The AI demo is easy. Getting it approved is the hard part.
Most firms have tried AI agents and few have scaled them. The reason is rarely the AI. It is what the firm needs before an agent touches real files: limits, approvals, required steps and a record.
AI agents stall between the demo and real use because a firm has to answer for what they do. Before an agent touches client files, the firm needs four things: limits on what it can change, a person approving anything that leaves the firm, required steps it cannot skip, and a record of every check. Build those first and agents get approved.
Why do AI agents stall after the demo?
The demo shows an agent doing a job in minutes. Then someone asks what happens when it is wrong, who approved what it sent, and how the firm would show a regulator. Those questions usually have no answer yet, so the project waits. The AI was never the hard part.
The numbers say the same. In IBM’s 2026 Tech Leader Study of 2,000 technology executives, 77% said AI adoption is outpacing their governance, and only 11% said they feel prepared for the scale of AI agent deployment coming in the next year. In McKinsey’s 2025 State of AI survey, 62% said their organizations are at least experimenting with agents, but in any given business function no more than 10% said they are scaling them.
What does a firm need before an agent touches real files?
- Limits on what it can change. An agent that can read a file but not change it is a different risk from one that can. Start with read-only access, and widen it one action at a time.
- A person approves anything that leaves the firm. Emails, texts, forms and ads are drafted by the agent and sent by a person. The agent never sends on its own.
- Required steps it can’t skip. If a disclosure has to be signed before advice, the file can’t move to advice until that step is recorded. Moving it anyway takes a stated reason, and the supervisor sees it.
- A record of every check. What was checked, what was caught, which rules applied, who approved, and who overrode a step and why. That record is what an examiner or auditor asks for.
Why keep the rules outside the agent?
Agents change. The model behind them changes every few months, and a team may run several. If the firm’s rules are written into each agent, every change means re-checking every agent. If the rules live in one place outside them, every agent reads the same rules, and changing a rule changes it everywhere. That is how external guardrails work.
How does Broker AI do this?
- Drafts are checked before anyone sees them: missing required wording, banned promises, a licence number left as a placeholder, a commercial email without an unsubscribe line. Signs of payment-redirection fraud stop the draft for a person to decide.
- File steps are a gate. The brokerage and the broker set the steps each file goes through. A file can’t move past a required step that isn’t recorded, and an override is logged with its reason.
- The agent works on the broker’s own AI account. Broker AI keeps the rules, the steps and the record, never the client file.
- Every check goes on record, with the time, the channel, the rules that fired and the policy version.
We’re building agents for a whole mortgage file on top of this now. What we’re building.
General information, not legal or compliance advice. Rules differ by state, province and regulator; check with your compliance team or counsel.
Questions people ask
- Why do most AI agent projects fail to scale?
- Usually not because the AI can’t do the work, but because the firm can’t yet show who approved what it did, what it was allowed to change, and what was checked. Surveys find adoption running ahead of governance.
- What does human in the loop mean for AI agents?
- A person approves each action that leaves the firm, such as a client email or a form, before it goes. The agent prepares the work; the person decides.
- Should an AI agent have write access to our systems?
- Start with read-only access, then add the specific actions the agent needs, one at a time, each behind an approval. Broad write access is what stops legal review.
By DisruptionIO, the team behind Broker AI. Published October 28, 2026.