We start with a fixed-price pilot of 6 to 10 weeks on a single workflow.
We judge the pilot against a figure you already report, such as cycle time or exception rate. The time your staff spend checking and correcting the agent counts against the result, which stops an agent that merely moves work around from looking like a success.
What makes a workflow a good candidate, when we advise against a pilot, and the systems it can connect to.
We look for a workflow that takes your staff at least 5 hours a week. Below that, checking the agent's output can eat up everything it saves, and we'd rather tell you before you pay for a pilot than after. Most pilots connect two or three of your systems, often the ERP alongside Checkmk or your SIEM.
A workflow is a good candidate when
The figure it should move is already in your reporting
You can pull 20 to 50 past cases from it, including the difficult ones
Its process owner can give two mapping sessions of 90 minutes
We advise against a pilot when
A system involved cannot give the agent an identity of its own
A person must make the decision by law and there is no preparation work to take over
Systems a pilot can connect to
ERP
CRM
TMS
Checkmk
SIEM
PIMS
Grafana
Document stores
Internal APIs
Databases
Engagement sequence
Pilot, 6 to 10 weeks
Stages
Each stage in order with its typical duration, and the point where invoicing starts.
The pilot runs through the first five stages, and we start invoicing when you accept our offer at the end of week two. The durations are what we typically see with two or three connected systems. With more systems, expect the upper end of each range.
We map the workflow with you and pull three months of your figure.
Those three months become the baseline we measure the pilot against. Your 20 to 50 past cases become the evaluation set, and we run every change against it, which is how we catch a fix that quietly breaks a case that used to work.
You receive
A workflow map and the baseline figure
Your team
Two mapping sessions of 90 minutes with the process owner
20 to 50 past cases, including the difficult ones
An hour with a specialist for each system involved
Marbell
Maps how staff and the agent will authenticate to each system
Lists the decisions that must keep a human approval step
A lead engineer and an integration engineer. The lead engineer writes your first reply and stays your contact until the review, so you never have to explain the workflow to someone new.
We work remotely, come on site for the mapping sessions, and speak German and English.
Your security team reviews the data path before you commit.
Your security team gets the architecture diagram and a written data-flow description first. Then we send the written offer. It sets the target figure, the fixed price, a stop date and the agent's access level in each system. Because the target comes from your own reporting, nobody has to argue at the review about what success meant.
We quote Gate inside the project offer, and after the pilot it carries a yearly licence and support fee per deployment. Whether you need it at all depends on where the agent runs and which of your systems it has to reach. The Gate page explains how that works.
Yes. Tell us what you need, and within 2 business days we send a written change note with the new price and stop date for you to accept.
Can you name us as a customer?
Our GTC let us say that you are a customer, but a signed offer takes precedence over the GTC. If you'd rather stay unnamed, we write that into the offer.
Three or four of your staff do this during the build as part of their normal work. We time their corrections from the first week of testing and send you a weekly note that puts that time next to the figure. You'll see early on whether the agent is saving work or just moving it somewhere else.
You receive
An evaluation log for each week of the build
Your team
About 22 to 42 hours from the reviewing staff in total
A 30-minute check-in each week with the process owner
Marbell
Keeps the agent on copies of your records until Approve and deploy
Runs the evaluation set of past cases after every change
Expect about 40 to 65 hours from your team across the whole pilot, assuming two or three connected systems. Most of it falls on the reviewing staff during the build, and we only need your system administrators in the Approve and deploy week.
Staff time your team supplies during a pilot, by role
Your team's time
About 40 to 65 hours in total, most of it during the build
RoleWhenTime
Process ownerFrom Understand to ReviewAbout 5.5 to 7 hours
System specialistsUnderstand1 hour for each system
Security reviewerDefineAbout 4 hours
ApproverDefine and ReviewAbout 2.5 hours
Reviewing staff (3 or 4)BuildAbout 22 to 42 hours, shared between them
System administratorsApprove and deploy2 hours for each system
Your administrators give the agent its own identity in each connected system.
Because the agent never uses a staff login, you can see exactly what it can reach and revoke it. The administrators we train that week can pause the agent without touching anyone else's account.
You receive
A handover pack with these contents:
Runbook
Access register
Architecture diagram
Evaluation set
Configuration
Prompts
Your team
About 2 hours from an administrator for each system
An administrator who can pause the agent and fall back to the manual process
Marbell
Switches the agent from copies to your live records
Runs the evaluation set on the live setup
Trains your system administrators to pause the agent, within their 2 hours for each system
At the end of those two weeks we review the result with you. For 30 days after release, the engineers who built the agent handle its incidents. Whoever picks up a problem already knows which cases gave the agent trouble during the build.
You receive
An evaluation report, signed off by your approver at the review
Your team
A 90-minute review meeting with the approver and the process owner
Marbell
Gives a first response to a critical incident within 4 business hours
Lists the cases the agent passed to a person, with the reason for each
Reports the checking time and the monthly running cost
How checking time is counted against the time saved, and which figures a pilot can be judged on.
We take the correction time from the build and the two live weeks, subtract it from the time the agent saved, and compare what's left with the baseline from Understand. If the agent saves an hour but needs forty minutes of checking, the report says twenty minutes, because that's what your team really gets back.
If the pilot misses its target, the revision round is already in the price. The revised agent runs for two more weeks on live work before a second review. If that review misses the target too, you can end the pilot there, and the Stop terms apply.
Target met
Your team can take the agent over with the handover pack, or we can run it for you under a monthly support agreement with a 30-day notice period.
If the agent runs on Marbell Cloud, the pilot server moves to production without a rebuild or a new contract, and hosting is invoiced monthly.
Target missed
The revision runs on the same evaluation set as the first review, so the second review compares like with like and the revised agent can't pass on easier cases.
Stop
We return your sample records and delete our copies within 30 days.
You keep the evaluation report and the workflow map.
Your team removes the agent's identities, and we shut down anything running on Marbell Cloud.
Outcomes of the pilot review and what each means for you
Response times, support hours, how updates are tested, and the notice period.
Critical incident
First response within 4 business hours
Other requests
Response within 1 business day
Support hours
Monday to Friday, 08:00 to 18:00 Swiss time, in German and English
Model and prompt updates
Tested against the evaluation set before release
Notice period
30 days
Expand
After the review
Each further system needs about 3 hours of your team's time.
Systems the pilot already connected need nothing more from your administrators, and another team can reuse the connections to the systems both teams share. That's why a second workflow usually starts faster than the first.
The reply comes from the lead engineer who would run your pilot, with a proposed time for a first call and any questions about your systems that we need answered before it.