Most companies we meet have staff pasting customer details into personal chatbot accounts, with no policy and no training. We fix that in a few weeks: a policy people actually follow, accounts the company owns, and training built around real jobs.
This is for the owner or operations lead of a company with 10 to 200 people. Nobody decided to adopt AI. Somebody on the team just started using it, then everybody did, and now the company has a tool it depends on and does not control.
It gets signed, filed and ignored, and the pasting continues.
People get faster at putting the wrong things into the wrong tool.
Our stance is tool-agnostic. In practice, most rollouts we run standardize on one approved tool, usually Claude, on company accounts and company email, so the company owns the work and can turn access off the day someone leaves.
Five steps, in this order. Each one makes the next one stick.
Written with you, in plain language, with green, yellow and red data tiers and the approval levels for automations. It starts from our template and gets adapted to your business, not the other way round.
Seats on company email, admin in the company's hands, personal accounts retired for work. Offboarding becomes one switch instead of a hope.
A half day or a full day with the two or three people whose jobs change most, built on their actual work. Not a prompt-writing seminar.
Each role gets a short list: what to use AI for, what never goes into it, and what to do when unsure.
We look at who is using it, for what, and what moved. Then we adjust the training or the policy, not the people.
Staff had been using personal ChatGPT accounts for quotes and customer emails. We wrote their AI use policy, moved everyone onto company-owned Claude accounts on company email so the company owns the work, and trained two key people for a day each on their actual tasks. It started as a single automation build and grew into the policy, the training and several more projects.
We use the same setup we sell: company accounts, a written policy with approval levels, and a monthly read-back of what worked and what did not. We also publish our own site audit and our own AI visibility numbers, zeros included.
Ten pages, written for a company that has never had an AI policy. Green, yellow and red data tiers, an approved-tool slot, a section on client documents, the approval levels for automations, and an acknowledgement page. The PDF is fillable. The Word version is there for your own colours and wording.
No form, no email required. It is a starting point for your own policy, not legal advice.
Every automation we build gets an approval level before it goes live. Drafts, summaries and internal records run on their own, with a log. Anything that goes outside the company, such as an email or a quote, is approved by a person until there is a track record. Anything that moves money, changes where money goes, or changes who has access is approved by a person every time, and that level never relaxes.
People ask whether the newest AI models make that last rule unnecessary. The evidence says no, and the reason has little to do with how capable the models are.
In July 2026, security researchers at Zscaler ran 26 AI models as web browsing agents with payment tools and no spending limits, then pointed them at a page with hidden instructions to pay a $3 fee for a fake API key. Four of the 26 paid (Zscaler ThreatLabz, July 2, 2026). In March 2026, Palo Alto Networks' Unit 42 found live web pages carrying hidden instructions that tell an AI agent to send $5,000 to an attacker's PayPal account (Unit 42, March 3, 2026). Academic testing of Google's agent payments protocol, posted in September 2026, steered the agent's purchases with nothing more than ordinary product text, succeeding in between half and nine out of ten attempts depending on the attack (arXiv 2609.11757).
The bigger problem does not need a trick at all. The second-costliest crime in the FBI's 2025 Internet Crime Report, business email compromise, at just over $3 billion in reported losses, is typically an email asking you to update a vendor's bank details. An automation that reads that email and does exactly what it says has done its job perfectly and paid the fraudster. The control that reliably works is a phone call to a number you already had on file. No model can make that call for you.
Can money ever be automated safely? Yes, with the right steps. A person fixes the payee, the maximum amount and the frequency in advance, those limits are enforced by the payment system rather than by the AI, and the agent cannot change any of them. The person approves the arrangement once instead of each payment. Visa's Intelligent Commerce and Mastercard's Agent Pay are both built on exactly that model: a tokenized credential that carries the user's spending limits, merchant restrictions and approval rules, enforced on the network rather than in the prompt (Visa, Mastercard).
For most small and mid-sized businesses, that infrastructure is not realistic yet, and approving each payment costs a few minutes a week. So the rule applies: if an automation can move your money, a person says yes first, and the approval screen shows the amount, the recipient and what changed.
Put the policy in place before the training, and the company accounts in place before either. Staff need to know which data never goes into an AI tool, and they need a company-owned tool to use instead of their personal one. Then train the two or three people whose jobs change most, on their real work.
We write it with you, starting from our free template and adapting it to your business. It covers which tool is approved, green, yellow and red data tiers, how client documents are handled, the approval levels for automations, and an acknowledgement every employee signs.
Usually, yes. One governed tool on company accounts is easier to train, easier to audit and easy to switch off when someone leaves. Our stance is tool-agnostic, and in practice most rollouts we run standardize on Claude.
A few weeks from the first call to a trained team: the policy, the accounts and the role training. The 30, 60 and 90 day checks follow after that, so the whole arc runs about three months.
Move the work onto company accounts on company email and retire the personal ones for work. It is not a punishment. It means the company owns what gets produced and can turn off access the day someone leaves.
No. It is a starting point written for a company that has never had an AI policy. Have your own attorney review the finished version, especially if you work in a regulated industry.
Request a free 30-minute call. We'll tell you what your policy needs, which seats to buy, and who to train first. No pitch, no pressure.