A typical six-week AI enablement programme is less dramatic than the demos make it look. The team maps work, sets rules, practises on live tasks and leaves with a few routines it can run without outside help. Most of the value comes from decisions made between sessions.
This is a generic composite of what we've seen across teams. It contains no client story, sector detail or performance figure. The sequence changes with risk and readiness, but the shape is fairly stable.
Week 1 maps the work before choosing tools
The team lists recurring tasks, inputs, outputs, owners and review points. It then scores a short set of use cases by effort, usefulness and risk. The aim is to pick work that happens often enough to practise and is safe enough to learn on.
A good first week ends with two candidate workflows and a baseline. The baseline might be elapsed time, number of handoffs or rework causes. Avoid promising a productivity percentage. You need a starting observation that the team can measure the same way later.
Week 2 sets access, data rules and quality bars
The team agrees which tools are approved, who can use them and what must never be pasted into a prompt. It defines where source material lives and what requires a second reviewer. Owners write a short acceptable-use note in plain English.
Quality also becomes concrete. The group collects examples of acceptable and poor output, then turns those differences into a checklist. The current NIST AI RMF Playbook, updated in June 2026, organises suggested actions around govern, map, measure and manage. That is a useful order for a small programme too.

Week 3 is where the demo meets real work
This week usually goes wrong. Early prompts worked on tidy examples, then fail on an incomplete brief, a late change or a messy source file. Some people conclude that the tool is unreliable. Others quietly fix every output themselves and report that it works.
The programme should slow down here. Compare failures, separate instruction problems from missing context, and tighten the review checklist. Keep one workflow small enough to repeat several times. An OpenAI Academy guide from July 2026 recommends bounded, reviewable workflows with named owners and reviewers.
Our opinion is that week 3 is the real programme. A smooth third week often means the exercises were too clean.
Weeks 4 and 5 turn good attempts into team routines
In week 4, the team rebuilds the stronger workflow as a shared playbook. It defines the trigger, required inputs, prompt or skill, output format, reviewer and exception path. Another person runs it without coaching. Their confusion is useful test data.
Week 5 expands carefully. The team may adapt the same method to a second task or connect approved information sources. Access stays narrow. Human approval remains explicit. NIST's current AI RMF material says roles, responsibilities, periodic review and human oversight should be defined and documented. A workflow is not ready if nobody knows who can stop it.
The group also writes the exception path for a missing source, a policy conflict or an absent owner. A safe routine can pause cleanly. It should not force the next person to guess whether an output is good enough.
Week 6 tests ownership and defines done
The final week centres on handover. Team members run the workflow from a normal request, through review, to a usable output. They compare it with the baseline and record where manual work remains. They also choose a date for the first review after the programme.
Done means the team has approved access, a short data rule, named owners, two repeatable workflows, examples of acceptable output, a review checklist and a retirement path. It does not mean every marketer uses AI every day. Plenty of teams finish with two solid routines and a clear sense of where AI has no business being.
If you are planning this internally, book the week 3 troubleshooting session before the kickoff. If you want a facilitated version, AI Team Foundation is the Chazrt name for this kind of six-week setup. The useful first step is still the same: bring one recurring workflow and its messiest real input.