What this ranking scores
A human-in-the-loop step has two jobs that pull against each other. It has to be easy enough that the right person acts within the hour, and strict enough afterwards to prove who acted. One decision settles both: whether your approver needed an account. The eight platforms here answer it in four different ways.
Moxo's Description of Services defines a Magic Link as a link granting "account-free access to a specific assigned task", and the audit log in the AutomationAtlas test workspace labels that actor "Public link" rather than a name (read 15 September 2026).
Zapier puts both halves of the problem on the same help page. The Request Approval step's reviewer type can be set to "Anyone", which Zapier defines as "members outside of your Zapier account", while the Limitations section below states that "Reviewers must have a Zapier account" and that a reviewer who lacks one "will be prompted to create a new account" (help.zapier.com, read 21 September 2026).
Kissflow turns the same decision into a licence counter. Its portal documentation states that "If there's a shortage of licenses, the invited user will remain inactive and unable to sign in to the portal" (community.kissflow.com, read 21 September 2026).
Pega documents a fourth shape and then prices it. Directed Web Access issues an external assignment where "The link in the external Assignment does not require login, but the user can use it only once", and Pega's licence compliance page books those people as "processing performed by Web users (typically through the Directed Web access feature), classified as Invocation users" rather than as named users (docs.pega.com, read 21 September 2026).
Reaching a person and enrolling one are different acts. A comparison table treats them as the same feature. The currency changes each time, from friction to a seat to the strength of the name in the log, and Pega is the only vendor here that puts a licence class on the difference. This ranking scores that mechanism across five weighted criteria and nothing else.
The category boundary
This page scores the pause and what survives it. It does not score process control depth (retries, incidents, compensation and state management), which is scored on Best Process Orchestration Platforms 2026. Building the agent is out too: the model, the tool catalogue and the reasoning loop are scored on Best AI Agent Platforms 2026.
Camunda 8 shows why the boundary sits there. Its BPMN user task is the most first-class human step in this list and scores accordingly, while the durable engine underneath it, which is the reason most teams shortlist Camunda at all, is scored nowhere on this page.
The five criteria
Each platform was scored from 1 to 10 on five weighted criteria, with the total a weighted average carried to one decimal.
How the human enters the path (25%)
Whether the pause is a step type the platform ships, or an app, node or toggle added to an automation designed to run without stopping. Scored on where the control lives in the builder's own object model, read from vendor documentation or from the builder itself. The line that decides the score is whether the engine can finish the process without a person: an element the engine must wait on scores above a step type the product merely ships, and both score above a switch somebody has to remember to turn on. A 10 is a wait-for-human element that is the workflow engine's own primitive, so the engine cannot execute the process without pausing on it, as a BPMN user task is. A 9 is a named step type the product ships and the engine waits on by definition, such as a Moxo Approval step or a Pega Approve/Reject step. A 4 is a per-tool setting called "Require approval before running" that ships switched off, where the product's default behaviour is to act without asking.
What the approver must have (20%)
The account, licence or seat the deciding person needs before the approve button works, and whether somebody outside your company can act without one. Scored on primary vendor statements about accounts and licences, not on marketing claims about external collaboration. The distinction that sets the score is between reaching a person and enrolling them: a notification reaches any inbox, while a licence gate can leave an invited approver unable to sign in at all. A 9 is a time-limited link scoped to one task with no account behind it. A 3 is an approver who consumes a platform licence from a finite pool. A route that skips the login but still needs an administrator to provision the person scores in the middle.
Decision routes (20%)
What the process can do when the answer is no: send back for revision, let the approver choose the return point, run a designated rejection branch, end the run, reassign, or escalate on a timer. Scored on the named routes a builder can select, counted from vendor documentation or from the builder's own controls, rather than on what could be modelled with enough effort. A 9 offers four or more distinct named outcomes plus a timed escalation path. A 4 offers approve and deny, where deny cancels the action and informs the caller.
Decision record (20%)
What the log stores about a completed decision: whether the actor resolves to an identifiable person, which fields accompany the entry, how far back it can be retrieved, and which plan carries it. Scored on published field lists and retention figures, with plan gating recorded, because an audit log a buyer cannot reach on the plan that carries the approval step is not evidence of anything. The lookalike to separate here is the run log: a run log shows that a step completed, an audit log names who completed it. An 8 publishes the full field list including the actor. A 4 is a stated commitment to auditability with no published fields.
Cost per decision point (15%)
The meter that runs when a person is asked to decide: per flow, per task, per seat, per credit, or nothing extra on a licence you already hold. Scored on published prices only, naming the cheapest plan that carries the approval step, because that is the number a buyer has to model. No published price is scored as no published price rather than as an unknown that might turn out cheap, and a published price a buyer can model scores above silence even when the number is large. A 9 is a supported approval gate at no incremental cost on a licence already in the estate. A 2 is a platform publishing no price for any plan.
On the weighting
The weighting decides the top of this list. How the human enters the path at 25 percent, together with what the approver must have at 20 percent, rewards a platform assembled out of human steps with an account-free route for outsiders, which is why Moxo places first on a cost per decision point score of 4 out of 10, a figure three platforms here beat outright.
Move cost per decision point to 0.30 and how the human enters the path to 0.10, leaving the other three untouched, and the top of the table reorders: n8n takes first at 7.0, Moxo falls to a tie for second with Microsoft at 6.9, Zapier climbs to fourth at 6.2, and Pega and Camunda 8 tie for fifth at 6.1. Neither weighting is more correct in the abstract, and the weights here were fixed before any score was entered. This page weights the mechanism because the mechanism is the question it exists to answer, and a reader whose binding constraint is the invoice should read the cost per decision point scores rather than the totals.
An approval step proves someone was asked, not that anyone decided
Every platform on this list can show that a person was asked, and none of the five criteria tests whether that person weighed anything before saying yes. In our view that gap is what human-in-the-loop has to mean now: once an agent drafts the thing being approved, the failure to fear is not a missing step but a reviewer who approves what the machine wrote without reading it.
The EU AI Act names the tendency. Article 14(4)(b) requires that the people assigned to oversee a high-risk AI system be enabled "to remain aware of the possible tendency of automatically relying or over-relying on the output", which it calls "automation bias" (artificialintelligenceact.eu mirror, read 23 September 2026). A log built to answer "who approved?" cannot speak to that. What speaks to it is the interval between the item reaching the reviewer, or being opened, and the decision landing, together with how often the reviewer changes or refuses what they are shown.
You can test for it in a few minutes on any product you can get an approval step running in. Approve one test item after reading it properly and a second the instant it arrives, then export the decision record and see whether anything in it tells the two apart. If nothing does, the product can prove that a person was asked but not that anyone looked, and a year of approvals landing seconds after assignment, none edited and none refused, will read in that log exactly like a year of careful review. This is a practitioner's check rather than a sixth criterion, and it moves no score on this page.
What a decision point costs
Approval gates are metered on at least five incompatible bases, and no vendor reconciles its meter against anyone else's. The published figures, each against the cheapest plan that carries the gate:
- Microsoft Power Automate: approvals ride a standard connector, and Microsoft states that "any license that grants access to Power Automate and the ability to use standard connectors is sufficient to create approval flows", which includes Office 365 licences carrying Power Automate rights. Premium is $15.00 per user per month paid yearly. Copilot Studio agent flows meter separately, at 13 Copilot Credits per 100 agent flow actions, with capacity sold in packs of 25,000 Copilot Credits at $200.00 per pack per month (learn.microsoft.com and microsoft.com, 14 September 2026).
- Zapier: the Human in the Loop prerequisites are "A paid Zapier account" and the note that it "is a premium app on Zapier", with Professional at $19.99 a month billed annually. Zapier states that "Only successful Human in the Loop actions will count toward your task usage" (read 21 September 2026).
- n8n: the meter is the execution rather than the decision. Cloud Starter is 20 euros a month billed annually for 2,500 executions, and the self-hosted Community Edition carries no licence fee (n8n.io/pricing, 14 September 2026).
- Moxo: the meter is the flow. Team is $5,000 a year for 100 flows and $100 of AI a year with unlimited seats, and the audit log, the AI agents, Agent Foundry, webhooks, REST APIs and MCP all start on the custom-quoted Scale plan (moxo.com/pricing, 22 September 2026). The $500 monthly option read here on 15 September 2026 has been withdrawn, leaving a disabled billing toggle and a "Book a demo" button on every plan, so the cheapest way into Moxo is a year's commitment.
- Pega: the meter is the case, and the list price is published on AWS Marketplace rather than on pega.com. Workflow Automation is $162,000.00 a year for up to 120,000 cases with $1.35 per case beyond the allowance, rising to $1,188,000.00 a year for up to 1,000,000 cases at $1.19 per case (read 21 September 2026). It is the only per-case rate card in this ranking, and the only six-figure floor.
- Camunda 8, Kissflow and ServiceNow: no plan price is published. Camunda offers free local development, an Enterprise licence for Self-Managed production and a 30-day SaaS trial before a sales conversation. Kissflow's pricing page offers "Pricing built around the value you receive", "Fixed annual agreements" and "No surcharges for AI or API usage" (read 21 September 2026).
Three meters deserve a second look before a contract. Moxo's $100 of AI a year has no published conversion to the credits its product logs against a flow, so an allowance and a consumption figure cannot be reconciled from public pages. Microsoft's is stricter: once prepaid Copilot Credit capacity is exhausted, new agent flow runs are blocked, which turns a credit shortfall into a stopped approval rather than an overage line. Pega's counts cases, not decisions. A process carrying four approvals and a process carrying one draw down the same allowance at the same speed.
Considered and not scored
- Relay.app is excluded because the product is gone. Its own shutdown notice set deletion of paid accounts and all associated data after Monday 14 September 2026 at 23:59 PT. Its documentation had described the most complete edit-before-approve path any vendor here published, letting a reviewer approve the output as it stood, end the run, "Manually edit/override the AI output", or re-prompt the model. Teams moving off it should read What should teams do now that Relay.app is shutting down?
- Make was not scored because its Human in the loop app is not generally available. Make's own app documentation describes it as "currently in closed beta" and limited to invited customers on the Enterprise plan (read 14 September 2026).
- Pipefy fits this boundary, since guest users act inside a process without an account and plans count admins and members rather than guests. It sits outside this edition's scored set.
- Salesforce Agentforce records generative AI audit data in Data 360, including accept, regenerate, modify and decline actions, and is a genuine fit for this category. It also sits outside this edition's scored set.
- Temporal, Prefect, Apache Airflow and Windmill are durable execution engines and are scored on Best Durable Workflow Engines for Production in 2026.
- Lindy, CrewAI, Tines and Bardeen build or host agents rather than gating them, and are scored on Best AI Agent Platforms 2026.
- SuiteDash, Clinked, Content Snare and Dock are client portals. A portal shows an outsider a document and a status, and this list scores what happens when that outsider has to decide something inside your process.
Verification
Every figure on this page was read from a vendor-owned surface, with 21 September 2026 as the retrieval date unless another date is stated against the figure. The surfaces checked were vendor pricing pages, vendor legal pages including Moxo's Description of Services, and primary product documentation on help.zapier.com, docs.n8n.io, docs.camunda.io, learn.microsoft.com, community.kissflow.com, docs.pega.com and servicenow.com/docs, together with the Moxo template builder on 14 and 15 September 2026 and Pega's AWS Marketplace listing on 21 September 2026.
Two figures a buyer might expect are not published. Pega's Case history publishes its field list without a retention period beside it, and Pega's published rate card prices the case rather than the decision, so a cost per approval has to be derived from a volume assumption rather than read off a page. Both are stated in the Pega entry rather than estimated.
Where a vendor's own surfaces disagree, both readings appear on the page and the disagreement is stated, rather than the more flattering figure being chosen. Three examples appear on this page. Zapier's Human in the Loop page offers a reviewer type of "Anyone", defined as "members outside of your Zapier account", and states under Limitations on the same page that "Reviewers must have a Zapier account"; both are quoted in the Zapier entry. Microsoft's approvals page states "There are four approval types you can use" directly above a table with five rows (ms.date 29 April 2026). Relay.app's homepage and its documentation reversed the free and paid shutdown dates against each other, and the homepage reading is used here because it matches the company's own announcement.
Prices and plan gates move faster than mechanics. Before relying on a cost figure, re-read the vendor's pricing page and confirm which plan still carries the approval step, because on this list the plan that carries the gate and the plan that carries the audit log are frequently not the same plan.
Editor's Note: The weighting decides the top of this list. Move cost per decision point to 0.30 and how the human enters the path to 0.10, and n8n takes first at 7.0, with Moxo and Microsoft tied for second at 6.9, so we would point a reader whose binding constraint is the invoice to the cost per decision point scores rather than the totals. The check we would carry into any evaluation sits outside the scores: approve one test item after reading it properly and a second the instant it arrives, then export the decision record and see whether anything in it tells the two apart. The EU AI Act calls the tendency behind a rubber stamp "automation bias", and the check takes a few minutes. — Rafal Fila, ShadowGen