Build an AI workflow with clear inputs, checks and approvals

Step 1

Decide which steps need AI

Not every step deserves a model. Separate the steps that follow a rule from the ones that need interpretation, and put a person where a mistake would cost something. Trigger, router, reasoner, tool and evaluator are one useful way to name the steps. They are a design vocabulary, not a standard.

Where the work falls

A framing used on this page, not an industry standard. It asks two questions: how hard is it to move the data, and how much judgment does each step need?

• Simple movement · Fixed rule

  Rule-based integration platforms are usually enough.

• Messy screens · Fixed rule

  UI automation (RPA) can drive legacy interfaces that have no API.

• Simple movement · Judgment needed

  One AI step inside an assistant or a single tool is often enough.

• Many systems · Judgment at several steps

  A multi-step AI workflow with checks. This is what the rest of this page covers.

A workflow runs the same steps each time, while an agent chooses some of its own at run time. For how ZeroTwo runs and supervises agents, see agent execution and supervision. If you are still choosing between kinds of tool, start with automation tool categories compared.

Step 2

Pick one process to start

Automating five things at once hides which one is failing. Choose a process that meets most of these tests, and design that one end to end first.

Step 3 · Conceptual design

Worked example: inbound lead to a reviewed reply draft

The workflow turns a sales email into an enriched record and a reply draft, and ends with a person. For a template built around the research half of this flow, see lead research and enrichment.

This is a design, not a recorded run. No outputs, timings or model comparisons are claimed. The email and schema beside it are invented for illustration. Which model suits which step is for you to test on your own examples; this page does not rank them.

1. Step 1 · TriggerRule

   A message reaches the shared sales inbox

   The workflow starts on a new message that passes a simple filter: not from your own domain, not an auto-reply.

   Fails when

   The same message is delivered twice.

   Control

   Store the message ID and skip IDs already processed.

2. Step 2 · ReasonerAI

   Extract the fields into a fixed schema

   A model reads the email and returns company, contact, request type, team size and requested date in the schema shown beside this list.

   Fails when

   The model invents a value the email never stated.

   Control

   Require the word unknown for absent fields, validate the output against the schema and keep the source email next to it.

3. Step 3 · ToolLookup

   Look the company up in your CRM

   A plain read from the system you already use, matched on the sender's domain.

   Fails when

   No match, or several.

   Control

   Treat no match and several matches as their own branches that go to a person.

4. Step 4 · RouterRule

   Branch on completeness and request type

   Complete and a known type leads to a draft. Missing details lead to a draft that asks a question. Pricing exceptions, legal and security questions skip drafting and go to a person.

   Fails when

   A request type nobody planned for falls through.

   Control

   Make the default branch send to a person.

5. Step 5 · ReasonerAI

   Draft a reply from approved facts only

   A model writes a reply using a short set of facts you supply, such as your product description and price list, and says plainly when something is not covered.

   Fails when

   The draft promises a feature, date or number that is not in the approved facts.

   Control

   Supply the approved facts as the only source and tell the model to leave gaps visible instead of filling them.

6. Step 6 · EvaluatorCheck

   Check the draft before a person sees it

   Rules run first: required elements present, no figures or dates outside the approved facts, no links outside an allowlist. A second model or rubric review can be added on top.

   Fails when

   A check that never fails.

   Control

   Test it against drafts you wrote to be wrong.

7. Step 7 · Human gatePerson

   A person reviews, edits, sends or rejects

   The reviewer sees the source email, the extracted fields and the results of the checks. The workflow ends at a reviewed draft. Nothing is sent automatically.

   Fails when

   Drafts wait with nobody reading them.

   Control

   Name the owner and a time after which the item escalates.

{
  "company": "string | unknown",
  "contact": "string | unknown",
  "request_type": "pricing | demo | support | other",
  "team_size": "number | unknown",
  "needs_by": "date | unknown",
  "missing": ["string"]
}

Step 4

Build it in ZeroTwo

This page does not include a recorded run of the example. Below is what ZeroTwo documents for recurring, supervised work, followed by what to confirm in your own account before you rely on it.

Step 5

Validate success and failure paths

Define what correct means before you turn anything on: which fields must be right, which may be unknown, and what must never appear in a draft. Then try to break it.

Build a test set you can review by hand

Include typical inputs, plus one of each failure above: missing data, a duplicate, hostile text and an unavailable connector. Size the set to what you can read yourself. There is no number that is enough for every process, so add cases until new ones stop surprising you.

Step 6

Measure cost and useful work

Run a bounded pilot and keep a log. The number to track is cost per accepted run: everything the workflow cost in the period, divided by the runs a person accepted.

Context from two primary sources

of agentic AI projects predicted to be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls.

40%+

Gartner press release, 25 June 2025 (a forecast), read 5 October 2026

AI high performers say they fundamentally redesigned workflows because of AI, against about one in four other respondents. This is an association in a survey, not proof of cause.

3 in 4

McKinsey, The state of AI in 2026 (survey of 1,719 respondents), read 5 October 2026

respondents say their organization is limiting AI use because of operating costs, including tokens.

1 in 5

McKinsey, The state of AI in 2026, read 5 October 2026

Our view, not a finding: starting small and tracking cost per accepted run is how you avoid the cancellation pattern in the first figure. Neither source tests that.

Step 7

Operate it safely over time

A workflow that passes its pilot still drifts. These six habits keep it trustworthy after launch.

Questions

Frequently asked questions