Agentic workflow examples, drawn step by step.
Nine processes from six industries. In each one, code keeps the order, AI reads and judges where a rule cannot, and a person signs off where the stakes call for it.
An agentic workflow is a business process whose steps are fixed in advance, with AI placed inside the steps that need reading, judgement or writing. The examples below are drawn the way they run: what starts them, which steps are code, which are an agent or a decision, and where a person steps in.
Updated · 5 sources
What are some agentic workflow examples?
Typical agentic workflow examples are business-customer onboarding in banking, claims intake in insurance, prior authorization review at a health insurer, invoice-to-payment in finance, freight invoice audit in logistics, supplier non-conformance in manufacturing and support ticket triage. Each keeps a fixed order of steps, uses an AI agent to read documents or free text, makes one or two bounded decisions with a confidence, and hands exceptions and high-stakes calls to a person.
The five workflow patterns, and which ones the nine examples use.
Anthropic’s engineering guidance separates workflows — models and tools orchestrated through predefined code paths — from agents that direct their own process. It names five workflow patterns. Each example below uses one or more of the first four.
- Prompt chainingFixed steps in sequence; each model call works on the last one’s output. Extract, then check, then draft.
- RoutingClassify the input, then send it down the path built for that kind. A claim photo and a police report take different steps.
- ParallelizationIndependent checks run side by side, or the same check runs several times and votes.
- Evaluator-optimizerOne model drafts, another checks against criteria, and the draft is revised until it passes.
- Orchestrator-workersA model splits the task at run time and hands parts to workers; the parts depend on the input.
Where each pattern shows up below
- Prompt chainingPrior authorization: extract the request, then compare it with the policy criteria, then draft the summary.
- RoutingClaims intake and support triage: the category decides which queue and which steps follow.
- ParallelizationBusiness onboarding: sanctions screening, ownership analysis and document checks run at once. Freight audit: every charge line is checked side by side.
- Evaluator-optimizerSupport replies: a second check confirms every statement against the cited article before sending.
- Orchestrator-workersNone of the nine needs it: their parts are known before the run starts. It fits research-style work, where a model decides at run time what the parts are.
Agentic workflow examples in banking, lending and finance operations.
Financial processes are rule-heavy and document-heavy at once. Code holds the rules exactly; agents read what arrives — certificates, statements, invoices — in whatever layout it comes.
1 · Business customer onboarding (know your customer)
- Application arrives
Forms, certificate of incorporation, ownership chart and IDs land in one package.
- Read the entity
An agent extracts legal name, officers and every owner with their stake, citing the page.
- Owners found, names screened
Code multiplies stakes through each layer of the chart to list anyone at 25% or more; sanctions and politically exposed person checks run by API for every name at once.
- Rate the risk
A decision places the customer on low, medium or high, with a confidence.
- Officer signs off
High risk, any screening hit or an unsure rating goes to a compliance officer’s review.
2 · Loan application pre-check
- Package sorted
Pay slips, bank statements and tax returns arrive together; each file is classified and missing types are listed at once.
- Extract income and debts
An agent reads amounts, periods and employers from each document.
- Apply credit policy
Code computes ratios and compares them with the bank’s own policy.
- Flag exceptions
A yes-or-no decision: does anything in the file need an underwriter’s judgement?
- Underwriter decides
The underwriter sees the extracted figures beside the source pages.
3 · Invoice to payment
- Invoice arrives
From a mailbox, a portal or a shared drive.
- Extract
An agent reads supplier, totals and line items from any layout.
- Three-way match
Code matches the purchase order and the goods receipt, line by line.
- Pay, hold or reject
A decision picks one, with a probability for each.
- Holds reviewed
Holds and unsure answers go to accounts payable; the rest is posted.
Agentic workflow examples in insurance claims and healthcare prior authorization.
Both start with a bundle of mixed documents and end with a decision someone must be able to explain. The workflow keeps the deadline and the audit trail; agents do the reading.
4 · First notice of loss — claims intake
- Claim reported
An email or a portal form arrives with photos, a police report and an estimate; each file is typed and follows the steps built for it.
- Extract the loss
An agent reads date, place, parties and damage from the form and the report.
- Check the policy
An API call confirms cover was in force on the loss date.
- Score severity
A decision places the claim on a severity scale and answers yes or no on fraud indicators.
- Route
Simple claims go to fast track; severe or flagged ones to an adjuster’s queue.
5 · Prior authorization review at a health insurer
- Request received
Through a portal, an electronic interface or a scanned fax. Code starts the deadline clock and escalates before it lapses.
- Extract the request
An agent reads diagnosis, procedure and the clinical notes.
- Fetch the criteria
Code retrieves the plan’s medical policy for that procedure.
- Compare, criterion by criterion
An agent marks each criterion met or not, quoting the passage in the notes.
- Clinician decides
A clinical reviewer sees the evidence map and decides; every denial carries its specific reason.
Why a person stays in these two. A claim payment and a medical decision both need a reason a regulator can read. The workflow gives the reviewer the extracted facts with their sources; it does not take the decision away from them.
Operations examples: freight audit, supplier quality and contract renewals.
Back-office operations are where volume meets variety: thousands of carrier invoices, inspection reports and contracts, each in its own format, checked against tables and dates that code handles exactly.
6 · Freight invoice audit (logistics)
- Carrier bills
A freight invoice arrives; each charge line — base rate, fuel, accessorials — becomes its own check.
- Read the charges
An agent extracts codes, weights, distances and amounts.
- Match the contract
Code compares each line with the contracted rate and the shipment record.
- Classify the gap
A decision labels each difference: valid extra, billing error, or proof needed.
- Dispute drafted
An agent drafts the dispute; an analyst sends it, and the approved amount is paid.
7 · Supplier non-conformance (manufacturing)
- Inspection fails
An incoming inspection logs a failed lot.
- Read the report
An agent reads the inspection report and the supplier’s past cases.
- Grade the defect
A decision assigns a defect type and a severity level.
- Trace the lots
Code finds every lot from the same batch and where it went.
- Request drafted, then approved
An agent drafts the corrective action request; a quality engineer edits and sends it, and a timer watches for the reply.
8 · Contract renewal watch (legal and procurement)
- Every Monday
A schedule starts the run.
- Find what is due
Code queries contracts ending in the next 90 days.
- Read the terms
An agent reads notice periods and auto-renewal clauses.
- Report
A templated report lists what needs action, by owner.
- Owners act
Each owner gets their list, with the clause quoted.
Support ticket triage: an agentic workflow that checks its own replies.
Customer service is a common first agentic workflow. This example uses the evaluator-optimizer pattern: one step drafts a reply, a second step checks every statement in it before anything is sent.
9 · Support ticket triage and reply
- Ticket arrivesEmail, chat or web form.
- CategorizeA decision returns product area and urgency, each with a confidence.
- RetrieveThe steps for that category fetch the matching help articles and the customer’s account.
- DraftAn agent writes the reply, citing the articles it used.
- CheckA second step confirms each statement against its cited article and the company’s tone rules.
a reply that fails the check is revised, at most a few times; then a person takes it
Where the run ends
- Passed, low riskSent, and the ticket closes with the reply on record.
- Passed, account changeQueued for a support agent to approve before anything changes.
- Unsure or failedThe draft and its failed checks go to a person, so they start halfway.
How to pick the first agentic workflow for your own team.
The examples that work share a shape: a known process, varied inputs, a clear answer someone can check. The ones that stall usually picked a process nobody could describe, or one with no way to measure the result.
- Draw the steps firstIf the team cannot draw the process on a whiteboard, it is not ready to automate.
- Mark each stepCode where a rule decides; an agent where text must be read; a decision where one judgement is needed.
- Keep sums in codeTotals, dates, ratios and thresholds belong in code, where they are exact.
- Make unsure an answerEvery AI step needs a confidence, and a path to a person when it is low.
- Collect fifty real casesWith the answers a careful person would give, before the first build.
- Name the ownerSomeone owns the exceptions queue from day one.
How Botminds runs these agentic workflow examples as XFlows.
On Botminds each example above is an XFlow: a graph of steps on one canvas, started by a document, an event, a schedule or the API, with agents and typed decisions placed exactly where judgement is needed.
Read the full design in agentic workflows, then build one in the Academy: the workflow pattern, invoice settlement, support intake triage and loan underwriting.
Agentic workflow examples: common questions.
The technology behind itAgentic workflowsFixed steps, AI where judgement counts→What are some agentic workflow examples?
Common agentic workflow examples include business customer onboarding, loan pre-checks, invoice to payment, insurance claims intake, prior authorization review, freight invoice audit, supplier non-conformance, contract renewal watch and support ticket triage. Each has fixed steps, AI where documents must be read or a judgement made, and a person for exceptions.
What are agentic workflows in AI?
Processes whose steps and order are set in advance, with AI agents doing the steps that need reading, judgement or writing. Anthropic describes workflows as systems where models and tools are orchestrated through predefined code paths, as distinct from agents that direct their own process.
What is the difference between an agentic workflow and an AI agent?
In an agentic workflow the business fixes the steps and AI works inside them. An AI agent chooses its own next step and tools as it goes. Use a workflow when the process is known and the inputs vary; use an agent when the path changes every time.
What are examples of agentic AI in banking?
Business customer onboarding, where agents read ownership documents and code finds the beneficial owners; loan pre-checks, where agents extract income and code applies credit policy; and payment exception handling. A compliance officer or underwriter keeps the final decision on high-risk cases.
What are examples of agentic AI in healthcare?
Prior authorization review, where an agent compares clinical notes with a plan’s medical criteria and quotes the evidence for a clinician; claims and referral intake; and document-heavy administrative work. Medical decisions stay with clinicians, supported by the extracted evidence.
What are the most common agentic workflow patterns?
Prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer — the five named in Anthropic’s guidance. Most real workflows combine two or three: route by document type, chain extraction and checking, and add an evaluator before anything goes to a customer.
How do I choose the first process to turn into an agentic workflow?
Pick one with high volume, varied documents, a process the team can draw, and answers someone can check. Collect real cases with known answers before building, so you can measure the first version.
Do agentic workflows replace people?
In these examples they take the reading, matching and drafting, and route exceptions and high-stakes decisions to people with the evidence attached. The people decide fewer, harder cases, with the facts already gathered.
- Anthropic: Building effective agents (December 2024)anthropic.com
- Gartner: Over 40% of agentic AI projects will be canceled by end of 2027 (June 2025)gartner.com
- Gartner: Agentic AI will resolve 80% of common customer service issues by 2029 (March 2025)gartner.com
- CMS: Interoperability and Prior Authorization Final Rule (CMS-0057-F) fact sheetcms.gov
- FinCEN: Customer Due Diligence (CDD) Final Rulefincen.gov
See it on
your own work.
Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.