{ }Flows as code
Each workflow is a versioned spec with schemas, tools and policies. Changes are diffed, reviewed and rolled back like software.
Agent Flow · by Explore AI Inc. · Irvine, California
Operations on autopilot.
Judgment stays human.
Agent Flow builds and runs AI agents that take the repetitive work off your team: invoices, claims, referrals, orders, quotes and lead follow-up. They work inside the systems you already use. Every step is traced, and every risky action waits for a person.
- ~4 weeks pilot to production
- Shadow mode before any autopilot
- Your cloud or ours
runs today0
touchless0%
median handle time0s
cost / run$0.00
awaiting approval0
carrier invoicesbills of ladingcertificates of insuranceFNOL claimspatient referralseligibility checkspurchase ordersRMAslease applicationsW-9sEOBsprior authorizationsAR follow-upsvendor onboardingquotesdispatch tickets
01 · The film
Five workflows, start to finish, in under three minutes.
Freight audit, patient referrals, wholesale orders, a sales quote and a new web lead. Narrated in English and Mandarin. Every screen is the product you can try below, running on demo data.
02 · Live demo
Skip the slide deck. Watch an agent do the work.
Pick a workflow and press run. The agent reads the document, calls tools, checks your rules, and stops for your approval when money or risk is on the line. You make the call.
inputdocument
agent trace0.0s
outcomeidle
Demo data. Companies, people and documents in this playground are fictional; timings and costs are representative of a production run.
03 · In the browser
No API? It works the website the way your team does.
Terminal and carrier portals, payer portals, government sites, webmail. The agent signs in with credentials from your vault, clicks, types, reads and downloads, and stops for you before anything is submitted or sent.
04 · Use cases
Where agents pay for themselves.
We studied the public websites of 126 Southern California companies, from property managers to medical labs, and listed the repetitive work they describe. Pick an industry to see one of those workflows before and after.
Minutes are typical per-item estimates that show the shape of the work, not measured results. Your teardown measures your own.
Not only lower cost. Faster revenue.
The same agents work the front of the business too: the replies, quotes and renewals that win deals when they go out quickly.
05 · How we deploy
From messy process to production agent in about four weeks.
No platform migration, no rip-and-replace. We start from one workflow your team already runs and the systems it already touches.
-
00
Teardown
30 minutes · free
We map one workflow end to end: volumes, systems, exceptions, and who signs off. You get an automation-potential score and a fixed quote within two business days.
-
01
Build
weeks 1–2
Connectors, extraction schemas and your rules written as code. We build a golden test set from your own past cases, so accuracy is measured, not promised.
-
02
Shadow
week 3
The agent runs beside your team on live work and changes nothing. Every case is scored against what your people actually did.
-
03
Autopilot
week 4 onward
Low-risk cases go touchless. Exceptions reach a person with the evidence already assembled. Thresholds widen only when the numbers earn it.
teardown0%
shadow0% · measured
autopilot wk 1~45%
autopilot wk 6~80% target
Illustrative touchless-rate targets. Your pilot reports the real numbers on your own cases.
06 · Platform
Engineered like infrastructure, not a chatbot.
Flows are versioned code. Policies are testable functions. Every run leaves a trace you can replay. Models are routed per step, so you pay frontier prices only where reasoning is needed.
CONTROL PLANE
approvalsaudit logevals & regression gatescost meterreplayRBAC & SSO
⇄Per-step model routing
Small, fast models extract; frontier models reason over exceptions. Cost per run is metered and capped per flow.
✓Evals before every release
Golden sets built from your history gate each deployment on field accuracy and decision accuracy. No silent regressions.
§Approvals by policy
Dollar thresholds, confidence floors and sensitive fields route work to the right person in Slack, Teams or email, with the evidence attached.
≡Full trace, full replay
Every prompt, tool call, diff and decision is stored. Open any run, see exactly why it happened, and replay it against a new version.
⌘Works where APIs don't
Supervised browser actions handle carrier, payer and county portals that have no API, inside the same guardrails.
# flows/freight-invoice-audit.yaml
flow: freight-invoice-audit
version: 14
trigger:
inbox: ap@yourco.com
match: { attachment: pdf, sender_in: carriers }
steps:
- extract: { schema: carrier_invoice, model: fast-extract, min_confidence: 0.92 }
- tool: tms.get_load # rate confirmation + stops
- tool: terminal.gate_events # real arrival / departure times
- reason: { model: frontier, task: reconcile_accessorials }
- policy: ap_variance # escalates to a person
- act: erp.create_bill
- notify: carrier.dispute_email
guardrails:
pii: redact
max_cost_per_run: $0.25
approvals: slack:#ap-approvals
evals:
golden_set: evals/freight_2026q3.jsonl # 1,200 past invoices
gate: { field_f1: ">=0.97", decision_acc: ">=0.99" }from agentflow import policy, Allow, Escalate
@policy("ap_variance")
def ap_variance(run):
"""Pay small, explainable variances; escalate the rest."""
variance = run.billed_total - run.expected_total
limit = max(150, 0.02 * run.expected_total)
if run.min_confidence < 0.92:
return Escalate("low extraction confidence")
if abs(variance) > limit:
return Escalate(
f"variance ${variance:,.2f} over ${limit:,.0f}",
approvers=["ap-lead"],
evidence=run.citations,
)
return Allow(){
"run": "run_7Hq2x", "flow": "freight-invoice-audit@14",
"duration_ms": 41230, "cost_usd": 0.061,
"spans": [
{ "step": "extract", "model": "fast-extract", "ms": 3810, "fields": 23, "min_conf": 0.96 },
{ "step": "tms.get_load", "ms": 420, "status": 200 },
{ "step": "terminal.gate_events", "ms": 1160, "via": "browser" },
{ "step": "reason", "model": "frontier", "ms": 6240, "tokens": 5812 },
{ "step": "policy.ap_variance", "result": "escalate", "variance": 76.5 },
{ "step": "approval", "by": "ap-lead", "waited_ms": 27900 },
{ "step": "erp.create_bill", "ms": 610, "bill": "BILL-20417" }
]
}run_7Hq2x41.2 s · $0.061
extract
tms.get_load
gate_events
reason
policy
approval
create_bill
notify
machine time13.3 s
waiting on a person27.9 s
tool calls4
tokens5,812
Most of the wall-clock is the human approval, by design: the agent prepares the evidence, a person makes the call.
07 · ROI
Price it against the work, not the seats.
Move the sliders to match one workflow on your team. The estimate uses the Production plan below and a conservative ramp.
Net first-year savings$0
Hours returned / year0
Labor value / year$0
Agent Flow, year one$0
Payback–
Estimate only. Uses the starting Production price scaled by complexity, savings that ramp to the share above over three months, 48 working weeks, and the starting pilot fee credited toward Production. Payback includes the pilot month. Your teardown gives you a quote on your real volumes.
08 · Pricing
Priced to your workflow, not a price list.
No two back offices are alike: volumes, systems, exception rules and risk all differ. Every engagement starts with a free teardown and a fixed written quote, so you see the number and the expected payback before you commit.
Teardown
$0
30-minute working session
- One workflow mapped end to end
- Automation-potential score
- Fixed written quote in 2 business days
Pilot
from $15,000
fixed scope · one workflow · 4–6 weeks
- Connectors to your systems
- Golden test set from your history
- Shadow-mode accuracy report
- Credited toward Production if you continue
Production
from $4,500 / agent / mo
quoted per workflow · annual
- Hosting, models and 24/7 monitoring
- Approvals in Slack, Teams or email
- Monthly eval and savings report
- Changes shipped within 2 business days
Enterprise
Custom
several departments or regulated data
- Deploy in your VPC or on-premises
- SSO, SCIM and custom retention
- Uptime and response-time SLAs
- Dedicated agent engineer and roadmap
what drives your quote
Volumeitems per month, and how spiky
Systemshow many, and API or portal-only
Riskmoney movement, PHI, regulated steps
Exceptionshow often a person must step in
Ways to work with us
A fixed-price build, then one yearly fee for hosting, monitoring, model updates and changes as your process evolves.
We run, watch and improve the agents for a monthly fee. Scope changes ship within two business days.
One-time delivery into your cloud with source code and runbooks, plus optional yearly support.
09 · Security & control
Your data, your systems, your rules.
Deploy where you need
Our US-hosted cloud, your AWS, Azure or GCP account, or on-premises for regulated data.
No training on your data
Your documents are never used to train shared models. Model providers are used under no-retention terms where offered.
Sensitive-data guard
PII and PHI are detected and tokenized before reasoning steps; only the fields a step needs are revealed.
Least-privilege connectors
Scoped credentials kept in your secrets vault. Write actions are allow-listed per flow.
Audit everything
Immutable log of every read, write, prompt and approval, exportable to your SIEM.
Kill switch
Pause any flow instantly and fall back to your human queue, with nothing lost in flight.
10 · FAQ
Questions operators ask us.
Do we need to change our software?
No. Agent Flow works through the systems you already use: APIs where they exist, supervised browser actions where they don't, and your existing inboxes and shared drives.
What happens when the agent is unsure?
It stops. Confidence floors, dollar limits and sensitive fields are written as policies; anything outside them goes to a person with the evidence and a recommended action already prepared.
How do we know it's accurate before it goes live?
Two ways: a golden test set built from your past cases gates every release, and a shadow week compares the agent with your team on live work before it is allowed to act.
Will this replace our staff?
It replaces the copy-paste. Most teams use the hours to absorb growth without new hires, clear backlogs, and move people to exceptions and customer work. How you use the capacity is your call.
Which models do you use?
We route per step across leading commercial and open models, and can run open models inside your environment when data cannot leave it.
Who is behind Agent Flow?
Explore AI Inc., an Irvine, California research and product company building adaptive agents. See explore-ai-inc.com.
11 · Get started
Bring one workflow.
Leave with a plan.
Tell us what your team does over and over. In 30 minutes we'll map it, score it and tell you honestly whether an agent will pay for itself.
- emailcontact@explore-ai-inc.com
- based inIrvine, California · on-site across Orange County, LA and San Diego
- companyExplore AI Inc.