The models.
Which providers and models exist in the registry, and which steps may use them.
AGENTFLOW · WORKING WITH US
We send an engineer. You pay for the result.
One of our engineers joins your team, builds the first workflows with the people who own the process, and puts them into production. Then they stay, to run them, watch them and make them better. Before any of it starts, we agree on the number that has to move. That number goes in the contract.
$50,000 in POC credits for qualified use cases with an executive sponsor.
HOW IT GOES
Build it. Run it. Make it better.
Three stages, and one agreement before them. Our engineer is there for all of it. So are your people, which is how they learn it.
- Before day oneAgree on the number
One call with the person who owns the process, the executive who sponsors it, and someone from AI engineering. We pick one process and write down what success means: the number that has to move, where it is measured, and by when.
- Days 1 to 3Build and deploy
Our forward-deployed engineer joins your channels and your standup. They connect your data and your semantic layer, then build the first workflow in chat with the process owner steering. It is in production before the week is out, on real data, with real checks. The next few follow in weeks two to four.
- From the first runRun and monitor
The workflows run on their schedule or their trigger, and our engineer watches every run with your team. Once a week you get one page, with the number we agreed at the top.
- Every week afterImprove
Checks get tightened against real runs. Costs come down. The next process joins the first. By the tenth workflow, your team is writing them in hours.
AFTER IT IS LIVE
Then we stay.
Most projects end at launch. This one is measured after it. Our engineer stays to run the workflows with your team, watch every run, and keep bringing down what each one costs and how long it takes. Every run writes its cost, its time and the result of every check to a ledger in your own storage. Once a week we put that on one page and go through it with you.
WEEKLY REVIEW · WEEK 67 workflows in production
The number held. Two runs stopped themselves. One change is proposed.
The number
June flux package signed one day after close. Agreed: one day or less.
Runs
31 this week, across 7 workflows. 29 delivered. 2 stopped by a check.
Stopped
Account 6120 moved $410K and the journals explained 71% of it. That is under the 90% rule, so it went to a person. It was a misposted accrual. The other was a bank file that arrived late.
Accuracy
96% on your 140 test cases. Agreed bar: 90%.
Cost
$0.31 a run at the median. Cap: $0.40.
Proposed
Run the summary step on a smaller model. 38% cheaper, same score on the test set.
Accept the changeSee the runsNot now
One page a week. The number first, then everything that could move it. Illustrative.
HANDS-ON, FROM THE FIRST WEEK
The best way to learn agents is to ship one.
In the first week our engineer writes and your people steer. In the second they write together. After that your people write and ours reviews. By the tenth workflow your team is writing them in hours, and ours is reading over a shoulder. We stay on the hook for the number. Your team takes over the keyboard.
ONE FINANCE TEAM · THE FIRST TEN WORKFLOWSwho wrote it · time to production
- 01Flux reviewour engineer3 days
- 02Accrual checkour engineer2 days
- 03Intercompany tie-outtogether2 days
- 04Bank reconciliation exceptionstogether1 day
- 05Payables aging reviewyour team1 day
- 06Revenue cut-off testyour team6 hours
- 07Prepaid amortization checkyour team5 hours
- 08Fixed-asset additions reviewyour team4 hours
- 09Vendor master change audityour team3 hours
- 10Board variance summaryyour team3 hours
The byline changes at the fifth workflow. The hours keep falling after that. Illustrative.
OUTCOME-BASED PRICING
The agent has to prove its work. So do we.
Software is usually sold by the seat and services by the hour, and both get paid whether or not anything changes. Our contract is tied to the result we agreed before day one. It starts as one page, signed by your sponsor and ours, naming the number, where it is measured, and the date.
SUCCESS MEASURE · ONE PAGEsigned before day one
Flux review: from five days to one morning, by the second close.
Process
Month-end flux review. Every account that moved more than the threshold, explained from the journals.
Owner
The controller.
Sponsor
The CFO.
Measured
Days from close to a signed flux package, read from your own close calendar.
Must hold
Totals tie to the trial balance. Documented drivers explain 90% of every movement. Under $0.40 a run.
Paid
Build fee due when the number is met. Run fee due only in the months it holds.
Signed · CFOSigned · e6data
Proof of concept funded by $50,000 in POC credits. Illustrative; commercial terms to be confirmed.
The fee for the build is earned when the number is met. After that you pay to keep it met, and the run fee is due only in the months it holds. The ledger sits in your storage, so the measurement is yours. Model spend is billed by whichever provider you route to; compiled workflows call a model only where judgment is needed, so that side of the bill stays small, and you can see it per step.
POC CREDITS
The proof is on us.
If your use case qualifies, we fund the proof of concept. The credits cover our engineer's time and the platform while we prove the result on your data. You bring the process, the people who own it, and a sponsor who can say yes. We take on six of these a quarter, because each one gets a named engineer.
WHAT QUALIFIES
A proof of concept that ends in production.
The credits go to use cases that can show a result quickly and are meant to be kept.
The use case
- A named owner: a process or a metric that belongs to someone who can say what right looks like.
- A result your own data can measure: days to close, hours returned, dollars recovered.
- Work that repeats, every week or every month, so there is something to measure inside the first month.
- Data we can reach: read access to what the process uses, in a deployment shape your security team accepts.
- A plan to keep it: if the number is met, the workflow stays in production.
The executive sponsor
- Leads the function the process belongs to: finance, operations, supply chain, customer success. A vice president or above.
- Signs the success measure. One page, before day one.
- Gives it thirty minutes every two weeks, to see the number and clear whatever is in the way.
- Makes the call at the end. If the number is met, a yes or a no within two weeks.
We ask for a sponsor for the same reason AgentFlow asks for a signature. Someone has to own the result.
WHO IS IN THE ROOM
A small team on each side.
It works when the person who owns the process is in it. Everyone else is there to clear the path.
From you
- An executive sponsor. Signs the success measure. Thirty minutes every two weeks.
- The process owner. A couple of hours a week. They decide what right looks like.
- One or two people from the team that will use it. They are writing workflows by week two.
- Someone from AI engineering or the data platform. They grant access, register models and tools, and review what gets built.
- A security contact. Once, early, to choose the deployment shape.
From us
- A forward-deployed engineer. One named person, the same one from the first call to the last. They work in your channels, on your data, inside your boundary.
- An engagement lead. Owns the plan and the success measure with you.
- The product team, on call. If a workflow needs something AgentFlow does not do yet, they hear it directly.
THE BUSINESS WRITES. ENGINEERING GOVERNS.
What AI engineering keeps.
The tools.
What is registered, which MCP servers are connected, which of their tools are exposed, which hosts can be reached.
The boundary.
Where AgentFlow is deployed, and what leaves it.
The review.
Every workflow is a versioned JSON document in your storage. Read it, diff it, review it like code.
“I've signed a lot of software contracts. This was the first one with my own number in it.”
Questions before you start
What does outcome-based mean here?
We agree on a number your business already tracks, that the process can move, and that your own data can measure: days to close, hours of analysis returned to the team, chargebacks avoided, minutes from event to containment. Our fees are tied to that number. If we cannot find one together, it is the wrong first use case, and we will tell you so.
Who decides whether the number was met?
You do. It is measured in your systems, by the method written on the success measure. The ledger behind it sits in your storage, with every run, every check and every cent.
What if it isn't met?
Then the build fee is not due. We will show you what we learned and what we would change, and you decide whether to go again.
What do the POC credits cover?
Our engineer's time and the AgentFlow platform for the proof of concept, up to $50,000. Model and cloud spend stay on your own accounts. On a compiled workflow they are small.
Why do you ask for an executive sponsor?
A proof of concept ends with a decision, and someone has to be able to make it. The sponsor signs the success measure, sees the number every two weeks, and says yes or no at the end.
What do you need from us to start?
A sponsor, a process owner, read access to the data the process uses, and a security contact to choose the deployment shape. If you already have a semantic layer, we bind to it. If you don't, we define the measures the first workflows need as part of the work.
Can the first workflow be one we already run somewhere else?
Yes. Bring a workflow you already run, on whatever you run it on. We will rebuild it beside your team and put the two receipts next to each other.
Can we take over the running ourselves?
Whenever you like. Our engineer moves to an on-call role. The workflows, the run history and the ledger are already in your storage, and your team has been writing workflows since week two.
Can we run it inside our own cloud?
Yes. The proof of concept runs in the same deployment shape you would use in production: our cloud over a private link, your cloud account, inside your VPC, or on-prem. The architecture does not change between them.
TALK TO US
Bring a process and a sponsor.
We will bring an engineer and the first $50,000. Start with the process that is done by hand, matters every month, and makes someone nervous.
Read the docsThanks.
An actual person replies, usually within one working day.