Feature · AI Workflow

AI workflow software with the model wired into the step, not stapled on through a webhook.

Classify inbound messages, extract fields from PDFs, draft reply emails in the rep voice, score leads against ICP, summarize long notes, and let Strkr AI decide which branch of the flow to take. Every step is a native node with the same observability as a database write, the same retry policy as an outbound API call, and no second subscription to maintain.

Start free trial See pricing
What an AI workflow should feel like inside a CRM

Nine specific jobs the model should do without a workaround.

AI workflow has become a crowded phrase that mostly points at two shapes: a chatbot bolted on next to a tool, or an orchestration tool that treats the model as an outbound HTTP call. Neither shape is what a revenue team actually needs on a Tuesday. The pragmatic version is a workflow builder where the model is a first-class step with the same contract as any other node. It receives typed inputs. It emits typed outputs. It respects retries, timeouts, idempotency, and the audit log. It participates in the same transaction as the database write that follows it. The checklist on this page covers the nine jobs reps and admins have asked us for since the first beta of flows. Strkr ships every one of them as a native step, not as a webhook pattern. The admin building a flow drags an AI node onto the canvas the same way they drag a Create Task node, and the model understands the records on the row the flow is working.

Classify

Label an incoming record from a short list.

Point a Classify step at any string on any record, pass a label list in the config, and the step writes back the chosen label with a confidence score. Inbound support emails get a severity. Fresh leads get an industry. Call transcripts get an outcome tag. The admin picks a threshold. Below it, the flow routes to a human triage branch instead of guessing.

Extract

Pull structured fields out of unstructured text.

Give the Extract step a JSON schema and an input string. It returns a validated object. Pull purchase order number, due date, and ship-to address from an email. Pull revenue, headcount, and industry from a company About page. Pull next steps and owner from a meeting note. The schema is the contract. The step refuses to emit shapes that fail it.

Draft

Compose a first-pass message in the rep voice.

The Draft step reads the record on the row and composes an email, SMS, or note in the configured tenant voice. A follow-up draft after a demo. A renewal pitch after a usage dip. A thank-you after a closed-won. The flow queues the draft on the owner for a review rather than sending blind. Humans edit the last five percent; the model writes the first ninety-five.

Score

Return a number plus the reasons that drove it.

Score a lead against an ICP shape, a deal against a comparable-cohort pattern, or a renewal against a churn signal library. The step emits a number between zero and one hundred, plus the three strongest contributing factors in plain English. Both land on the record. The admin filters a list view by score and a sales manager reads the receipts next to it.

Summarize

Compress a long record into a two-paragraph brief.

Collapse a quarter of meeting notes on an account into a two-paragraph briefing. Fold ninety days of support tickets into the current risk picture. Turn a lengthy inbound email thread into a bullet list a new rep can read in forty seconds. The step accepts a target length and a required-sections list so the output shape stays predictable across runs.

Decide

Pick a branch from a labeled set of choices.

Give the Decide step a question, a set of candidate branches, and the data on the row. It picks one. Route a lead to the right queue when geographic rules are not enough. Pick a renewal play from an eight-item playbook. Route an inbound email to Sales or Support when the subject line is ambiguous. The decision lands in the audit log with the model reasoning attached.

Compare

Rank two options against a stated criterion.

Compare two proposed meeting times against a stated priority, two candidate accounts for an enterprise motion, or two draft emails against a tone target. The step returns the winner and a short reason. Flows that previously branched on hand-written if statements can defer to the Compare step when the rule is softer than code can comfortably express.

Verify

Check a draft before the flow sends it.

The Verify step reads a draft and a rule set, then returns pass or fail with a reason. Verify an outbound email against brand voice. Verify a quote number against a discount policy. Verify a renewal narrative mentions the three required talking points. Verify a scheduled SMS respects the quiet-hours config. If it fails, the flow routes to human review automatically.

Chain

Feed one AI step straight into the next.

Extract fields from an inbound email, classify the extracted intent, draft a response conditioned on the classification, score the draft against a tone rule, then verify the final output. The chain runs in one flow with typed hand-offs between nodes. Each step keeps its own idempotency key so a retry on step four does not re-run steps one through three.

The anatomy of a Strkr AI step

What makes a model call feel like a database write.

The thing that breaks most AI workflow builders is that the model call looks nothing like the rest of the canvas. It has a different error model, a different retry policy, a different observability story, and a different billing meter. The admin ends up building a parallel universe around every AI node. Strkr AI steps carry the exact contract of every other node on the canvas. Typed inputs, typed outputs, a known timeout, a known retry curve, a visible run log, a line item in the audit history, and no credit meter sitting next to the step counting prompts. The AI node is one more node, not a different animal.

Typed in, typed out

Every port carries a schema.

Drag an AI step onto the canvas. The input ports accept typed values from upstream nodes. The output ports emit values the downstream nodes can safely read. The admin never passes an untyped blob through the step and prays. A node emitting an email draft lets the downstream Send Email step read the subject and body directly, with validation at the boundary.

Idempotent by default

Retries do not fan out side effects.

Every AI step carries an idempotency key derived from the input hash and the flow run id. A retry on a transient error returns the earlier result rather than re-prompting the model and double-charging the step. The downstream nodes see the same output they would have seen on the first attempt, so retries never create duplicate drafts, scores, or decisions.

Observable

Every run shows inputs, outputs, and token count.

Open the run log on any flow. Click an AI node. See the exact input the step received, the exact output it emitted, the elapsed time, the token count, and the model that handled the call. Debugging an AI step feels like debugging a stored procedure. Nothing is a black box. Nothing lives in a vendor dashboard three logins away.

Timeout and fallback

Slow calls do not stall the flow.

Every AI step carries a configurable timeout. On breach, the step emits a fallback value defined in the config and the flow continues. A classify step with a five-second timeout falls back to the default label and routes to a human review branch. The flow never hangs on a model that is having a slow day.

Guardrails inline

Policies run before the output is handed off.

Attach a guardrail block to any AI step. Block outputs that mention a banned term, that include profanity, that exceed a length cap, or that fail a regex check. The guardrail runs in the same transaction as the step. A failing output never reaches the downstream node, which keeps the compliance posture coherent across the whole canvas.

Audit trail

Every call is attributable to a flow run.

The audit log carries one row per AI step execution: who triggered the flow, what the input was, what the model emitted, which tenant policy applied, and how long the call took. Compliance walks a flow quarter from the log, not from a vendor export. The audit view ships inside Strkr; nothing sits in a second console.

No credit meter

The seat price is the full price.

Strkr AI steps do not consume a per-call credit meter that bills on top of the seat. The admin building the flow is not budgeting prompts against a monthly cap. The flow runs millions of times a month without a procurement conversation about overage. The pricing page is the whole pricing story.

Permission-aware

The step reads what the actor can read.

An AI step inside a flow runs with the effective permissions of the service account on the flow. It cannot read records the service account cannot read. It cannot write fields the service account cannot write. The model inherits the row-level and field-level permissions on the tenant, not a wide-open bypass that leaks data across teams.

Tenant-local prompts

Your prompts never train anyone else model.

Prompts, inputs, and outputs from Strkr AI steps are tenant-isolated and are not used as training data for the shared model. The admin opens the compliance page, sees the data handling posture in one line, and the legal review on a new flow closes in minutes rather than weeks.

AI workflow patterns the Strkr team sees most

The ten flows reps and admins build first.

When a Strkr tenant turns on the AI steps for the first time, there is a predictable set of flows that get built in the first week. Admins at a 40-rep SaaS inbound-heavy sales floor and admins at a 200-person services firm converge on surprisingly similar patterns because the underlying jobs are the same. The list below is the ten patterns we see most often, each one shippable in under thirty minutes on the canvas and each one running on real records inside the pipeline from day one.

Inbound triage

Classify, score, enrich, route.

A fresh web lead lands. The flow extracts a company domain, enriches firmographics, classifies the industry, scores against the current ICP, and routes to the right queue. An SDR who used to spend four minutes per lead on triage opens an already-tiered inbox. The flow runs end to end in under eight seconds.

Inbound email router

Sales, Support, Billing, or human triage.

A shared inbox collects messages from a dozen customer channels. The flow classifies each one into Sales, Support, Billing, or Other, writes a short extracted summary onto the record, and routes to the matching queue. A seven-person team that used to spend an hour each morning on inbox sorting opens an already-sorted inbox at 09:00.

Deal risk scorer

Nightly rescore, three-sentence why.

A nightly flow walks every open opportunity and emits a risk tier plus three reasons: last-activity age, stakeholder density, discount depth. The reasons land on the deal record. A manager sorting the Monday pipeline review opens from the top of the risk list and skips eighty percent of the healthy deals that do not need coaching this week.

Churn watcher

Signals compound into a tier change.

A churn flow runs across every active customer every night. Usage trend, support ticket volume, invoice aging, and stakeholder churn each contribute a signal. The flow lifts an account from green to yellow when at least three signals move together, so a single bad week in one metric does not red-flag a healthy book.

Note enricher

Call recap into structured outcome.

A rep clicks the mic icon after a call and dictates for ninety seconds. The flow transcribes, extracts next step and owner, drafts a follow-up email in the rep voice, updates the stage recommendation on the deal, and writes a two-paragraph recap for the manager. The rep spends twelve seconds reviewing instead of twelve minutes composing.

Renewal composer

Draft the pitch 120 days out.

Every active contract triggers a flow when the end date crosses the 120-day window. The flow pulls usage over the term, lists wins since the last renewal, flags open support items, and drafts a renewal narrative in the CSM voice. The draft lands as a task on the owner. The CSM edits, sends, and the motion starts four months early rather than four weeks late.

Expansion play fire

White space signal into a book meeting.

When a customer hits a seat ceiling, adds a specific stakeholder title, or activates a feature flag that correlates with cross-sell, the flow fires. It drafts an email, books a slot on the exec sponsor calendar, assigns a task to the AE, and logs the play on the account. The AE sees the next step pre-queued.

Invoice extractor

PDF in, structured record out.

A vendor emails an invoice as a PDF. The flow runs an Extract step with a schema, pulls invoice number, amount, due date, line items, and ship-to, writes to the Finance object, and routes to the correct approver. Finance stops hand-keying thirty invoices a week. Errors go to a review queue instead of straight-through.

Lead reactivation

Old closed-lost into a fresh conversation.

A quarterly flow walks every closed-lost opportunity older than nine months, scores reactivation likelihood against the current product shape, and surfaces the top fifty for the SDR layer. Each one carries a drafted opener and the one feature flag that changed since the deal lost. Dead pipeline becomes a sourced lane again.

Governance for AI workflows

The config pane that keeps compliance in bounds.

Admins who run AI workflows at scale eventually ask the same four questions. Which steps are allowed to send outbound messages without a human review. Which steps are allowed to touch which fields. Which steps are allowed to run on which object types. And who gets to approve a change to an AI step in production. Strkr answers all four from a single Governance pane on each flow and a tenant-wide policy inherited by every flow in the workspace. The admin configures the posture once. New flows inherit it. Flows that attempt to drift from it fail validation at save time.

Human-in-the-loop toggles

Mark the steps that need a review.

Every AI step carries a review toggle. On, the step emits a task on a configured owner, pauses the flow, and resumes on approval. Off, the step runs through. The admin uses the toggle to require review on outbound messages while leaving internal scoring steps straight-through. The policy pane shows every flow with the toggle state at a glance.

Field-level allowlists

Say which fields the model may write.

A tenant admin configures an allowlist per object: which fields an AI step is allowed to write, which fields it is not. The validator refuses to save a flow that writes outside the allowlist. Compliance stops auditing by reading YAML; they audit from the pane that lists allowed fields in plain English.

Approval workflow

Changes to prod flows need a second pair of eyes.

Flag any flow as production. Changes to that flow require a second admin to approve the diff before the version promotes. The diff view shows the before and after of the prompt, the schema, the retry policy, and the guardrail config. Nothing slides into prod with one keystroke.

Prompt version history

Every prompt change, with a diff.

Prompts behave like source code. Every save creates a version. The admin pulls up any prior version, sees the diff against the current one, and reverts in two clicks. Audits walk the prompt history alongside the flow run history, so compliance can connect a disputed output to the exact prompt that produced it.

Role-scoped execution

Who may run which flow.

Flows carry an execute role. An admin flow runs for admins only. A user flow runs for anyone with the standard CRM seat. A system flow runs on schedules under a service account. The role model is identical to the role model used across the rest of the product, so there is one place to think about who may do what.

Rate caps

Cap how often a flow may fire.

Attach a rate cap to any flow. Fifty runs per minute per user. Five hundred runs per hour per tenant. The cap stops a misconfigured upstream webhook from firing a million runs on a Saturday. The cap also stops a runaway loop. The default caps ship conservative so admins have to deliberately raise them.

PII masking

Fields the model never sees.

The tenant admin marks fields PII. AI steps running against records containing those fields receive the record with the PII fields masked. The step still processes the row and still writes the allowed outputs; it just never sees the fields the policy forbids. The posture is enforced in the step, not in a wrapper the admin can forget.

Residency pinning

Keep the run in region.

AI steps can be pinned to a region so the request, the response, and the logs stay inside that region. A tenant with an EU residency requirement pins every flow to the EU region, and the step validator fails any flow that attempts to escape it. The admin configures the posture once and every new flow inherits it.

Export controls

Log who looked at AI step inputs.

Every view of an AI step input in the run log is logged. A compliance officer reviewing an investigation sees exactly who read the prompt and the output, when they read it, and from where. Insider risk reviews walk the log for anomalous reads, which is the same posture every well-run SOC walks for database queries.

How AI workflows plug into the rest of Strkr

The step talks to every object, every surface, every tier.

The payoff of a native AI step shows up in the second week. Admins stop building glue between the model and the records; the step is already wired. A flow that scores a lead can write back to the lead on the next line. A flow that drafts a renewal pitch can post the draft to the account activity timeline, log the task, and schedule the follow-up in one canvas. Strkr builds that integration natively because the AI step reads from and writes to the same object model the rest of the product does.

Records in context

The step reads the row it works.

When a flow fires on a record change, the AI step on the canvas gets the record as structured input. Account name, owner, open opportunities, last activity, custom fields. No separate lookup node, no manual join. The step reasons about the row directly, which keeps the canvas short and the output accurate.

Messaging as a native action

Draft, verify, send through Strkr Messaging.

An AI draft feeds a Send SMS node that routes through Strkr Messaging. Opt-outs, consent, quiet hours, and audit all apply. No carrier shows up by name in the UX. The compliance posture on outbound messages is the same posture every other written channel on Strkr respects, enforced inside the flow.

Forecast integration

Scored deals change the forecast number.

An AI risk score written to an opportunity flows into the forecast rollup immediately. The sales leader looking at the forecast sees the model-called number alongside the rep-called number without a nightly batch job stitching them together. The whole loop closes inside Strkr.

Reporting on the output

Every AI field is a report column.

Any field an AI step writes is a reportable field. Build a dashboard on lead score distribution. Build a weekly report on classified ticket mix. Pin a chart to the home. The AI output is first-class data in Strkr, not a side table the BI tool has to join against. The reports also respect field-level permissions the admin set in Governance.

Project workflow steps

Sprint, release, and backlog triggers.

AI steps inside flows wired to the Projects module can summarize a sprint, score an incoming bug, draft a release note, or classify a backlog item. The step reads the project row with the same contract as it reads a CRM record, so the Projects admin building an automation uses the same canvas and the same AI primitives.

Documents module read

Context-aware search across the wiki.

An AI step can retrieve context from the Strkr Documents wiki. A support draft reads the latest version of the troubleshooting guide. A sales draft reads the current objection-handling note. The flow stays inside Strkr; nothing crosses to a third-party retrieval service holding a copy of the wiki.

API ingress

Flows receive external events over the API.

An external system posts an event over the Strkr API, the matching flow fires, and the AI step runs on the incoming payload. A product telemetry pipeline feeds a usage-drop signal into a Draft step that queues a save play. The API path accepts the same auth and the same rate limits as any other API consumer on the tenant.

Scheduler and recurring runs

Flows that run on cron, not on triggers.

Schedule a flow nightly, hourly, or at a specific hour each weekday. The AI step inside runs on that cadence. A weekly flow that composes a Monday pipeline brief for every manager runs at 06:30 Sunday evening; by Monday morning every manager has their pre-read in the inbox.

Flow version promotion

Dev, staging, prod with real promotion.

A flow has a dev, a staging, and a production lane. Changes to the AI step prompts move through the lanes with approvals at each boundary. Prod never surprises the team; staging always sees the exact flow version prod will see next. The promotion is a feature, not a Pulumi script the admin maintains on their laptop.

AI workflow software that treats the model like any other step on the canvas.

Classify, extract, draft, score, summarize, decide, verify, compare, and chain. All typed, all retried, all logged, all guardrailed. On every paid tier with no credit meter. Open a trial, drag a Draft step onto a flow, and watch it reason about a real record in your pipeline in under a minute.

Start free See flows docs
Common questions

What buyers ask about this feature.

What is an AI workflow and how is it different from a workflow with a webhook to an AI service?

An AI workflow treats the model as a first-class step inside the automation canvas, with typed inputs, typed outputs, retries, idempotency keys, timeouts, guardrails, and observability on par with every other node. A workflow that calls an AI service via webhook treats the model as an opaque outbound HTTP call, which means retries can double-send, outputs are untyped blobs the next step has to parse, and the audit trail lives in two places. Strkr ships native AI steps so the admin drags a Classify, Extract, Draft, Score, Summarize, Decide, Compare, or Verify node onto the canvas the same way they drag a database write, and every step honors the same contract the rest of the flow does.

Does Strkr charge per AI call or meter credits for AI workflow steps?

No. Strkr AI runs on every paid tier with no per-call credit meter. The seat price covers model usage inside flows at normal operating volume. Admins build flows that run millions of times a month without a procurement conversation about overage, which is the main reason comparable AI workflow builders quietly stall out six months after launch. Rate caps exist as a safety mechanism against misconfigured upstream webhooks firing in a loop, but they are a protection against runaway flows, not a billing meter.

Which AI step types does Strkr ship, and what do they do?

Nine native step types: Classify (label from a short list with confidence), Extract (pull structured fields from unstructured text against a schema), Draft (compose email, SMS, or note in the tenant voice), Score (number plus three reasons), Summarize (compress long text to a target length with required sections), Decide (pick a branch from candidate options), Compare (rank two options against a criterion), Verify (check an output against rules before hand-off), and Chain (feed one AI step directly into another with typed hand-offs). Every step runs under the same retries, timeouts, idempotency keys, guardrails, and audit logging as every other node on the flow canvas.

How are permissions and data residency handled for AI workflow steps?

AI steps run under the effective permissions of the flow service account, which means they inherit row-level and field-level permissions from the tenant. A step cannot read a record the service account cannot read and cannot write a field outside the admin-configured allowlist. Prompts, inputs, and outputs are tenant-isolated and are not used to train the shared model. Fields marked PII are masked before the model sees the row. Residency pinning keeps the request, response, and logs inside a chosen region, and the step validator fails any flow that attempts to escape the pinned region.

Can we require human review on specific AI workflow steps before the output is sent or written?

Yes. Every AI step carries a human-in-the-loop toggle. When on, the step emits a task on a configured owner, pauses the flow, and resumes when the owner approves or edits the output. Admins commonly flip the toggle on for outbound messaging (email, SMS) and leave it off for internal scoring and classification. Flows marked production require a second admin to approve changes, prompt version history records every edit with a diff, and export controls log who viewed which AI step input, so compliance has a complete audit trail without maintaining a parallel system.

What are the most common AI workflow patterns teams build first in Strkr?

The top patterns our tenants build in the first week: inbound lead triage (classify, score, enrich, route in under eight seconds), inbound email router (Sales/Support/Billing/human), deal risk scorer (nightly rescore plus three reasons), churn watcher (multi-signal tier changes), note enricher (dictation to structured outcome), renewal composer (draft pitch 120 days out), expansion play fire (white space signal to booked meeting), invoice extractor (PDF to structured record), and lead reactivation (old closed-lost to fresh conversation). Each ships in under thirty minutes on the canvas and runs on real records on day one. All of them inherit the governance posture configured on the tenant.

Related features

Pair this with...

Try it free. Bring your team next week.

No sales call, no migration consultant, no four-month implementation. Enter your card, get 14 days of the full Pro tier, cancel any time before day 14 with zero charge. Spin up a workspace, import your CSV, and have something useful before lunch.

Start free trial See pricing