Book a call
Max Tymoshyn, founder of Norml Studio
Max TymoshynFounder, Norml Studio

AI workflow design and implementation, with failure handling and tests built in

We take one process your team wants to run with AI and turn it into a working workflow. First we settle what starts it, which inputs it reads, what the AI step may decide alone and what happens afterwards. Then we build it and test it on real examples, including incomplete and failed ones, before it handles live work.

A few clients we’ve worked with

99Legal / Kremenchuker Law GroupAstraZenecaCoxitDiet vs DiseaseFlorida seal — Regional Counsel projectMetroplex HomebuyersNasdaq Private MarketOffice of Criminal Conflict and Civil Regional Counsel, Second Region of Florida — Ita M. NeymotinPolaclubThe Farm Soho

What we do

Norml designs, builds and hands over one AI workflow at a time, together with its test set and documentation. After launch, any change goes through the same tests before it reaches the live workflow, whether your team makes it or we do.

What one AI workflow includes

Nothing is connected until your process owner approves a written specification. It decides what the build contains and what the tests check.

Solution
Automations
Tools
n8n, Make, Zapier, Claude
Trigger
An agreed event starts the workflow, such as a submitted form, an incoming email or file, a new or changed record, or a scheduled run. It passes the inputs named in the specification, and incomplete inputs stop at a check before the AI step.
Outcome
The AI step returns one of its allowed results, uncertain or high-impact cases go to a named reviewer, and the actions that follow run with retries, duplicate protection and an alert to the step's owner if something fails. Every run leaves a record the team can check.
  • A written workflow specificationIt names the event that starts the workflow, each input and where it comes from, the output and the person who owns it. It also says what counts as an incomplete input and where such a case goes, so gaps are handled by a rule instead of being guessed by the model.
    01
  • An AI step with set limitsThe AI step does one job, such as classifying a request, reading fields from a document or drafting a reply, and may only return results from an agreed list. It works from the source material you approve. A person decides when the model is unsure, or when a result would reach a customer or change a financial record.
    02
  • Downstream actions with failure handlingEach action after the AI step, such as updating a CRM record, creating a task or sending a message, retries temporary errors a set number of times. A run ID stops a repeated event from creating a second record or message. When retries run out, the run stops in a visible state and the step's owner gets an alert.
    03
  • A build on n8n, Make or ZapierWe choose the platform during design, based on the systems it must connect, where the data may go and who will maintain it. The AI step runs on Claude unless the project needs another provider. Accounts and API keys stay with the project.
    04
  • A saved test setReal examples, each with the result a person would expect, including incomplete, conflicting and repeated inputs. The workflow runs against the whole set before launch and again after every change, so a fix in one step does not break another unnoticed.
    05
  • Documentation and run recordsA diagram and step-by-step description of the workflow, with its prompts, rules, known limits and how to pause it. Each run is logged with its input, the AI result and the actions taken, so anyone checking a complaint can see what happened.
    06
Josiah BallCo-Founder & CEO, HOMEEC

“Norml Studio's professionalism and responsiveness were impressive.”

Clutch review
Lucas SeyhunCEO, The Farm Soho

“They were proactive in suggesting improvements we hadn't considered.”

Clutch review
Gaius CooperDirector, Streamline Pump Solutions

“They delivered what they said they would, on time.”

Clutch review
Emma EstefaClient Success Manager, Metroplex Homebuyers

“They focused on how to make our daily operations easier, not just on building features for the sake of it.”

Clutch review
Bret MillerManager, Door Knock Pro

“They understood that the app needed to be practical for someone sitting in a car, not just look good on a desktop.”

Clutch review
Saima AhmedFounder, GoPetZen

“Their level of dedication to the success of the project and GoPetZen's business as a whole is unique.”

Clutch review

What your team provides before the build

  • A process and an ownerThe process the workflow should take over, and one person who can settle its rules and accept the result. If you deliver the project for your own client, tell us who approves on their side.
    01
  • Real examplesReal inputs with the outcome a person produced for each, including the awkward ones. Remove personal details first where your data rules require it, and confirm we may process the rest.
    02
  • Invited access to connected systemsInvitations to the systems the workflow reads from and writes to, and to the automation platform if the project already uses one. Every account stays under the project's control.
    03
  • A model account in the owner's nameAn API account with the AI model provider, held by the business that will run the workflow, so usage, billing and data settings stay with that business.
    04
  • Review and alert rulesWho reviews uncertain results, who receives an alert when a step fails, and which actions must never run without a person's approval.
    05

The five stages of one AI workflow build

  1. [Write the specification with the owner]

    We go through the current process with its owner and the real examples, then write the specification: starting event, inputs, AI decision, actions, review points and failure paths.

    Step 01
  2. [Test the AI step on its own]

    Before any system is connected, the AI step runs on the saved examples and its results are compared with the expected ones. This settles what it may decide alone and what goes to review.

    Step 02
  3. [Build with test records]

    We connect the systems using test records or a separate workspace, so no customer receives a message and no live record changes during the build.

    Step 03
  4. [Run the failure cases]

    We send repeated events, missing fields and rejected updates through the workflow and confirm that retries, duplicate protection and alerts behave as the specification says.

    Step 04
  5. [Check the first live runs together]

    The workflow goes live while its owner checks the first runs with us. You receive the documentation, the test set and a walkthrough of how to pause, check and change it.

    Step 05

The people who do this work.

Meet the whole team

No account layer between you and them. These are the Norml specialists this work runs through.

Max Tymoshyn, founder of Norml Studio

Let's chat

“I've spent 10 years working across development and marketing. I started as a developer, but design, SEO and advertising were already part of the job. That taught me to treat a website as one system: the business, the content, the design and the code have to work together. At Norml, I lead the architecture behind our work and stay involved wherever strategy, design, development and search meet.”

Max TymoshynFounder & Architect

From the Norml blog.

Questions about an AI workflow build

Do we need a finished workflow design before we start?
No. Bring the process, a handful of real examples and the result you expect. Writing the specification is the first part of the work. If the examples show the AI step cannot handle part of the process well enough, that part stays with a person or gets a review step.
Can you use the automation platform we already have?
Often, yes. We check whether it can retry a failed step, keep a log of each run and hold a case for review. Where it cannot, the specification says so and names where that part should run.
Can the AI step use a model other than Claude?
Yes, when the project already pays for another provider or its data rules require one. That model runs against the same saved examples before it is connected to live work.
What happens when a step keeps failing?
The run stops with its input and the error attached and stays visible until someone deals with it. After the cause is fixed, for example an expired connection or a changed field in the destination, the owner resubmits the run, and actions that already succeeded are not repeated.
Will the AI get every result right?
No. Checks after the AI step catch missing or malformed values, unsure cases go to a person, and the test set shows before launch where the model gets things wrong. We share those test results; we do not promise an accuracy figure.
Can you build the workflow inside a project we deliver for our own client?
Yes. We work in the accounts the project uses, through invitations from whoever controls them, and follow your process for reviews and approvals. The documentation names who owns each part after handover, so your client's team knows who handles a failed step.
Can you keep improving the workflow after launch?
Yes, as separately agreed work. A change such as a new category, a revised prompt or another destination is tested against the saved examples before it replaces the live version, and the documentation is updated with it. Your team can make changes the same way.
Where does our data go?
Before the build we agree which fields the AI step may read, and anything it does not need is left out. Data passes through the automation platform and the model provider on accounts the project owns, under those providers' terms. Your business accepts those terms and stays responsible for its data-protection obligations; we review the data settings with you before live data is used.
Does this include choosing which processes to automate?
No. Choosing which processes across a business to automate is a separate assessment. Policies, prices and approval rules come from the process owner; the workflow applies them but does not set them. A workflow that needs a new application or a customer-facing interface is scoped on its own.

Read the full automation case studies

The case studies show these automations within the wider work we did for those clients.

See all case studies