Good AI work is people-centric.
Agents take on whole jobs. People decide, and sign off the last step.
Head of AI Operations. AI systems and platform leader with 10+ years across AI, AdTech, media and creative technology.
Before: everyone had AI tools. Nobody had handed one a whole job.
In each room an agent took on a whole job. In each one, a person still signs off the last step.
I chair the monthly AI steering group with the CEO and heads of department. I wrote the 2026–27 strategy: stop handing people AI tools, start delegating whole jobs to agents.
97% weekly AI use today. Target: 90% owning work that runs without them.
Signed off by the CEO and heads of department, every month.
Read the case studyI designed and ran the agency’s first AI hackathon. Forty staff, the CEO among them, in five teams with no engineers, launched working products in six hours. Leadership credited it with winning over the sceptics.
Then came an AI academy inside the platform, with practice clients, AI marking and certification, rolled out team by team.
40 people, 6 hours from no engineers to working products.
Signed off one team at a time, through certification.
Read the case studyI built the platform staff use to work with AI every day, hands-on in Python and TypeScript on Google Cloud. It has 65 tools and 10 connections to the systems the agency already runs on. Anyone can ask the client data warehouse a question in plain English, scoped to the client in front of them, hundreds of times a day.
≈500 hours saved a month, across 40 monthly users and 1,000+ agent conversations.
The platform drafts. A person sends.
Read the case studyI built the system the agency’s agent teams run inside. Each team gets its own sandbox. Each agent gets its own key, its own spend cap and its own short list of sites it may reach. Read-only comes first, and write access is earned.
In paid media, a lead agent reviews every draft before a person sees it. Staff approve, reject or queue agent work on one page.
Every agent stops before the final click.
Nothing is published or sent until a person approves it.
Read the case studyInvoices now route themselves into the accounting system. Weekly timesheets are drafted from Slack, calendar and email, ready for staff to check. An AI PA runs the CEO’s weekly round-up, reminders and calendar triage, and is technically unable to send as the person it serves.
10 → 4 days, the median time for an invoice to reach Accounts.
Staff sign off their own timesheets.
Read the case studyTools that judge social creative concepts against live Meta performance data, find past creative and recommend which ads to pause, and draft LinkedIn posts in each leader’s own voice.
Every idea checked against live performance data.
Every post passes an anti-slop check, then a person approves it.
Read the case studyEvery room now runs through one operating system, under four rules.
97% of staff use AI every week. Target: 90% owning work that runs without them.
Not built: a creative-scoring tool. A blind test showed it predicted ad performance no better than chance. I advised against buying it.
AI systems and platform leader with 10+ years across AI, AdTech, media and creative technology. I pair a product manager’s eye for how teams really work with production engineering.
Outside work: creative technology, AI art and open-source AI tools.
mjameswilliams@proton.me LinkedIn
They sit behind a personal access link. Ask for one, or enter the code you were sent.
Request access