Home / Products / Process Runner

Process Runner™

Turn how your work gets done into a process that runs.

Process Runner turns a workflow into a process file: a list of steps for AI, for people and for other systems. Each step finishes only when its evidence is in, and every run keeps a record of who did what, with which AI, at what cost.

Start a Conversation See examples

Run record Complete

Process: Idea Review, version 1.4.0
Run: inst_idea_review_8c41d2e0
Steps: 13 of 13 complete
AI models: 4 vendors: Anthropic, OpenAI, Google, xAI
Decided by: The operator, signed in: 4 ideas carried forward
AI cost: $0.41 across 11 AI steps
Time: Oct 27, 2026, 2:14 to 2:31 PM ET
Ledger: Every entry hashed and linked to the one before it

Example run record. Every run keeps one.

The idea

A process is a file.

Most work already follows a process, but it lives in people's heads, in email threads and in whichever chat window was open that day. Process Runner writes it down as a file: the steps, who or what does each one, what each step has to produce, and what happens next.

The file is portable and versioned, so a process runs the same way for one person or a whole team, and it gets better as you change it. You build it in plain language.

Process Runner overview: inputs such as documents, spreadsheets, databases and APIs flow through defined, routed and human-reviewed process steps into structured data, reports, approvals and an audit trail

How it works

Four ideas, every run.

1
Write the process

List the steps, the roles, what each step must produce and what happens next. A step in Process Runner is called a gate.
2
Each gate does its part

An AI model, a person, a prebuilt agent or another whole process does the work of a gate. Some gates run side by side.
3
Nothing moves without evidence

A gate completes only when what it must produce is in: an answer, a file, a signature, an approval. Then the gates that depend on it unlock.
4
Every run is on record

Who did what and when, which AI model, how many tokens and what it cost, and every approval. Memory carries what was learned into the next run.

The building blocks

Five kinds of gates.

AI gate

An AI model reads what earlier gates produced and writes this gate's evidence: a summary, a decision, extracted data, a draft.

Person

Human gate

Waits for the right signed-in person to answer, approve or sign. Nothing that needs a person is completed by an AI.

Agent

Agent gate

A prebuilt agent does a known job: reads a PDF, searches the web, buys a domain, sends the result where it belongs.

Process

Sub-process gate

Runs another whole process as one step, so a good process can be reused inside a bigger one.

Parallel

Parallel gate

Runs several pieces of work at the same time and waits for all of them before the process moves on.

Examples

Processes you can run.

Review an idea with four AI vendors

Before planning anything, AI models from four vendors each look at an idea from a different angle, one after another, and a person decides what goes forward.

AI Load what past reviews learned
Person Describe the idea and the goal
AI Four vendors, four lenses: capability, experience, integration and new uses
AI Rank the ideas and group them into phases
Person Pick the ideas to carry forward
AI Save the learnings to memory

Result: a ranked slate of ideas from several vendors, and the person's pick, on record.

A customer asks for a refund

A customer asks a store's Relay helper for a refund. Two AIs from different vendors check it, the customer's account age and history count too, and a person decides only the close calls. The same process can open a new account.

Person The customer asks Relay for a refund
AI Relay gathers the order, the amount, the reason, and the customer's account age and history score
Parallel Two AIs from different vendors each score it against the refund policy, with their reasons
Agent All three pass: the refund is issuedBoth AIs score 90% or higherAccount at least 6 months oldHistory score 80 or higher

Person Any one falls short: a person approves or declines, seeing the scores and every reason
AI Relay tells the customer the outcome

Result: clear refunds go through on their own, close calls go to a person with the reasons in hand, and every score and decision is on record.

Door takeoff from construction plans

An estimator's door takeoff, read from the drawings and the specifications instead of typed in by hand.

Person Upload the plans and specifications
AI Read the door schedule, opening by opening
Agent Apply the estimating rules and cross-check schedule against spec
Person Approve every line that was flagged
Agent Export the estimating spreadsheet

Result: the takeoff in the estimator's own spreadsheet layout, with every flagged line signed off.

An AI agent pays for what it uses

An outside AI agent buys something from a site, inside the site's rules, and both sides keep a receipt.

Agent The agent says who it is and who it acts for
Agent The site's terms and price are checked
Agent Funds are held from the agent's wallet
Agent The work is delivered and the funds are released
Agent Both sides get a signed receipt

Result: paid, on the site's terms, with a receipt either side can check. More about Agent Payments

Inside a process

One refund, gate by gate.

This is the refund example above. Each gate says who does it, what it must produce and what it waits on. Two AIs from different vendors score the request separately, so no single model decides alone. When both score 90% or higher, the account is at least 6 months old and the customer's history score is 80 or higher, the refund goes through. When any one falls short, it waits for a signed-in person who sees the scores and every reason.

Change the threshold or the policy and the next run follows it. Every run records the version it ran.

# Refund: two AIs check it, a person takes the close calls

gate 1    The request
  done by       Relay, the helper
  must produce  the order, the amount and the reason
                the customer's account age and history score
                the conversation, private details taken out
  waits on      the customer's request

gate 2    The check, both AIs at once
  done by       two AIs from different vendors, separately
  must produce  each one's score against the refund policy
                each one's reasons
  waits on      gate 1

gate 3a   The refund
  runs when     both AI scores are 90% or higher
                the account is at least 6 months old
                the history score is 80 or higher
  done by       the payments agent
  must produce  the refund receipt

gate 3b   The decision
  runs when     any one of those falls short
  done by       a person on the team, signed in
  sees          every score and every reason
  must produce  approve or decline, and why

gate 4    The answer
  done by       Relay
  must produce  the outcome, told to the customer
  waits on      gate 3a or gate 3b

# On record: every gate, score, reason and decision,
# and the version of the process that ran.

Any AI model

Don't marry the model.

Models change. The process stays the same.

Run a whole process on one model, or give every step its own: a five-step process can use five different models. Process Runner works with the major vendors and the rest, including models running on your own machines. You choose which ones are allowed.

Google Gemini
OpenAI
Anthropic Claude
xAI Grok
DeepSeek
Kimi
Local models

Pin a model to a step, or let the router start with the cheapest model that can do the job and move up when it can't. If a vendor is down or out of credit, the step moves to another. Every call's model, tokens and cost go into the run record. Read why.

Cost your processes

Know what each process costs, to the token.

Every run records each AI call: the model, the tokens in and out, and what it cost, with a running total for the run. Add it up per step, per run or per process, so you know what a process costs before you sell it or scale it.

Set limits. Give a process, or a single step, a cost limit and a token limit. A run that reaches its limit stops and waits for a person instead of spending more.
Fall back to cheaper models. Within a step's limit the router picks a model that fits, starting with the cheapest one that can do the job, and moves up only when it has to.
Daily caps. Automated work runs under daily caps, with a switch to stop it.

Run cost: Idea Review Under limit

Step	Model	Tokens	Cost
Load past learnings	Gemini Flash-Lite	6,120	$0.001
Lens 1: capability	Claude	11,840	$0.092
Lens 2: experience	GPT	12,310	$0.064
Lens 3: integration	Gemini	13,020	$0.041
Lens 4: new uses	Grok	13,550	$0.058
6 more AI steps	mixed	61,900	$0.154
Total	4 vendors	118,740	$0.41

Example run cost. Limit for this process: $1.00 a run.

People stay in charge

Governed, not just automated.

The right person signs

A human gate checks who is signed in, not a name someone typed. The person who asks for something is not the one who approves it.

Stalls get noticed

A step that fails is tried again, and a step that keeps failing is flagged to a person instead of being dropped.

A record that shows changes

Each run keeps a ledger whose entries are hashed and chained, so an edited, deleted or reordered entry shows.

Memory with a source

What a run learns is saved with where it came from and how long it holds, and the next run starts with it.

Your data, your vendors

Choose which AI vendors may process your data. If a step fails over, it never lands on a vendor you ruled out.

Portable and versioned

A process can be exported as a signed package and imported elsewhere, and its version travels with every run.

Built on Process Runner

Our own products run on it.

All products

Put your process on Process Runner

Let's discuss your project.

Basic to enterprise pricing available.

Start a Conversation