Activepieces combines fixed flows with agents for variable work

Activepieces brings visual automation and AI agents into the same project environment. A fixed flow can receive an event, transform data and update another application; an agent can handle a less predictable part of that process and return structured information. The useful design question is where judgment belongs. Keeping routine operations explicit and confining AI to a well-defined decision makes the result easier to maintain and its credit usage easier to understand.

01 / ProductFlows and agents solve different parts of a process

A flow begins with a trigger and continues through configured actions. Triggers can be scheduled, webhook-based or tied to an application event. Actions can communicate with services or run code. The flow-building guide² describes this simple model, which works well when the procedure is already known and the main task is making it repeatable.

An agent instead receives a job and a set of tools, then decides which tools to use as it runs. Activepieces documents agents that can use application actions, reference documents and tables, and operate as a step inside a flow. An agent belongs to a project, which determines the connections, flows and files it can reach. The agent overview³ makes project placement part of the design, rather than merely a way to organize the sidebar.

The distinction is practical. Checking whether a form contains an email address is a fixed rule. Interpreting an unusual supplier description may benefit from AI. Writing an approved supplier identifier into a database is again a fixed operation. Putting all three inside an open-ended agent would make the procedure less predictable without necessarily improving the ambiguous middle step.

02 / AudienceA fit for teams willing to own their procedures

Activepieces suits operations teams with a technical collaborator, internal tools teams and developers who want a visual workflow environment with a self-hosting option. It can be attractive when many steps are straightforward but a small part requires classification, extraction or choosing between a limited set of actions. The buyer should have someone who owns the process after the first successful demonstration.

The strongest alternative depends on that ownership model. Read n8n's blueprint when deployment control and a technical workflow platform are central. Read Make's blueprint when a visual automation service is the main requirement and the team wants to compare how branches, usage and maintenance behave in practice. A connector count alone will not establish which product makes a particular workflow easier to operate.

A weak fit is a team expecting the agent to resolve an undefined business policy. If nobody agrees what makes a supplier eligible for a category, the software cannot supply an authoritative rule. Clarifying that rule is often the most valuable part of the pilot. Once agreed, it becomes possible to decide which parts should be fixed logic and which genuinely require interpretation.

03 / WorkflowA supplier-intake flow with one bounded AI task

Consider a proposed supplier-intake workflow. A form submission creates a candidate record with a stable submission identifier, company domain and requested category. Fixed steps check required fields, normalize the domain and search the existing supplier register. Exact matches are marked for review rather than automatically creating a second company. These steps should remain visible and deterministic.

Only then pass the supplier description and the approved category definitions to an agent. Ask it to return a proposed category, the supporting phrase and an uncertainty reason when the description does not fit. The flow can store those fields and create a review task.

Give the agent fewer tools than the whole project uses

The classification step needs reference material and a structured response. It does not inherently need the ability to approve a supplier or edit payment details. Place the agent in the appropriate project, then review which tools and connections are actually necessary. A useful test is whether the same proposed category can be produced without granting any destination write capability to the agent itself.

After review, a fixed flow can apply the approved category and record the reviewer and source submission. This separates the model's interpretation from the organization's decision. When someone challenges a category later, the team can inspect the original description, the model's proposed reason and the human approval independently, instead of trying to infer all three from one final database value.

Publish changes deliberately

Activepieces keeps published versions locked. Editing a published flow creates or uses a draft, while the existing published version continues to be the released definition. The version-history guide⁴ is useful here: an editor working on a revised classification prompt should not assume that saving the draft has changed the procedure used by incoming submissions.

For the pilot, keep a small set of difficult supplier descriptions and compare the proposed change against that set before publication. Include a company that offers two categories and one that uses vague marketing language. The purpose is to detect a changed decision boundary, not to manufacture a headline score from a few convenient examples. Record the published version alongside the review artifact when the process needs that traceability.

04 / PricingCurrent pricing uses credits across runs and AI

The pricing page¹, accessed 15 September 2026, now describes a credit system. One flow run uses one base credit regardless of its number of steps. AI work adds credits, with listed model tiers and agent actions using additional units.

The same page lists additional usage on Plus and Team at $0.007 per credit. With a supported bring-your-own-key setup on Plus or above, an AI step uses one platform credit while model charges go to the model provider separately. That shifts part of the bill; it does not make the model execution free. Compare the combined cost for your actual prompts and outputs.

For supplier intake, estimate a normal accepted submission, an ambiguous submission and a duplicate separately. A duplicate rejected before AI processing should have a different cost profile from an agent that calls several tools to resolve an unclear description. Put explicit limits around how much investigation is useful. A supplier record that needs a person can be routed to that person instead of consuming an open-ended sequence of model calls.

The Free plan's daily allowance also changes the meaning of availability. The Free plan stops running work when its daily allowance is exhausted. An active toggle is therefore insufficient evidence that an intake process is keeping up. Monitor completed records and unresolved submissions, especially during a campaign or import that concentrates work into one day.

05 / DistinctionsSelf-hosting changes the work your team owns

The architecture documentation⁵ describes an application service, workers, sandboxes, an execution engine, Postgres and Redis-backed jobs. This provides useful visibility into the operating model. It also explains why a self-hosted installation is more than a container that happens to display the same builder as the cloud product.

A team choosing self-hosting must own database and queue durability, file storage, upgrades and worker capacity. The open-source automation core and the paid agent or enterprise layers should be evaluated as separate entitlements. Check the actual edition for the capabilities you plan to use instead of assuming that every feature shown on the cloud pricing page is included in a free self-hosted deployment.

This can still be the right tradeoff. An internal platform team may already operate the required infrastructure and prefer keeping integration execution within its environment. A department without that support may gain little from avoiding a cloud subscription if it creates a new service nobody is prepared to restore after a failure. The comparison is total operating responsibility, not license price alone.

06 / QuestionsRecovery guarantees have a precise boundary

Activepieces' crash-recovery documentation⁶ says that checkpointed completed steps are skipped when a run resumes. It describes the recommended production configuration and notes that queue durability depends on Redis. That is a meaningful mechanism, but it should not be read as a universal promise that every external side effect can happen exactly once under every failure condition.

In the supplier workflow, consider a destination that accepted a create request immediately before the worker lost contact. If the step's completion was not recorded, the integration needs a way to recognize the already-created supplier. A stable submission identifier and a destination lookup can resolve that uncertainty. This is a property of the business operation, not something a visual success indicator can establish retroactively.

The pilot should also demonstrate a changed connection, a malformed agent response and an unavailable destination. Check that the review queue preserves the submission and identifies the next action. A robust intake process is allowed to stop on an uncertain case; it should not silently turn uncertainty into a new supplier or lose the record while showing that the flow is still enabled.

07 / DecisionChoose the balance between fixed procedure and AI

Activepieces deserves consideration when a team wants explicit workflows, selective agent behavior and a choice about where execution runs. Start with a process whose ordinary path is understood, then assign AI only the part that benefits from interpretation. The resulting flow should still make sense to the person responsible for its business outcome.

A good first proof is one supplier intake that produces an explainable review record, a duplicate that produces no second supplier and a revised draft that leaves the published process stable until release. Use that evidence to compare cloud credits, self-hosting effort and alternative products. Expand the agent's role only when it improves a defined part of the procedure.