Skip to content
ConsultEvo

June 2026 Zappy Award Winners: Lessons in Shared Context

Zapier named Eric McNulty of Mercari, Blair Mishleau of Patreon, and Amber Sharp of Tivly as its June 2026 Zappy Award monthly winners. Their reported examples point to a shared operating principle: make company context available where work happens, give AI a bounded job, and keep a person accountable for exceptions.

This is an implementation-oriented interpretation of those examples, not a reproduction of the winners’ systems. Reported outcomes below come from Zapier-published material and are not independent audits. The control patterns and schemas are proposed adaptations unless explicitly attributed to a published workflow.

The practical question is not simply whether AI can answer or classify. It is whether the surrounding workflow supplies the right context, rejects unsafe outputs, records what happened, and gives a named owner a useful fallback.

Trigger and context AI job Validation and fallback Destination
Support conversation with queue, channel, message, and policy Classify and summarize eligible cases Check restricted topics, category, confidence, and placeholders Configured helpdesk action or CX review
Employee question and permitted internal sources Retrieve and summarize relevant guidance Check access, freshness, conflicts, and evidence Slack answer or knowledge-gap task
Account history and period-specific onboarding metrics Translate verified facts into a handoff draft Check identity, periods, units, and missing fields Review task or approved CRM record

Mercari: route support cases through rules before asking AI to classify

Zapier’s Mercari customer story says the company handles about 120,000 support tickets monthly. It reports that roughly 47,000 are resolved entirely through Zapier, first replies go out on about 97% of tickets, CSAT rose by one point on a five-point scale, and more than 3,000 agent hours are freed each month. These are customer-story figures attributed to Mercari through Zapier, not independently audited benchmarks.

Zapier describes a roughly 90-step master Zap that uses Google Sheets for prompts, category definitions, policy logic, and response guidance. Customer message content is scrubbed for personally identifiable information, and low-confidence cases go to a human. A Google Apps Script interface allows nontechnical teammates to run the system. The full production workflow, sheet schema, category taxonomy, and sub-workflows are not public.

Zapier also publishes a guarded support-orchestration template inspired by Mercari. It uses Dixa, Google Sheets, AI by Zapier, and Code by Zapier. It is a partial template shell, not Mercari’s production workflow: implementers must supply category-specific Sub-Zaps, field mappings, policies, thresholds, and destinations. Its masking covers common email and phone patterns, not every sensitive identifier.

A reliable adaptation separates deterministic eligibility from probabilistic classification. A restricted queue, payment dispute, account-access issue, or regulated topic should be blocked before AI is asked to classify it. AI can then classify and summarize eligible cases, while structured validation decides whether an approved action is possible.

01Receive the conversationA documented template pattern starts with a new Dixa conversation. Retrieve the stable conversation ID, message, queue, channel, priority, and status, then verify the available trigger fields in the connected account.
02Apply policy before AIExclude restricted queues and topics first. Match exactly one active policy row, record its version, and minimize or mask sensitive text before sending eligible content to AI.
03Return bounded fieldsAsk AI by Zapier for an approved category, subcategory, confidence value, short summary, recommended action, and human-review flag. Named output fields are a documented capability, but the schema here is a proposed contract.
04Validate and routeReject unsupported categories, malformed confidence values, missing policy versions, and unresolved reply placeholders. A valid eligible result can follow a configured Dixa action; failed, restricted, and low-confidence cases go to the designated CX queue.

The public template supports routing through configured paths and category-specific sub-workflows, including reply, tagging, closing, transfer, and escalation patterns. Treat confidence as a routing signal, not proof of correctness: test thresholds against human-reviewed cases and track false acceptance as well as escalation volume. Extend masking for addresses, order numbers, account identifiers, and other fields present in your data.

Patreon: make internal knowledge findable without forcing an answer

Zapier reports that Blair Mishleau built “‘Treon,” a Slack-based assistant that searches internal resources such as help articles, Figma files, go-to-market briefs, documentation, and a glossary. The article says answers were grounded in source material, conversations were logged to identify knowledge gaps, employees could flag inaccuracies, and resources were regularly reviewed. Zapier reports more than 1,300 employee questions answered since launch.

The reviewed sources do not independently document the assistant’s retrieval architecture, connectors, permission enforcement, visible citation format, or question-count method. The safe conclusion is therefore about the reported pattern, not a technical specification.

A proposed internal-assistant flow is: receive an employee question, retrieve permitted source passages, check access and freshness, generate an answer with source identifiers when the evidence supports it, and otherwise create a knowledge-gap task. If sources conflict or a source is inaccessible to the requester, the assistant should return no authoritative answer rather than expose restricted information or produce plausible prose. Teams designing this kind of bounded system can review AI agent design with defined responsibilities and human oversight.

A no-answer result is a control

When evidence is missing, conflicting, stale, or unauthorized, return answerable=false, record the reason, and route the question to the relevant knowledge owner. Trust improves when the assistant has a clear stopping condition.

Tivly: turn sales history into a reviewable client handoff

Zapier describes a three-stage idea at Tivly: triage inbound messages, route new-client questions to a chatbot using organizational knowledge, and assemble account history and onboarding insights into a handoff scorecard that translates sales shorthand into plain language. Zapier reports more than 200 onboarding intelligence reviews in six months, onboarding volume up by more than 165%, and transition preparation falling from about 25 hours monthly to under one hour. The source does not define the growth denominator, comparison period, or time-study method, so these figures are not comparable benchmarks.

The reviewed sources do not document Tivly’s exact source applications, CRM destination, field mappings, five-step handoff agent, or Claude data controls. The following is a hypothetical handoff design, not a reconstruction of Tivly’s system.

  1. Collect verified inputs: use a stable account ID, reporting-period start and end, campaign metrics with units and timestamps, leads purchased, conversion definitions, and approved recent notes. Keep source applications as systems of record.
  2. Generate a review draft: ask AI to summarize only supplied facts, identify risks and missing fields, and suggest next actions. A proposed typed result can include account_id, period_start, period_end, summary, risks, missing_fields, recommended_next_actions, and source_run_id.
  3. Validate the period: check account identity, date boundaries, units, source timestamps, and required values. If a weekly campaign total is combined with a different conversion period, create a review task instead of a definitive scorecard.
  4. Assign approval: save a draft or task to the verified destination. The incoming account owner approves consequential conclusions; sales or onboarding operations corrects source-data problems.

For client-period records, an illustrative key is account_id + period_start + period_end + workflow_version. It identifies the declared weekly or period review, not a person, CRM contact, aggregate metric, or automation run. Preserve a separate zap_id + run_id for each execution and a separate source event identifier for the handoff trigger.

Controls that make shared context dependable

Use a small output contract that downstream steps can test instead of asking AI for free-form prose alone. A proposed cross-workflow contract may include an approved category, confidence in a declared range, recommended action, requires_human, policy_version, source_ids, and missing_information. Structured fields make validation possible, but they do not guarantee correct values.

Treat Google Sheets policy content as controlled configuration: use stable headers in the first row, an owner, active status, version, effective dates, and a review date. If zero or multiple active rows match a policy key, fail closed and route the case to the configuration owner. Test mappings after structural changes, and keep policy rows separate from ticket events and automation logs.

Keep records at their natural grain. A ticket event, generated message, AI observation, citation, account-period review, and monthly aggregate are different rows. Illustrative keys are:

  • Ticket event: source_system + ticket_id + event_id.
  • Generated message: source_system + ticket_id + message_id.
  • Automation run: zap_id + run_id.
  • AI observation: run_id + model_provider + prompt_version.
  • Citation: answer_id + citation_sequence or answer_id + canonical_source_id + source_locator.
  • Account-period review: account_id + period_start + period_end + workflow_version.
  • Aggregate metric: metric_name + source_system + aggregation_start + aggregation_end + definition_version.

These are proposed implementation keys, not vendor-published schemas. They prevent multiple citations, prompt versions, replays, and reporting periods from overwriting one another. Store aggregate metrics separately from raw events, and retain the definition version for measures such as “resolved,” “first reply,” or “onboarding volume.”

Zapier documents trigger deduplication for certain records within a Zap, but that is not end-to-end idempotency. A Tables lookup followed by create can race when concurrent runs both find no record. For consequential records, use a destination-enforced unique key or a transactional upsert where available. If neither capability is verified, describe the write as duplicate-prone, add reconciliation, and retain the run ID and source event ID for investigation.

Assign exception owners by workflow: CX operations for policy and configuration problems, the support queue for case review, a knowledge owner for unsupported answers, and the incoming account owner for handoff approval. When an outbound or consequential action must wait, configure an explicit approval step. Zapier’s Human in the Loop documentation describes pausing a Zap for approval, rejection, or data changes; it is not automatic review of every downstream action.

Before launch, verify
  • Source access is approved for both the automation and the requesting employee.
  • Policy rows have an owner, version, effective dates, and a tested no-match route.
  • AI outputs pass allowed-value, type, range, and required-field checks.
  • Every exception has a named fallback owner and a defined approval gate where needed.
  • Concurrent writes use destination-enforced uniqueness or a transactional upsert.
  • Event, run, citation, account-period, and aggregate records use separate grains.

For teams formalizing policy, routing, and exception ownership, Zapier automation design and implementation is a relevant service area.

What the June award timing means now

The June article listed July 24, 2026, as the annual submission deadline. The current official 2026 rules instead state that the annual contest ended July 31, 2026, and that monthly-award submissions may continue through December 31, 2026, subject to modification or discontinuation. Use the official rules for current eligibility and timing.