistant received and which tools your account actually has.
Then give it a bounded real job and inspect the finished result in the system that owns the work. Count the owner time needed to explain, correct and deliver it. An assistant can write a useful report and still need more supervision than the role can justify.
You can begin with an assistant you already have. The larger goal is models working from shared knowledge, with one accountable owner for each outcome and evidence the next worker can read. Every additional tool should help that operation work better.
Experience reported: October 1, 2026. Public source checks: October 2, 2026 UTC. Scope: one onboarding report and one account capability inquiry; no functional phone trial.