Staff have completed general AI courses and still ask what is appropriate in their own role.
A licence was bought, or a session was run, and the material did not match the documents on anyone's desk. Completion is low and nothing in the work has visibly changed.
Among UK charities, 76% now use AI tools, up from 61% a year earlier, while the barriers small charities name most are limited digital skills (48%) and lack of training (44%) source. The constraint is not access to tools. It is knowing which tasks to give them and how to check what comes back.
What you hold at the end
A working playbook for each department, and a change in what the team does next week.
A department playbook: five tasks to hand to AI, three never to, the prompt patterns that work on that team's files, the four checks before anything is sent, and what to log.
Staff working inside the rules, because the page is cleared against your own policy before it is handed over.
An attendance record and per-team material, which is what an evidence file needs.
A measurable change: the tasks on the page are the ones we time before and after.
We scope training by where a team starts and where it needs to reach. The days below are delivery days for the
programme itself, before department labs. Levels 1 and 2 are starting points, not destinations: reaching level 3
is the shortest engagement we take, whether the group begins at 1 or at 2.
Level
What the person can do
Days to reach
What we plan against
1
Not startedLevel 1No working use of an assistant. Staff here often hold the strongest reservations, which is useful: the objections are specific and can be answered with their own documents. Nothing to unlearn.Has not used an assistant for work.
starting point
Where people begin, not something we train to.
2
OccasionalLevel 2Occasional use with no method. The output is unpredictable because the input is, nobody can say which prompts worked, and the data rules have not been learned. This is where most organisations actually sit.Asks for a summary, a search or a first draft, without a method. Results vary by person and by day.
starting point
Where people begin, not something we train to.
3
StructuredLevel 3Understands what a model is, its context window, model tiers, web search versus the model’s own knowledge, and the organisation’s data rules. Writes a full brief (role, audience, purpose, constraints, format, sources), uploads the source rather than describing it, iterates, and runs the four checks: figures, facts, language, defensibility. Proof: three real tasks delivered with the checks recorded.the floorWrites a full brief, grounds it in the real document, verifies before anything leaves the team.
3 days
Time saved on the tasks in scope: 20 to 30%
4
Set upLevel 4Context engineering: quality depends on what the model has in front of it, not on clever wording. Projects hold the reference documents, custom instructions hold the standing rules, Skills hold the house method, one read-only connector brings in the files. Proof: a Project and a Skill a colleague can use without you, at the same quality.Loads context once (Projects, custom instructions, Skills), so nothing is briefed twice and quality stops depending on who asked.
5 days
Time saved on the tasks in scope: 30 to 40%
5
DelegatingLevel 5Agentic AI: an agent plans, acts with tools, checks and repeats until the job is done. The person briefs a whole job with stopping rules, runs two or three workstreams in parallel, checks output against a small test set, and sets permissions deliberately. Understands browser agents, MCP connections at the user level, and blast radius. Proof: one multi-step job delegated end to end, with a test set and a permissions note.Hands whole multi-step jobs to agents, supervises rather than types, runs workstreams in parallel.
6.5 days
Time saved on delegated workstreams: 40 to 50%
6
BuildingLevel 6Designs Skills and agents for someone else: inputs, method, outputs, refusal cases, an operating guide and a named owner. Knows MCP and connectors at the organisation level, evaluation with rubrics and test sets, prompt injection in depth, and how to choose models and vendors. The right target for champions and team leads, not for everyone. Proof: the department’s set-ups survive a week without you.Builds Skills, agents and connections for colleagues; reviews their output; owns the operating guide.
8 days
Gain is team-level, not a personal percentage
7
At the terminalLevel 7Terminal-based agentic work with no coding background. Claude Code works on folders and files rather than one conversation: renaming, restructuring, classifying or reviewing hundreds of documents in one run, the same review across a year of reports, one repeatable pipeline with a dry run and a log. A CLAUDE.md carries the standing instructions; hooks and permission modes limit what the agent can touch. Proof: one pipeline on real files that a second person has run.Runs agentic work on folders of files from the terminal (Claude Code), without writing code.
10 days
Task-dependent; batch work can exceed any chat figure
What the percentages are, and are not.
Planning assumptions, not promises, and they apply to the tasks in scope, never to a whole job. Anchors: 40% less
time on professional writing tasks at higher rated quality Noy and Zhang, Science 2023; 25% faster and 12% more tasks on realistic consulting work Dell’Acqua et al., HBS 2023; 14% more output across 5,179 support agents, 34% for the least experienced Brynjolfsson et al., NBER 2023.
Level 3 sits at the lower edge of those studies because verification is added to the work. Levels 4 and 5 are
extrapolations: no published field experiment has yet measured persistent set-ups or delegated agents at this
scale, so they are planning figures that each engagement replaces with its own before-and-after timing at ninety
days. Levels 6 and 7 pay off through other people and through volume, so no personal percentage is honest.
Why the days are not linear.
Level 3 is the longest step because it carries all the background: what a model is, the brief, the checks, the
data rules. Every later level assumes it and adds one capability. Level 4 is where the time starts compounding,
because nothing is briefed twice; level 5 changes what a person can hold at once; levels 6 and 7 are worth buying
for a few staff rather than for everyone. The same study that measured a 25% speed gain found AI users 19 points
worse on a task chosen to sit outside the tool’s competence HBS, 2023, which is why every level above 2 is taught with its own refusal cases.
What each level covers
The minimum sufficient background, and what to skip for now.
Each level names what you must understand, what you do, the tools, and what the free content pushes at you that you do not need yet. The full page for each level is free: the method, a cheat sheet, example prompts and the day-by-day plan.
3Structured
UnderstandWhat a model is, the context window, model tiers, web versus own knowledge, your data rules.
DoThe full brief, the source document, iteration, the four checks, a one-line log.
ToolsA chat assistant with file upload, from the approved list.
Skip for nowPrompt frameworks with acronyms, API keys, agents, automation, benchmarks.
Levels are set per group, not per organisation. A fundraising team at level 2 and a leadership team at level 4 are
trained separately and scoped separately. Find your level → · Estimate a level in the estimator →
Examples
The general module, and what a department lab covers.
The general module sets the shared rules. Each lab applies them to one department's real tasks and documents.
General module · all staff
Seven parts, one sitting
The rules in practice. What your policy permits, with which data, on which tools.
Delegation. Which tasks to hand over, which to keep, and the red lines for beneficiary and donor data.
Description. Three prompt patterns: role, audience, constraint and format; give the source document; let the tool interview you first.
Discernment. The four checks before an output is used: figures, facts, language, defensibility.
Diligence. What to log, what to disclose, who owns the result.
Practice. Two exercises on tasks every department shares, on your approved tools.
Close. The shared playbook and where questions go afterwards.
Department lab
Programmes and monitoring
Tasks: donor narrative reports from field notes and the previous report; indicator wording for a logframe; synthesis of survey free-text. Red line: no beneficiary names, case details or locations below district level in any tool.
Department lab
Fundraising and partnerships
Tasks: first drafts against a funder's template; eligibility checklists from a call for proposals; donor research summaries. Red line: no figure accepted without recomputation against the budget file.
Department lab
Finance and administration
Tasks: budget narratives from the spreadsheet; procurement comparison tables; procedure drafts. Red line: no bank, payroll or supplier identifiers in any prompt.
Department lab
Human resources
Tasks: job descriptions and interview question banks; staff-facing policy explanations; onboarding material. Red line: no candidate or employee personal data, and no screening decision delegated to a tool.
Department lab
Communications and advocacy
Tasks: press releases and social copy in the house voice; translation drafts for review; briefing notes from long reports. Red line: nothing published without a named owner and a source check.
Department lab
Leadership and governance
Tasks: board paper summaries; decision memos with options; scenario notes. Red line: strategic and personnel decisions are not delegated, and AI use in board papers is disclosed.
Legal, protection and safeguarding, IT and procurement are scoped the same way: a short task inventory with the team lead, then a lab on those tasks.
How we get there
A task inventory, a lab, and a page written with the team.
Task inventory. A short interview with each team lead: what the department spends its week on, and which of it repeats.
General session for all staff on the shared rules and the four competencies.
A 90-minute lab per department on that team's real tasks and real documents, run by the person who scoped the engagement.
The playbook is written in the lab, with the team, then cleared against your policy.
Everything follows the four competencies of the AI Fluency framework published by Anthropic with Rick Dakan and Joseph Feller source. The framework is public, so the method can be checked against it.
Typical duration: Three to ten delivery days depending on the level, spread across two to six weeks, plus a lab for each department.
Scope and quote
Delivery days, then a written quote.
The number of days follows the level you are training to (three days to level 3, five to level 4, 6.5 to level 5, eight to level 6, ten to level 7) plus the department labs. Online or in person; in person, we arrange the venue, hotel and hospitality. Priced per engagement. You receive a written quote after a free scoping call. Book the scoping call →
Course for individuals and organisation representatives: three delivery days to level 3, for a group of up to nine. Higher levels add the days in the ladder above. Each participant leaves with the playbook for their own tasks and an attendance record.