The ten ticket types that fill the inbox, and the three among them the agent may finish without asking anybody.
afka uses analytics cookies to understand how visitors use the site. No ads, no data sales.
The working manual for setting a client team up with the afka support agent: an agent that resolves tickets end to end across chat, email, voice, SMS and WhatsApp, configured for the whole team, with a human rhythm the team keeps after you leave.
A company whose support volume has outgrown the team but not yet justified a support organization. The signs are easy to hear in one call: the founder answers tickets in the evening, the queue grows on weekends, and the same ten questions consume most of the day.
Your first working artifact is a table with four rows: who talks to the agent, and where. Draw it with the client in ten minutes, because every later setting hangs off it.
The owner talks to the agent in Slack and by voice, and holds the autonomy dial.
The support lead works inside the help desk the team already uses and owns the escalation line.
Every team member can write to the agent, or dictate to it, in the afka composer: configuring the agent IS talking to it, not filling forms.
The meeting agent joins the weekly support call on Zoom, Google Meet or Microsoft Teams and turns what was said into finished tasks.
Three roles, agreed out loud in the first session. The approver receives every gated action in the approvals inbox in the web app and as a card in Slack, and a decision resumes the work in under two seconds. The teacher is whoever may say remember this for next time and accept what the agent learned. The audit reader opens the log each week and sees who did what, when, and at what cost. One person can hold all three roles; a team of eight should not let that happen.
One support agent inside the client workspace: connected to the help desk and the order system, answering in the brand voice, taking real actions such as order status, refunds within policy and account changes, with the support skill pack filled in from your interview, spending caps enforced in code, and every action in the audit log.
An afka skill pack has two layers, and the difference is the whole safety story. Guidance is plain language: the tone, the refund philosophy, the phrases the brand never uses. Guardrails are code: the refund cap in dollars, the categories that always wait for a person, the disclosures that must appear. A guardrail beats any instruction, from anyone, every time, and when the two layers disagree the more conservative action wins. Your interview fills both layers: it gathers exactly these answers from the client, and they go straight into the pack:
The ten ticket types that fill the inbox, and the three among them the agent may finish without asking anybody.
The refund cap per ticket and per day, in numbers the client says out loud.
The sentences that must never be said to a customer, and the disclosure that must always appear.
The named human for a customer angry beyond a template, and the context that should arrive with the escalation.
Connections run through managed connectors: a popup opens, the client authorizes their own tool, and afka never receives a password. The support set is the help desk, the inbox and the order system, from more than three thousand supported apps. The connectors screen shows every connection's status, a broken one says reconnect-needed and one click repairs it, and disconnecting is revocation: it takes effect within one task cycle.
Sit with the team for one hour and list the ten most frequent ticket types with their real volume. This list decides what the agent handles first, what stays with a person, and what your skill-pack interview digs into.
A new workspace picks the support function and produces a first finished artifact from seeded sample data in under a minute. Show it before any tool is connected: the team sees the whole loop, plan, action, record, before anything real is at stake.
Start on Review: the agent acts, and refunds, sends and account changes wait for approval. Agree the spending caps out loud, for example 100 USD per refund and 500 USD per day, write them into the guardrail layer, and show the team that the caps are enforced in code rather than in the wording the agent is given.
Run one real ticket end to end in front of the whole team: the agent reads the history, answers in the brand voice, takes the action, and the audit log records every step. The moment the support lead checks the answer against policy and finds nothing to fix is the moment the deployment becomes theirs.
Walk the approver through the approvals inbox, the audit reader through the log, and the teacher through the skill library. Leave the weekly checklist behind you and put the first weekly call in the calendar with the meeting agent invited. Your handover artifact is the rhythm the team keeps without you.
The deployment holds because the rhythm is small enough to keep.
Daily, five minutes: the approver sweeps the approvals inbox. An empty inbox is information too: it means the spending caps are set right.
Weekly, on the support call: the meeting agent attends, decisions become tasks, the audit reader walks the log, and the team agrees what graduates from Review to Auto. Trust widens the allow-list; the audit never weakens.
Weekly, ten minutes: the teacher reviews the skill library, accepts what the agent learned from corrections, and archives what went stale.
Every first session produces the same three questions. Give the real answers, because they are better than reassurance.
On Review, nothing customer-facing leaves without a person approving it. On Auto, the agent acts only inside the allow-list and caps the team set, and every action is in the audit log with who, what, when and cost, so a mistake is found in minutes, not discovered in a quarterly review.
In about a quarter of a second, and that number is measured, not promised. Pause is enforced at five separate chokepoints, and disconnecting a tool revokes access within one task cycle.
The workspace is isolated per business at the database layer, skills the agent learns never leave the workspace, and credentials are injected at the moment of a call: the model never holds them.
The three situations you will actually meet: a connector shows reconnect-needed, which one click repairs; the client wants everything stopped, which is the pause button and it is real; and a ticket needs a human, which is the escalation line with full context attached, the named person from your interview. None of the three is an incident. All three are the system working.
You earn 30 percent of the client subscription for twelve months from the first paid invoice, with first-touch attribution for 90 days. Your implementation fee is your own business and stays yours. afka publishes no income projections: the terms are the terms.
No. A new workspace picks the support function and gets a first finished artifact from seeded sample data in under a minute. Connectors then widen what the agent can reach.
Hard spending caps enforced in code: per refund, per day, and per count. A cap that can be talked around is not a cap, so the spending caps do not live in the wording the agent is given.
The first session produces the first resolved ticket. The rhythm, one short daily sweep and one weekly review, is set in the same session, and the week-one graduation to Auto for the safest ticket types is a decision the team makes on the first weekly call.
The program terms, the agreement and your link are on the deployer page.
The deployer program