The board asked for "something with AI" next quarter. You have three vendors on the calendar, no numbers in hand and an operation that can't stop for experiments. The risk isn't the AI going wrong — it's spending 30 days on demos and ending up not knowing whether it's worth it.

This plan is for the operations director who wants an answer in a month: start small, measure for real and decide on executions and cost, not slides.

Week 1: the trial with your own key and a process with real pain

The Zihin Trial lasts 15 days, asks for no card and includes 100K complimentary Zihin tokens, and you can connect your own AI provider key (BYOK). If the complimentary tokens run out before the period ends, an agent with a fixed model and no fallback chain keeps running on your key; when the 15 days end, you need to choose a plan. In practice, that changes the conversation with finance: usage on your key is billed by the provider, with no Zihin markup, and the platform is a separate cost — zero during the test.

The Trial includes 1 agent, 1 admin and 1 editor/member, 100K tokens per cycle, 24-hour sessions and access to Zihin Chat, the app the team uses to talk to the agent. It's small on purpose: it fits one process. Choose it well.

Three criteria for choosing the process:

  • The pain is measurable today. Someone already complains about it in numbers: hours per week, a ticket queue, a missed deadline.
  • The answer depends on data that lives in a system. Orders, tickets, customer records, contracts. If the information is "in Carlos's head", it's not the first case.
  • Getting it wrong is cheap. Start with lookups and triage, not irreversible actions.

Warning: Don't start with the process most visible to the board. Start with the one that has an owner, data and pain. An agent that answers "what's the status of order 4,512?" in seconds is more convincing than an ambitious pilot that never reaches the end of the trial.

Week 2: one data source, read-only

The agent is only useful when it reaches the system where the work happens. But in week 2 there's a single rule: connect one source, and only for reading.

In Zihin, the agent uses the tools you make available — REST endpoints configured as tools with authentication and defined parameters, or MCP servers. An integration can be read-only, and writing is never automatic: the agent only does what the exposed tool allows. For ERPs and databases, the production path is Zihin Replay, which keeps a read-only copy in your network, queried through the MCP Bridge — the agent never receives database credentials. That feature is part of the Business plan, so in the Trial the source is usually an API of the system or an MCP server your team already exposes. See the paths at /integrations.

Credentials live in the platform's encrypted vault (AES-256) and are resolved by the authorized integrations, not pasted into the prompt.

What to deliver by the end of the week: an agent that answers real questions about the chosen process, querying the system, for 2 or 3 people in the operation.

Week 3: who approves what

Before expanding, define the limits. Two mechanisms address most of an operations director's concerns:

Human approval. You define the scope of actions that require approval, and the step stays pending until the responsible person decides — right in the chat. One user can approve the action triggered by another; responsibility sits between people, not just between a person and an agent.

A spending cap per agent. A dollar limit per agent. When it hits the cap, the agent stops on its own, even when running on your key. That takes the question "what if it burns everything overnight?" off the table.

Add to that the four-level access control — owner, admin, editor and member — enforced in both the UI and the API. The difference between the two most used roles is simple:

  • Editor creates and edits agents, including flows, connections and configuration.
  • Member uses ready-made agents (chat, WhatsApp, Slack) and doesn't see configuration, prompts or other people's conversations.

Both count as plan users. In the Trial you have one editor/member seat; this is the time to decide who will be the agent's technical owner and who will be the operations user.

Tip: Write the approval policy in one sentence before configuring it: "every reply that mentions an amount above X goes through management". If the sentence doesn't fit on one line, the process isn't clear enough for an agent yet.

Week 4: measure executions and cost, then choose the plan

Time to look at the numbers. The console shows the account's consumption; each execution's session can be replayed with phases, tool calls and per-call latency. With that, three questions close the pilot:

  1. How many executions did the process generate in two weeks, and how many needed human approval?
  2. How much did it cost in tokens on the provider's bill, and what does that represent per execution?
  3. How much time did the operation stop spending on the chosen process, in the metric defined in week 1?

With the answers, choosing a plan stops being a guess. The prices published at /plans:

Core Pro Business
Price $89/month $269/month On request
Agents 5 20 Unlimited
Admins · editors/members 2 · 10 5 · 50 Unlimited
Included tokens 5M 15M 15M
Triggers Webhook, email and scheduling + multichannel (WhatsApp, Slack, Teams) + on-premise MCP Bridge, Tunnel, SSO, SLA
Sessions 7 days 14 days 30 days

A quick guide:

  • Core if the pilot became one or a few processes, triggered by webhook, email or schedule, with up to 10 people using it. Includes RBAC, credential vault, basic telemetry and onboarding with virtual training.
  • Pro if you need to put the agent on WhatsApp, Slack or Teams, or want Zihin Watch — weekly summaries and monthly reports on executions, costs and points of attention by email. It also brings a managed token pool on top of your own key.
  • Business if the data lives in an ERP or database behind the firewall (Replay, MCP Bridge and Tunnel), if SSO and audit logs are required, or if the contract calls for an SLA and dedicated support. Pricing on request.

If you're torn between two, upgrade later: an upgrade takes effect immediately, with the prorated difference charged at the switch, self-service plans have no lock-in, and the published add-ons cover gradual growth — a Booster of 1 million tokens for $10 per pack, an extra admin for $15/month and a pack of 10 editors/members for $10/month.

What not to do in the first 30 days

  • Don't connect three systems at once. Each integration requires checking compatibility, authentication and scope. One at a time.
  • Don't start with writes. Irreversible actions come in after the approval process has been tested on reads.
  • Don't expect "automatic learning". The agent follows configurations and instructions defined by the team; memory keeps context between sessions, but that isn't training and doesn't guarantee identical answers. Review the results.
  • Don't treat the trial as a demo. Treat it as week 1 of the operation. The same roles, limits and approvals you configure now are the ones that apply in production.

By the end of the month you'll have an agent in a real process, a cost per execution, a written approval policy and a plan chosen by usage — not by promise. If you'd rather do the journey with support, Zihin structures onboarding in four phases: diagnostic, plan, execution and ongoing support. Questions about the trial, keys and roles are in the FAQ.

Book a diagnostic.

Originally published in Portuguese on the Zihin blog. I work at Zihin, which connects AI agents to a company's systems and processes, with permissions, human approvals and cost tracking.