2026 Zappy Award Winners: What They Built and How to Adapt the Workflows

The 2026 Zappy Award winners are Ethan Schwandt of Jobber; Carlos Robledo of Hims & Hers and Doug Hamilton of Klaviyo; Ignacio Piñeiro of Galgo; Justin Hallman of Youtech; Jocelyne Mendez-Guzman of BioRender; Eric McNulty of Mercari; Nicholas Francoeur of Compucom; Ariel Chen of Figma; and Belinda Jimenez of Program Manager.

The examples cover recruiting, enablement, delivery-evidence review, phone-revenue attribution, sales follow-up, support operations, equipment records, background-check monitoring, and project management. This guide answers who won, then turns the most documented examples into adaptable workflow designs with explicit fields, validation gates, event identity, and human ownership.

Zapier’s award overview and customer stories report results supplied by featured organizations. The reviewed material did not include raw datasets or independent audits, so the figures below are reported outcomes rather than forecasts or guarantees. Some winners have detailed customer stories; others are described only in the award overview.

Who won the 2026 Zappy Awards?

• AI Transformation Leader of the Year: Ethan Schwandt, Jobber.
• AI Rising Stars: Carlos Robledo, Hims & Hers, and Doug Hamilton, Klaviyo.
• Operations AI Builder of the Year: Ignacio Piñeiro, Galgo.
• Marketing AI Builder of the Year: Justin Hallman, Youtech.
• Sales AI Builder of the Year: Jocelyne Mendez-Guzman, BioRender.
• Customer Support AI Builder of the Year: Eric McNulty, Mercari.
• IT AI Builder of the Year: Nicholas Francoeur, Compucom.
• HR AI Builder of the Year: Ariel Chen, Figma.
• AI Builder of the Year: Belinda Jimenez, Program Manager.

This is a category list, not a ranking. The practical value is in the operating patterns, not in assuming that one team’s reported result will transfer directly to another organization.

What the winners’ workflows have in common

The strongest examples follow a repeatable operating shape: identify recurring work, capture a source event or record, use AI for a bounded interpretation task when necessary, validate the result with rules, write to a controlled destination, and send exceptions to a named person or queue.

Workflow	AI’s job	Validation or gate	Action and fallback
Jobber enablement	Support job-specific experimentation and recurring-work discovery.	Leaders and employees identify useful handoffs; details come from the award overview.	Build internal capability rather than relying on one expert.
Hims & Hers recruiting	Rex supports reports and recruiter requests across recruiting data.	Explicit rules identify stale candidates, overdue feedback, offers, and service-level issues.	Send results to Airtable and Slack; verify ATS fields and record IDs.
Klaviyo enablement	Summarize product updates and provide role-tailored guidance.	Adoption and tracker figures are award-overview claims, not a separately reviewed technical case study.	Place guidance where the talent team works.
Galgo evidence review	Evaluate delivery evidence against strict rules and classify exceptions.	Log verdicts and retain human audit; implementation details are award-overview claims.	Review invalid evidence instead of treating every result as final.
Youtech attribution	No AI is required for the core click-to-conversion match.	Require a GCLID, valid conversion action, timestamp, value, and currency where applicable.	Upload an offline conversion to Google Ads; investigate unmatched events.
BioRender follow-up	Extract discussion points, promises, and next steps from a Gong transcript.	Use verified roles, deterministic commercial templates, and representative approval.	Create a Gmail draft and notify the representative.
Mercari support	Classify and summarize using policy and category context.	Scrub sensitive data, check confidence and allowed routes, and exclude prohibited paths.	Route to a configured flow or human queue.
Compucom equipment records	Extract make, model, and serial number from photos.	Check format, site context, and inventory reference data.	Store only validated values; send mismatches to field operations.
Figma screening monitoring	Identify pending screening components and triage urgency.	Compare with Checkr, preserve event history, and retain human adjudication.	Escalate at-risk cases to senior approvers.
Program Manager projects	Assist resource assignment and project monitoring.	Project teams retain oversight; technical details come from the award overview.	Automate setup while keeping project decisions with the team.

Use AI for ambiguous language, classification, summarization, and unstructured extraction. Use deterministic logic for identifiers, dates, currency, exact pricing, eligibility, thresholds, deduplication, and system-of-record writes. AI by Zapier supports configurable output fields that can feed later workflow steps, but current fields, models, permissions, limits, and app actions should be checked in the current documentation.

Teams adapting these patterns may also consider Zapier automation services when the process, owner, and exception path are already defined. The goal is a controlled handoff, not an undefined instruction to let a model run the job.

Five workflow patterns with enough detail to adapt

1. Guarded support orchestration

Mercari’s customer story describes a governed orchestration layer for new Dixa conversations. The documented sequence retrieves ticket details, excludes unsuitable channels, loads policy and category context, scrubs personal information, classifies and summarizes the conversation, and routes it to a specialized path. Low-confidence cases and failures return to human agents.

Zapier also offers a public guarded-support template inspired by this pattern. It uses Dixa, Google Sheets, AI by Zapier, Code by Zapier, and Sub-Zap placeholders. It is an orchestration starting point, not Mercari’s production configuration. Users must provide categories, policies, queue identifiers, helpdesk logic, and the Sub-Zap workflows.

• Trigger and inputs: a new conversation with ticket ID, source event ID, channel, and message text. The support operations owner maintains policy and routing definitions.
• AI output: a proposed category, summary, confidence value, recommended route, and human-review flag. These are illustrative fields, not a Zapier-published schema.
• Gate: require approved categories and routes, numeric confidence within range, current policy context, and successful PII scrubbing. Keep prohibited channels, unsupported issues, missing context, and tool failures out of automated response paths.
• Destination: write the classification and route to the helpdesk or create an internal note, then send exceptions to the designated support queue.

Zapier reports about 120,000 monthly Mercari tickets and roughly 47,000 fully resolved through automation each month. Those are monthly aggregate figures, not a per-ticket accuracy rate.

2. Phone-revenue attribution with click-level identity

Youtech’s customer story describes capturing each paid-search GCLID, storing click-level data, receiving a later quality signal such as a qualified CRM lead, CallRail event, or signed client, matching that event to the original click, and uploading an offline conversion and revenue value to Google Ads.

• Click record: store the GCLID, click timestamp, client identifier, and normalized contact fields where available.
• Conversion event: store a separate source system, stable source event ID, conversion action, conversion timestamp, value, and ISO currency code where relevant. A click record, a conversion event, and a monthly revenue total are different grains.
• Validation: require a GCLID, verify that the Google Ads conversion action uses an Import from Click source, validate timestamp and currency, and apply any required delay after the click.
• Duplicate control: an illustrative event key is client_id + source_system + source_event_id + conversion_action. A phone number, date, or GCLID alone may collapse legitimate multiple funnel stages. If concurrent processing could create costly duplicates, enforce uniqueness in a database and use a transactional upsert. A search-then-create sequence is not race-safe.
• Fallback: keep missing-ID, unmatched, rejected, or not-yet-eligible events in an exception state for marketing operations. Do not invent an attribution link.

The Google Ads action accepts a GCLID, conversion action, timestamp, and optional value and currency. Zapier reports that Youtech attributed $213,000 in previously unconnected phone revenue to paid search. That is a customer-reported attribution result, not an independent audit.

3. AI-assisted sales follow-up as a reviewed draft

BioRender’s documented workflow begins with a Gong webhook and transcript. JavaScript restructures the transcript, AI extracts account details, discussion points, promises, and next steps, and role-based rules determine the appropriate follow-up. Deterministic code assembles exact commercial language, creates a Gmail draft, logs the result in Google Sheets, and notifies the representative in Slack.

• Required identity: retain a stable Gong call ID, participant metadata, call type, and recipient. The call ID should be stored with the generated draft ID and approval state.
• Bounded output: extract what was discussed and promised. Do not let the model invent prices, discounts, contract terms, or commitments.
• Gate: verify participant role and call type, check that the call ID has not already produced a draft, and require representative review before sending.
• Exception: if ownership or commercial terms are ambiguous, mark the draft for review and leave the disputed field blank rather than inferring it.

Zapier reports that the workflow reduced follow-up drafting from roughly 15 to 20 minutes to about three minutes. The published account describes draft creation, not automatic sending. Teams exploring AI agent services should preserve this separation between language assistance and controlled business terms.

4. Equipment-data extraction with reference checks

Compucom’s customer story describes technicians submitting equipment photos, AI extracting make, model, and serial number, and the result being checked against inventory data before storage. A separate integration reads spreadsheet scheduling data and creates Resource Guru bookings. The scheduling bridge is a distinct workflow, not a turnkey extension of image extraction.

• Trigger and inputs: a form submission with submission ID, technician ID, site ID, workstation number, and image reference.
• AI output: make, model, serial number, and an extraction status such as extracted, uncertain, or unreadable. Treat all extracted values as untrusted.
• Validation: check serial-number format, client and site context, and known device combinations in the inventory reference data.
• Write gate: store the values only after the reference checks pass. Unknown devices, blurry images, and multiple possible matches create a field-operations review task.
• Identity: retain the source submission ID or asset-event ID so a replay cannot create a second inventory record.

The customer story reports improvements in survey time and asset-entry errors, but those figures are published customer outcomes rather than independent measurements.

5. Human-led background-check monitoring

Figma’s customer story describes synchronizing Checkr status events into an onboarding view and append-only status log. Daily AI triage identifies the pending screening component and assigns urgency, while at-risk cases go to senior approvers. Human adjudication and a manual fallback remain essential.

• Event record: retain candidate ID, report ID, screening-component ID, status, event timestamp, start date, jurisdiction, and source event ID.
• AI role: identify which screening is pending and prioritize attention. AI must not make the hiring, adjudication, or adverse-action decision.
• Source-of-truth gate: compare the event with Checkr, test the status component by component, and preserve every status change rather than overwriting history.
• Event identity: a proposed key combines report ID, screening-component ID, and event version or timestamp. This is an editorial design recommendation, not a vendor schema.
• Exception: if the view disagrees with Checkr or a generic pending status hides the delayed component, pause automated escalation and use the manual process until reconciled.

Zapier reports that Figma reduced recurring monitoring to a daily scan of under 15 minutes while tracking more than 120 active candidates. The published account does not provide an independent measurement audit.

A practical build sequence: decide the job before adding AI

Start with one recurring handoff and its owner. Define the source of truth and the record or event that should cause the workflow to run before writing a prompt. Then specify the downstream fields and the safe failure path.

For AI outputs, retain provenance such as source system, source record ID, source event ID, received time, workflow version, prompt version, model identifier when available, confidence, review status, and final disposition. This makes a replay distinguishable from a new business event.

What the published results do and do not prove

Zapier’s award article and customer stories report outcomes supplied by featured organizations. The review found no raw datasets, independent audits, or source-system exports that establish causal impact, general accuracy, or repeatability.

For example, Zapier reports roughly 47,000 Mercari tickets fully resolved through automation each month. Youtech reported $213,000 in previously unconnected phone revenue attributed to paid search. Zapier reports Figma’s recurring monitoring fell to a daily scan of under 15 minutes. These figures have different units, denominators, and operating contexts. None should be presented as a forecast for another team.

The same discipline applies to the other published figures. Zapier reports that Jobber’s active builders grew by 122% in the first 30 days, that Galgo reduced invalid evidence from approximately 8% to below 2%, and that BioRender reduced drafting time to about three minutes. These are reported accounts, not independently audited benchmarks.

Awards, prizes, and submission details

The annual 2026 category awards and the monthly Zappy Builder Award are separate programs. Under the official 2026 rules, each annual category winner receives $5,000 and a swag pack valued at approximately $500. Monthly recipients receive $500, a swag pack, and promotion.

Zapier’s awards landing page describes year-round submissions for monthly recognition. The official 2026 rules contain potentially conflicting annual-contest dates: one passage says the contest ended December 31, 2026, while another says entries after July 31, 2026 were not eligible for annual awards. Do not infer a definitive annual deadline. Check the current rules or confirm directly with Zapier.

Choose the pattern that matches your operational bottleneck

• High support volume: investigate guarded classification and routing, with prohibited channels and low-confidence cases sent to agents.
• Phone-driven paid-search revenue: keep click records separate from conversion events and validate Google Ads requirements before upload.
• Repetitive sales follow-up: extract call details into a reviewed draft while keeping exact commercial terms in deterministic templates.
• Messy equipment records: use image extraction followed by format, context, and inventory checks before writing asset data.
• Time-sensitive, high-stakes monitoring: preserve event history, compare against the source system, and leave decisions with a named human approver.
• Recruiting or organization-wide adoption: review the documented Hims & Hers story for ATS reporting and Slack delivery. Treat Klaviyo, Jobber, Galgo, and Program Manager details as award-overview descriptions unless a separately reviewed implementation source is available.

Pick one process, identify its owner and source record, and write down what happens when the workflow cannot safely act. That decision is a better starting point than adding AI to every step.