AI harness · AI governance · Zurich and remote
You correct an AI mistake once. After that, the correction is a rule.
Anyone new at your company gets house rules, keys only to the rooms they need, and someone who reads their letters before they go out. An AI harness gives the AI exactly that, so it can take on recurring work such as answering client enquiries or preparing reminders. It remembers your decisions and keeps track of open tasks. A person on your team signs off anything binding.
Yannick Seelig, Wetzikon ZH.
- AI review and plan, fixed price
- CHF 1,900
Pilot price CHF 1,390 instead of CHF 1,900 (regular price), 2 pilot places left. Pilot place terms
- AI agents in your business, first workflow
- from CHF 6,900
- Further workflow, fully set up
- from CHF 4,900
- Review is credited
- if the setup is ordered within 60 days
What the AI gets on day one
The AI model writes the text. Your rules, locks and checks sit in the six parts around it, and they stay when the model changes.
01
House rules
Written rules tell the AI what it does, how it writes and what it never decides alone, such as sending something, making a commitment or deleting anything.
For example You strike the sentence "that is certainly deductible" from a draft once. From then on the rules say tax questions go to the client manager, and if the sentence turns up anyway, the check stops it.
02
Keys only for what it needs
The AI reaches only the systems and data its task needs. Everything else is blocked technically, not just forbidden, and before launch I show you, block by block, that access fails.
For example It sees a client's policies, but not the health details from their applications.
03
Check before it goes out
Whatever leaves the company is checked first, and figures and deadlines are matched against their source. Where it becomes binding, a person signs it off.
For example A service charge statement whose total does not match the books does not go to the tenants.
AI model
swappable when a better one comes along
04
Memory for decisions
Decisions, agreements and corrections are written down, and the AI reads them on every task.
For example Mr Meier chose in March not to raise his home contents cover. When he asks again in autumn, the AI knows why.
05
Tasks tracked
Open work is recorded with an owner and a next step, so nothing gets lost overnight or over the holidays.
For example A client does not send their salary statement. After ten days the follow-up sits with the person in charge instead of getting buried in the inbox.
06
A tested playbook per workflow
For every workflow I set up, I write a playbook and test it on finished cases from your daily work. It is included in the price of the workflow.
For example Before the AI prepares its first real rent reminder, the playbook runs over last quarter's reminders, and every result is compared with what went out then.
What changes day to day
Ways of working
Today Everyone asks the AI in their own way, and replies to clients sound different every time.
With an AI harness Everyone works with the same rules and playbooks, and every reply sounds like your company.
Data
Today Staff copy client data into their own AI accounts, and nobody knows what sits where.
With an AI harness The AI works through a company account and sees only what its task needs.
Corrections
Today You correct the same mistake again every week.
With an AI harness A correction one person makes holds for everyone from then on. The check looks for that mistake, and anything it cannot judge with certainty goes to a person.
What goes out
Today You find out whether a reply was right when the client writes back.
With an AI harness The check stops any figure or deadline that does not match its source before the client sees anything.
Holidays
Today Only the person who is away knows what is still open.
With an AI harness The stand-in sees every open task with its next step.
Your AI's job description
Every workflow gets a one-page job description, like a new hire. It shows what the AI does, what it may do and who signs.
Example accounting office, 15 staff (fictional)
Job description · AI for client enquiries
- Duties
- Answer client enquiries following the "enquiries" playbook, tested on 40 finished cases.
- House rules
- Name deadlines only with their source. Tax questions go to the client manager.
- Keys
- Reads the contact mailbox and the deadline database. Payroll and staff files are blocked.
- Check and signature
- Every deadline is checked against the deadline database. Replies that state amounts need a joint signature with R. Brunner.
- File notes
- Agreements with clients, such as who wants documents by post only.
- Open items
- Unclear enquiries sit with the client manager with a due date, visible to the stand-in too.
No sole signature
Offers and prices
You start with the review at a fixed price. Afterwards you decide whether to go on and with which workflow, and the plan states its price.
Step 1
1 to 2 weeks
AI review and plan
CHF 1,900
Pilot price CHF 1,390 instead of CHF 1,900 (regular price), 2 pilot places left. Pilot place terms
Every project starts here. Credited up to the price of the setup if you order it within 60 days.
With your team, I find the work that repeats every week, build a prototype for the most valuable item and write the plan for the harness.
- A workshop with your team and a ranked list of recurring work by hours saved
- A prototype for the top item, built on your real data
- Which data goes where, where it is processed, subscription or API, and roughly what the AI model costs per workflow
- How much time your team needs to set aside for the setup
- Draft house rules, access, checks and the test cases for the first workflow
Step 2
2 weeks
AI agents in your business
from CHF 6,900
Pilot price from CHF 5,190 instead of from CHF 6,900 (regular price, base scope), 2 pilot places left. Pilot place terms
The first workflow, built on the plan.
I set up all six parts for the first workflow and test them on your own cases.
- All six parts, in your accounts at the handover
- Playbook, tested on finished cases
- Trial run on real cases in the second week
- Onboarding for the people who sign off
Step 3
2 to 4 weeks
Further workflow, fully set up
from CHF 4,900
Pilot price from CHF 3,690 instead of from CHF 4,900 (regular price, base scope), 2 pilot places left. Pilot place terms
After the first workflow, each billed separately.
It uses the same rules and memory as the first and gets its own playbook. How long it takes depends on how many new systems it connects.
- Its own playbook, tested on your cases
- Access and checks for this workflow
- Tied into the shared rules and tasks
You pay the provider for the AI model directly, as a business subscription or per use through the API, and I estimate that cost per workflow in the plan. Maintenance from CHF 950 per month is optional. With it I adapt the rules and playbooks when your business or the AI model changes.
From the first call to the trial run
- 0130 minutes, free
First call
You tell me where your team uses AI today. I tell you which workflow to start with and whether the review is worth it.
- 021 to 2 weeks
AI review and plan
In the workshop your team shows me the workflows. I rank them by hours saved, build a prototype of the top one on your real data and write the plan for the harness. Management then decides whether to go on and which workflow comes first.
- 032 weeks
Setup and trial run
I set up the first workflow. In the second week it works on real cases, and it counts as live once it passes the test cases agreed in the plan and every case the AI does not handle itself reaches the right person.
- 042 to 4 weeks per workflow
Further workflows
You decide on each further workflow separately. It gets its own playbook and job description.
You work with me directly
I write the plan, set up every workflow myself and stay your contact afterwards if you take the monthly maintenance.
On your side it takes one person who decides and one who knows the work. The plan says how much time that second person needs, and with them I pick the cases I test the playbooks on.
The harness runs in your accounts and is yours. Rules and playbooks are written in plain language, your team can read and change them, and everything keeps running without the monthly maintenance. Because I do every setup myself, I take on only a few at a time and give you a firm start date in the first call.
Who the AI harness is for
A fit
- Companies where AI is already part of the day, but everyone uses it differently.
- Businesses with recurring cases that need specialist knowledge, such as an accounting office, an insurance broker or a property manager.
- Management that wants to give AI more work and first wants to know what it may see and who checks it.
Not a fit
- The AI should write to clients directly, with no check. I do not set that up.
- You want one workflow automated, such as connecting two programs, and need no rules for a whole team. That is what the AI automation offer is for.
- Nobody in the company wants to sign off replies or own a workflow. Then I tell you so instead of building the harness anyway.
Common questions about the AI harness
What is an AI harness?
An AI harness is the working setup that makes an AI model reliable in a company. The term comes from software development. The harness sets the rules the AI follows and what it may access, checks what goes out, records decisions and open tasks, and gives every workflow a tested playbook. The model writes the text, and the harness makes sure it fits your company.
What does an AI harness cost, and how long does it take?
The AI review and plan costs a fixed CHF 1,900 and takes 1 to 2 weeks. The first workflow (AI agents in your business) costs from CHF 6,900 and takes 2 weeks, each further workflow from CHF 4,900 in 2 to 4 weeks. Maintenance from CHF 950 per month is optional. If you order the setup within 60 days, I credit the amount paid up to the price of the setup.
What does it cost to run, and what changes for my team?
You pay the provider directly for the model, and the plan estimates that cost per workflow. It also says whether your existing AI subscriptions cover it. Your team keeps working in its mailbox, drive and business software.
Does the AI run on a subscription or on the API?
To keep costs down, I usually recommend a business subscription from the provider, with a licence for each person who works with the AI. On the API the company pays for each call separately. In return a workflow sends only the data it needs, and nobody copies anything into a chat window by hand. Where data protection or professional secrecy calls for more control, or a workflow runs at night on its own, I use the API. I set what fits your business per workflow in the plan.
Which data leaves the company?
In the finished setup, what a workflow sends to the AI model goes to the provider through an account in your name. The provider's terms exclude training on your data by default. Everything else stays in your systems. In the plan I record for each workflow which data goes where and where it is processed, so you can check it against your confidentiality duties and data protection law. For the tests I work with access you give me and withdraw after the project.
What happens when a provider retires its AI model?
Rules, playbooks and checks stay. A successor model takes over a workflow only once it passes that workflow's test cases.
Do I need an IT department for this?
No. If a workflow runs in the background, at night for example, a cloud service or an office computer is enough. Anything binding still waits for your sign-off, and I show two people how to stop and restart the workflow.
What do you do, and what does my team do?
I write the plan, the rules and the playbooks, set up access and checks and test everything on your cases. Your team brings finished cases, signs off what is binding, and changes rules itself or, with the monthly maintenance, tells me what should change.
Where should your AI start?
Bring a case the AI gets wrong today. In a free 30-minute first call I tell you which workflow to start with and what the review covers.