Safety
AI Agent Safety and Governance
DH79 treats AI agent safety as an operating model: scoped permissions, least-privilege access, human approval gates, draft-only modes for sensitive workflows, logs, monitoring and clear escalation rules. The aim is to make agents useful for research, drafting, preparation, admin and monitoring without handing over legal, financial, medical or commercial judgement.
Who this is for
- Businesses connecting agents to email, CRM, documents or calendars.
- Teams in professional services, healthcare, finance-adjacent or client-sensitive work.
- Founders who want speed without losing control.
The business problem
AI agents become risky when access, approval and accountability are vague. Safety must be designed before agents touch business systems. The important test is whether the work is frequent enough, valuable enough and controlled enough for an agent to help without hiding risk. DH79 starts with a narrow workflow because useful agents need clear inputs, clear outputs and a named human owner.
Example workflow
Define the workflow, decide what data is needed, grant only that access, keep sensitive outputs under human approval, log what happens and review exceptions monthly. The workflow is designed so the agent prepares, drafts, summarises or monitors, while a human remains responsible for approval where judgement, reputation, compliance or customer trust is involved.
What DH79 sets up
- Permission and access map.
- Human approval gates and draft-only modes.
- Logging, monitoring and review rhythm.
- Plain-English operating rules for the team.
What the AI agents can do
- Prepare drafts and summaries.
- Flag gaps and exceptions.
- Monitor repeated processes.
- Support human decisions rather than replace them.
What tools they can connect to
- Gmail, Outlook and shared inboxes
- Google Workspace, Microsoft 365, Notion, Drive and SharePoint
- HubSpot, Pipedrive, Salesforce or lightweight CRM systems
- Slack, Teams, calendars, task tools and internal knowledge bases
- Website CMS, spreadsheets, forms and reporting dashboards where access is scoped
What stays human
- Legal, financial, medical, HR and regulated decisions.
- Final approval of sensitive external messages.
- Commercial judgement and customer relationships.
DH79 deliberately avoids promising fully autonomous business judgement. The safest commercial gains usually come from agents preparing the work, making gaps visible and giving humans better drafts, summaries and reminders.
First 30 days
- Map the safety boundary.
- Launch with limited access.
- Review outputs and logs.
- Tighten or expand controls based on evidence.
Safety and GDPR-aware controls
- Least-privilege access.
- Human approval for reputation-critical work.
- Logs and exception review.
- Escalation rules for unusual or sensitive outputs.
Pricing and scope
DH79's managed package starts from £5,000/month inside an agreed operating scope. Work that needs unusual volume, specialist integrations or regulated review is scoped before launch so costs and responsibilities are clear.
How to judge whether this should be your first agent
A good first agent is not the most exciting idea in the business. It is the workflow with clear inputs, repeatable steps, visible mistakes and a human owner who can approve the output. For ai agent safety and governance, DH79 looks for a task where the agent can prepare drafts and summaries, connect only to gmail, outlook and shared inboxes, and leave legal, financial, medical, hr and regulated decisions with a person. That makes the pilot easier to measure and safer to improve.
- Bring two or three real examples of the current workflow, including a strong example and a messy edge case.
- Decide who owns approval, who receives the draft or summary, and what would count as a useful first-month result.
- Start with a draft, research, preparation, triage or monitoring task before allowing any agent to take external action.
FAQs
Can DH79 set up ai agent safety and governance without our team managing prompts?
Yes. DH79 maps the workflow, builds the agent instructions and private workspace, connects the agreed tools, sets approval rules, monitors usage and improves the system. Your team should understand the operating rules, but it should not have to manage tokens, hosting or prompt maintenance.
What should stay under human approval?
External messages, legal or financial commitments, sensitive client communication, medical or regulated judgement, unusual edge cases and anything that could affect reputation should remain human reviewed unless a narrower approval policy is agreed.
How quickly can the first workflow go live?
A narrow first workflow is normally designed during the first month. The first 30 days focus on workflow audit, data and tool access, agent build, controlled testing, team feedback and a decision on what to improve or add next.
How does DH79 reduce risk?
DH79 uses scoped permissions, least-privilege access, human approval gates, logs, draft-only modes for sensitive work, clear escalation rules and monthly review. The aim is useful operational leverage without handing important judgement to an unsupervised system.
Is this suitable for uk businesses that need safe managed ai agent workflows?
It is most suitable when uk businesses that need safe managed ai agent workflows have repeatable research, drafting, preparation, follow-up, admin or monitoring work and want a managed service rather than a DIY platform. If the first use case is too vague, DH79 starts by narrowing it into a controlled pilot.
Want to know which AI agents your business should build first?
Book a short AI operations call and we'll map the fastest, safest starting point for your business.
Book the AI operations call