Notes on putting AI to work.

Article 4 still applies after the Digital Omnibus, but no certificate is required. Who needs AI literacy training, what to cover and what to record.

AI saves individuals real time, then the team spends much of it checking, fixing and forwarding output that only looks finished. Here is where the hours leak, a five-day audit to find it in your own team, and the workflow changes that stop it.

Most AI projects fail for reasons that have little to do with the model: the wrong problem, no owner, no baseline, data nobody can reach. Here is what the research actually says, seven causes with a prevention step for each, and a pre-mortem you can run before you approve the budget.

Strategy consultancies, AI agencies, implementation partners, freelancers, your own team and vendor engineers all sell AI implementation, and each is good at something different. This guide compares them, shows what published rate data exists, and gives you 12 questions to ask before you sign.

The model is rarely the expensive part of an AI project. Here is what implementation costs in 2026, with the three cost buckets, current model prices, a worked example and a worksheet for your own numbers.

Most companies are deployers, not AI labs, and three parts of the EU AI Act already apply to them: the bans, AI literacy and the transparency rules. The heavier high-risk duties now start in December 2027, and here is what to do before then.

Human in the loop only protects you if the approval point sits on the right step and the review is real. Here is a grid for placing approval points, the research on rubber-stamping, the metrics that show whether the check works, and a template to document it.

Most published AI ROI figures are opinions collected in surveys. Here is how to measure the return on your first AI workflow yourself, with a two-week baseline, an honest formula, a board reporting rhythm and a template you can copy.

Most companies should buy commodity AI and build only where the workflow is their own, usually on top of rented models rather than from scratch. This guide sets out the three options, the evidence, the hidden costs and a scoring sheet you can use in your next meeting.

Annex 22 is still a draft, and the July 2025 text keeps generative AI out of critical GMP decisions while allowing it elsewhere under qualified human review. Here is the exact status, a task triage table, and what to document now so the final text does not force a redo.

Most AI roadmaps are a list of use cases with dates attached. This one is a 12-month plan in four phases, each with concrete outputs and a go or stop decision, plus a readiness check, a who-does-what table and a one-page template you can fill in this week.

Some companies that cut staff for AI have had to apologize, rehire or put humans back on the phones. Here is what the documented cases show, why AI-driven cuts go wrong, and the checks to run before you change headcount.

A chatbot answers, a workflow follows steps you set, and an agent decides its own next step. Here is when each one fits, why most valuable systems are workflows with an agent step and a human approval, and what to ask a vendor who says \"agent\".

An AI agent can be steered by any email, document or web page it reads, so its safety depends on what it is allowed to do, not on how clever it is. Here are the risks in plain language, a 15-point check before go-live, the questions to ask your vendor, and what to watch in the first week.

Most staff already use AI at work, often on personal accounts, and a policy that only forbids it gets ignored. Here is a two-page template in plain language, a one-page staff version and a two-week plan to roll it out.

Frontier models are converging in quality and falling in price, and every company can rent the same ones. The gap that matters now opens after the model, in how the work is redesigned, approved, adopted and measured.

Most teams have more AI ideas than capacity to build them. Here is a method for choosing the first one: a workflow with one owner, a baseline you can measure and a clear point where a person approves the result.