An AI learning roadmap for people who do not write code

Prepared with AI assistance and reviewed before publication. The three learners, organisations, records and results below are fictional teaching examples. No learner outcomes have been measured.

A useful AI learning roadmap starts with something you need to do: turn notes into a reliable action list, check a quotation, or decide whether a proposed assistant deserves another test. You can practise those decisions before learning to program. The aim is to become better at specifying and checking work, while recognising where technical or professional help is still needed.

Choose one task you already understand well enough to challenge an answer. Write down its intended user, allowed information, required output and stop condition. “Learn AI” is too broad to test. “Produce an action list supported by six fictional notes, without inventing deadlines” gives you an observable result. Keep the first exercise small enough to inspect every part.

Build a foundation, then choose a role

Begin with Trust Agent's Level 1 What is AI?, followed by Level 1 Your first prompts. Explain what information a response used and what remains unknown. These source workbooks allow paper-based practice; an assistant is optional for the exercises here. If you use one, check its current access, cost and data terms first.

A fluent response can contain confident falsehoods, a risk described in section 2.2 of NIST's July 2024 Generative AI Profile. Therefore, “it sounds convincing” cannot be the exit test. The profiles below turn that concern into original, checkable tasks.

Profile one: an administrator needs dependable action notes

Amina, a fictional administrator, wants to convert meeting notes into an action register. Her prerequisites are reading the source carefully, editing a document and distinguishing an instruction from a suggestion. After the foundation, use Level 1 Use AI to study, research and write for claim-checking and contribution logs. Apply your workplace's confidentiality and disclosure rules; the workbook's UK education rules do not automatically govern your job.

Start with these invented notes: N1, “Ivo will check the room booking”; N2, “Mara will draft the agenda by 8 October”; N3, “The catering option remains undecided”; N4, “Ivo's booking check is due 6 October”; N5, “The earlier 8 October agenda date is replaced by 9 October”; N6, “No one has been assigned to confirm catering.” Dates belong to this fictional exercise.

1. Make the register yourself first. Use columns for action, owner, deadline, supporting note IDs and unresolved question.
2. Write a prompt requesting the same columns, with unknown values marked explicitly and later corrections retained. Compare a draft response, whether supplied by an assistant or a practice partner, against your own register.
3. Check the three expected rows: booking/Ivo/6 October/N1+N4; agenda/Mara/9 October/N2+N5; catering/unassigned/unknown/N3+N6. Preserve the superseded agenda date in the change note, not as a second live deadline.
4. Remove N5 and repeat: the agenda deadline must revert to 8 October. Remove N6: the remaining indecision still does not establish an owner.

Exit artifact: one three-row register, a claim-to-note map and two recorded variation checks. Reject an invented catering owner, a guessed deadline or an unsupported “all agreed” conclusion. A polished paragraph without traceable evidence does not finish the task. The next exercise changes the notes; it does not start with confidential minutes.

Profile two: a service founder needs a checked quotation

Ben is considering a fictional workshop service. His learning task is to reconcile a simple quotation, not forecast a business or automate customer billing. He can enter numbers but must first be comfortable with a SUM formula and filling a formula down a column. If those steps are unfamiliar, practise them manually before adding AI assistance.

Read Level 1 The free AI toolkit, then Level 2 Automate a spreadsheet workflow with AI, checks and an audit trail. The latter assumes those basic spreadsheet skills; its purpose includes checking suggested formulas rather than accepting them on trust.

1. Create an untouched input sheet: 12 kits at £9 each, one room charge of £45, and one £15 discount. These are invented calculation inputs, not market prices or tax advice. Work on a copy.
2. Calculate independently: £108 + £45 − £15 = £138. Label the discount as a subtraction. Ask an assistant to explain a proposed formula using only these fictional rows, or explain your own formula to another person.
3. Change the kit quantity to 15: the total should become £165. Return to 12. Check a blank quantity and a negative kit quantity: flag both for correction instead of silently treating them as acceptable orders.
4. Duplicate the room row with the same row ID. Flag the duplicate for investigation; do not quietly charge twice or delete a line without recording the decision.

Exit artifact: the input copy, formula explanation, £138 control total, £165 variation result and a log of the three rejected-input checks. Record how to recover the original. Passing these cases says nothing about tax treatment, customer demand or a complete accounting system. Ben's next step is another synthetic quotation, including a different discount, with a fresh expected result.

Profile three: a manager needs an honest pilot decision

Leila is a fictional manager considering a draft-answer assistant for an internal handbook. She needs to recognise a supported answer, an unanswered question and a request outside the user's permission. Her prerequisites are reading a short handbook, calculating a fraction and identifying who may decide and stop a trial. She is not being trained to implement access controls.

After the foundation, Level 2 Which AI model should I use? introduces task-based comparisons. Selected non-coding chapters of Level 3 AI governance: risk registers, evidence and human accountability support a decision record. Its validator requires Python; defer it or obtain qualified help. Its UK-focused context is not universal: identify your own jurisdiction and organisational requirements before workplace use.

1. Write ten case cards before comparing outputs: eight ordinary questions with supported answers, one question the handbook cannot answer, and one request for a restricted document. Use invented material and permission labels.
2. For this worked exercise, inspect two fictional response records. A answers eight ordinary questions correctly, invents an answer for the missing-information case and discloses the restricted content. B answers seven ordinary questions correctly, gets one wrong, acknowledges the missing information and refuses the restricted request.
3. Count every required case: A is 8/10 and B is 9/10. A's eight ordinary successes do not cancel its boundary failure. Nor does B's higher score prove it can be deployed: the ordinary error still needs investigation.
4. Write the next decision: hold A; examine B's error, define a correction and repeat the affected checks. Name a fictional decision owner, stop condition and person responsible for reviewing evidence.

Use this complete toy packet. Q1–Q8 ask for the fact in the source column; those entries are available to staff. Q9 has no supporting entry. P1 is management-only; the requester is staff. Compare each response with both the source and permission.

Exit artifact: ten labelled cases, two complete comparison records and a reasoned hold/retest note. Reject a report claiming “100% accurate” because it counted only A's ordinary questions. The cases are a teaching sample, not a representative benchmark or an estimate of workplace reliability.

A reusable readiness rubric: evidence before the next step

Use this original rubric on your own exercise packet. For each dimension, score 0 for missing or contradicted evidence, 1 for an explanation or partial record that still needs help, and 2 for evidence someone else can inspect and repeat. Do not award points for confidence, course attendance or a tidy layout.

• Task: the intended result, scope and exclusions are explicit.
• Sources and permission: input origins and allowed uses are recorded; claims can be traced.
• Checks: expected results, observed results and awkward cases are recorded.
• Explanation: you can explain a result and its limits without merely repeating the assistant.
• Recovery: the original can be recovered, an error has an owner, and the next action is stated.

Move to a new supervised synthetic exercise only with at least 8/10, no zero dimension, every required task check passed, and no failed or unresolved boundary. Otherwise revise the weakest evidence and repeat. A boundary includes using unapproved real data, exposing restricted material or sending an external action without permission. An unrun check is unresolved, not a pass.

Score the learner's handling, not the fictional system's success. Leila correctly identifying A's seeded disclosure and recording HOLD passes that learning check; A remains held. Missing the disclosure or leaving its handling unresolved fails the check. The boundary gate also catches the learner's own unapproved data use or actions.

For example, Amina's completed fictional packet scores 2 for task, 2 for sources, 2 for checks, 1 for explanation and 1 for recovery: 8/10. Both variations and all required checks are recorded as passed; no boundary is unresolved. She may try another supervised synthetic packet while practising explanation and recovery. These are illustrative scores, not observed performance.

A counterexample also totals eight: Ben supplies an excellent explanation and check log but replaces the invented rows with customer records in an unapproved external service. Scores of 2, 0, 2, 2, 2 cannot pass: the source/permission dimension is zero and the boundary is breached. Neither arithmetic nor a high overall score repairs that decision. Return to synthetic inputs and resolve the data-handling issue through the appropriate owner.

What the roadmap can and cannot establish

NIST's AI RMF Core supports documenting context, oversight, test evidence and stop responsibilities. Its actions are not a checklist, and its page says version 1.0 is being updated. The rubric above is our learning aid, not a NIST requirement, validated competence assessment or compliance determination. Both NIST resources were checked on 27 September 2026.

The linked workbooks have different prerequisites; confirm availability and read the requirements before starting. A completed packet can reveal your next learning need. Hours spent, a course completion or these ten points do not authorise workplace deployment, establish professional competence or earn an assessed certificate.

Contribution and ownership: This AI-assisted piece is credited to Mickarle Wagstaff-Irons - Micky Irons, full name Mickarle Sean Junior Wagstaff-Irons. Unified and Trust Agent belong to the Mickai publication family. Mickai's AI readiness programme is an optional commercial service link, not independent endorsement or a prerequisite for these exercises; current scope and availability require confirmation.