AI Workflow Security: Protect Your IP Before You Scale

If you run an education or consulting business, your frameworks are what clients pay for. The day agents start running them, your IP has a new way out of the building. AI workflow security comes down to five habits. Lock your accounts. List every agent. Prove where your data came from. Keep a human check on anything that ships. Put your rules in the vendor contract. None of it needs an engineering team to start. All of it is cheaper before you scale than after.

What AI workflow security means when your IP is the product

Once agents run your frameworks, three questions decide whether your IP is safe.

• Who can get into your AI accounts?
• What can each agent read, write and send?
• Can you show, on paper, where your data came from and who approved the output?

If a buyer or an auditor asked tomorrow, could you answer all three?

Start with your accounts, not your agents

The moment you run AI across a team, your AI accounts become a target, the same as a bank login. We learned this the hard way once. Now we treat security as part of the cost of running AI, not an afterthought.

Three fixes, none of them technical:

1. Individual logins. No shared passwords. No shared magic links.
2. Two-factor on everything. Use an authenticator app, not SMS.
3. Hard spend caps. A breach should never be able to run up a bill.

The casual setup most teams run, one login passed around in a chat thread, is exactly what a breach feeds on.

Then look at what your team installs. A third-party skill or plug-in can quietly send data somewhere you never agreed to. Nothing should run until someone has read what it does. The same goes for AI features your team already uses without asking. Njin's piece on shadow AI already running inside your business covers how to find those tools before they find your client data.

Keep one list of every agent you run

For each agent, write down:

1. Who owns it.
2. What it is for.
3. Where its data comes from.
4. Which model it uses, and which version.
5. Who can access it, and how.
6. How much damage a wrong answer could do.
7. When it was last checked.

Do the client-facing agents first. The ones that write proposals, price work or answer support tickets carry the most risk.

This is not our invention. NIST's Generative AI Profile lists it as a governance action: list your generative AI systems and add them to an AI system inventory. It also asks you to set out who reviews what, and how often.

The sheet does double duty. When a buyer asks what AI you run, you show them the list. Our AI agent governance playbook walks through building it from scratch.

Prove where your data came from

Provenance means proof of origin.

CISA's AI data security guidance, published with the NSA, the FBI and international partners, names the core defences:

• Track provenance. Log where data came from and the path it takes through your system.
• Keep a signed ledger. Changes go into a log that is cryptographically signed and can only be added to, never edited.
• Sign your revisions. A digital signature shows who changed a dataset and that nobody tampered with it after.
• Encrypt and store it properly. At rest and in transit.

Poisoning a dataset is cheap. The same guidance reports that enough control of some web-scale datasets can be bought for about $1,000 USD, and in some cases as little as $60 USD. It calls that a viable threat from low-resource attackers.

For most founder-led businesses, the practical version is simpler. Label and version every file you feed an agent. Record who added it and when. Never plug in a third-party model or dataset until the vendor confirms in writing where it came from. Our AI data governance guide covers the setup step by step.

Keep your judgment in the loop

Your judgment is the asset. The AI is an enthusiastic intern. It works hard, and it needs checking.

We run a rule called 10/80/10. That is 10% planning, 80% the AI doing the work, and 10% a human making it right. The two 10%s are where the judgment lives. Nothing goes out without a human yes.

To make that check hold up under pressure:

• Sample, don't skim. Give reviewers a set number of outputs to check each week and a scorecard to check them against.
• Keep a decision journal. Write down why an output was accepted or rejected. Over time it becomes a record of your standard.
• Break the chain. If one agent hands straight to the next, one error flows through the lot. Add a check between steps.

Research points the same way. A 2026 paper in the Journal of Knowledge Management on how generative AI reshapes decisions argues that AI raises the bar on traceability, on being able to contest a decision, and on a human owning the reason for it. A 2025 study on contestable AI in criminal intelligence work built the same idea into a working tool. Investigators could verify and correct each AI prediction, and every version was logged.

Vet vendors before they touch your IP

Four things belong in every AI vendor contract:

1. A written ban on training their models on your data without your consent.
2. The right to have your data deleted on request.
3. The right to audit how your data is handled.
4. A written statement of where anything they supply came from.

The FTC has warned AI companies to keep their promises about customer data, including promises not to use it to train models. It also says that in past enforcement actions it made businesses delete products, models and algorithms included, that were built on data they got unlawfully.

If a tool you depend on was built on data it should not have used, that becomes your problem too.

Watch for personal data leaking through vendor tools as well. Our post on PII in AI covers how to stop it.

Operating controls that catch failures early

You would not hand a new hire every key on day one. Treat agents the same way.

• Least privilege. Each agent gets only the access its job needs.
• Short-lived credentials. Logins that expire fast, so a leak does not last.
• Logs tied to the inventory. Record prompts, model versions, outputs and anywhere data leaves your system.
• A test before go-live, and a way back. Run a quick check before any change ships. Keep a rollback for any agent that touches delivery.

Treat your logs like a flight recorder. You hope you never need them. Our post on layered AI agent security breaks this down agent by agent.

What to do when something goes wrong

Something will break. Decide the plan now.

1. Freeze. Stop the agent. Save every log before anything changes.
2. Contain. Revoke its keys. Block anything it could trigger downstream.
3. Fix. Roll back, find the root cause, and tell affected clients if the rules require it.

A saved log beats a fast guess every time.

Security is a layer, not a project

Bolting security on after launch is the expensive way to do it.

Orchestration means many agents working as one system. So one weak login, or one agent with too much access, puts the whole system at risk. That is why we don't run AI as a set of loose tools. We run an AI Operating System: frameworks, skills, memory and connectors, with security as one of the layers. A daily guard checks the stack for drift and flags problems early. Accounts are locked down. Every skill is vetted before it runs.

We build with Claude Code, and the order does not change. Rules first. Then the AI employees that carry your judgment. It is the approach behind the 90-day program. If you want to see how the layers fit, read how the AI Operating System works.

AI workflow security starts with a list of every agent your business runs.

James Killick

Find out where your risk sits

Locked accounts and a clean inventory are step one. Step two is knowing how much of the business still runs through you, because that is the work your agents will carry.

Take the Founder Bottleneck Assessment. It scores five dimensions in six minutes, then names your next move.

Sources

• AI Companies: Uphold Your Privacy and Confidentiality Commitments, Federal Trade Commission
• Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1
• AI Data Security guidance for data used to train and operate AI systems, CISA, NSA, FBI and international partners
• The invisible hand of generative AI, Journal of Knowledge Management
• Contestable AI for criminal intelligence analysis, Frontiers in Artificial Intelligence