5

Class 5: AI at Work

Class 5 of 8About 3 hoursNo coding

By the end of this class you will be able to

• Use an approved AI chat tool for six everyday tasks: drafting, summarizing, translating, analysing data, writing meeting notes and brainstorming, with a ready-made prompt for each.
• Say what typically gets lost or distorted in each task, and how to check for it.
• Run the full SitRep workflow in the chat tool, from raw field notes to a reviewed situation report.
• Decide which parts of a task to delegate to AI and which to keep, using a simple decision table.
• Build a small personal library of work prompts you trust.

Where we are

Class 3 taught you to brief the AI, and Class 4 to give it the right information. This class puts both to work on real tasks, ending with the SitRep workflow from start to finish. Class 6 then goes deeper into checking AI output, which you'll already be doing informally today.

1. The right mindset: you are still the author

Think of the AI as a very fast, very willing junior colleague who drafts beautifully and never says "I'm not sure". They can save you hours. They can also hand you a polished paragraph that contains a wrong date, a softened caveat or a figure from nowhere, and they'll never tell you which. Anything you send, sign or publish is still your work. The AI changes how you produce it, not who is responsible for it.

A useful habit before any AI-assisted task is to break it down, the way you'd plan work for a team:

1. What are the steps? For example: gather the facts, check them, draft, review, send.
2. Which steps is the AI good at? Usually: first drafts, restructuring, shortening, changing tone, spotting gaps, generating options.
3. Which steps need a human? Usually: deciding what is true, what matters, what is sensitive, and what is sent in the organization's name.
4. How will I check the result? Decide this before you start, not after the draft looks good.

This is the same "break it down → decide who does what → verify" thinking that runs through both courses. The rest of this class applies it to six common tasks and then to the SitRep.

Every task in this class starts with the same check

Is my tool approved for this material? Emails, donor information, meeting content and field data can all include personal or sensitive information. Anonymise where you can, and don't paste what the tool isn't cleared for. See the data rules on the About page.

2. Drafting: emails, talking points, donor updates

Drafting is where most people start, and where the time savings are clearest. The blank page is often the slowest part of writing; the AI removes it.

What AI does well: first drafts, alternative versions, adjusting tone ("warmer", "more formal", "shorter"), turning bullet points into prose and prose into bullet points.

What goes wrong: the draft sounds generic ("We are pleased to inform you…"), invents details to fill gaps (a date, a partner, a figure), overstates results ("significantly improved") or uses a tone that doesn't fit the relationship.

Draft a short email from me to [recipient and their role].

Purpose: [what I need them to do or know], by [deadline]. Relationship and tone: [e.g. we work together often; warm but professional / first contact; formal]. Key points to include, in this order:
- [point 1]
- [point 2]
- [point 3]

Rules: under 150 words; no information beyond the points above; if something needed is missing (a date, a name), put it in [square brackets] for me to fill in rather than inventing it.

The last rule is the most useful one in drafting. Square-bracket placeholders turn "the AI invented a meeting date" into "the AI showed me where I need to add a date".

For talking points, tell the AI who will speak, to whom and for how long: "Five talking points for our country director at a donor roundtable, 90 seconds in total, each point one sentence with one supporting fact from the notes below." For donor updates, a couple of past approved updates as examples will match your house style better than any description of it.

Know the Term: Zero-shot vs few-shot prompting

A zero-shot prompt asks the AI to do a task with instructions only and no examples: "Write a one-line summary of this report." A few-shot prompt adds a handful of worked examples (the "shots") of the input and the output you want, then gives the new input. The model spots the pattern in your examples, such as the length, tone, structure and level of detail, and copies it. Two or three good examples often do more than a paragraph of description, because style is easier to show than to explain.

Analogy: handing a new colleague two past donor letters and saying "one like these", instead of writing a page of style guidance.

Why you care: when the AI keeps getting the tone or layout "almost right", examples are usually the quickest fix. Just make sure they are good ones, because the AI will copy their flaws too.

3. Summarizing long reports, and what gets lost

A 60-page evaluation, a 30-page policy, a long email thread: summarizing is one of the most valuable things AI does. It is also where losses are hardest to see, because you haven't read the original.

What typically gets lost or distorted:

• Caveats and hedges. "Preliminary data suggest a possible increase" becomes "cases increased".
• Numbers and their units. Households become people; "up to 400" becomes "400"; a percentage loses what it's a percentage of.
• Minority views and disagreement. A summary tends to report the majority position and drop the partner who objected.
• What the document doesn't say. A summary rarely tells you "the report has no data for Dorra".
• Anything in the middle or the annexes, especially in long documents (Class 4, Section 7).

So ask for a summary that exposes what it did and didn't cover:

Summarize the report below for [reader], who needs to [decide / know what].

Format:
1. Key points: 5 bullets, one sentence each.
2. Figures: a table of every key figure with its unit and the page or
   section it comes from.
3. Caveats: any limitations, uncertainties or disagreements the report
   itself mentions, in its own words.
4. Not covered: important topics a reader might expect that the report
   does not address.

Use only the report. Keep hedging words ("preliminary", "estimated", "up to") exactly as they appear.

Report:
"""
[paste or upload the report]
"""

Then spot-check: pick two figures from the table and find them in the original. If either is wrong, check them all.

4. Translating and multilingual work

Between widely used languages (for example Arabic, Chinese, English, French, Russian and Spanish), AI translation is now good enough for many everyday purposes (as of 2026): understanding an incoming document, a first draft of routine correspondence, checking a colleague's translation. It is not a replacement for professional translation of official documents, and your organization may have rules about which is which.

What goes wrong: official terms are replaced by natural-sounding but non-standard ones; numbers, dates and names shift; tone changes (a firm request becomes a polite suggestion, or the reverse); formal registers differ between languages and the AI may pick the wrong one.

Three moves make AI translation safer:

1. Give it your terminology. A short glossary of the terms that must be translated a particular way (check them against your organization's terminology list or an official terminology database).
2. Ask it to flag, not just translate. Terms it wasn't sure about, idioms with no direct equivalent, and places where the tone had to change.
3. Check with a back-translation or a second reader. Ask a new chat to translate it back, then compare meaning, or have a colleague who reads the language review it.

Translate the text below from [source language] into [target language]
for [audience], in a [formal / neutral / warm] register.

Use these fixed terms:
- [term] → [required translation]
- [term] → [required translation]

Rules:
- Keep all numbers, dates, names and place names exactly as in the
  source.
- After the translation, list: (a) any term you were unsure about,
  (b) any phrase where you changed the tone or structure, and why.

Text:
"""
[paste text]
"""

To check an existing translation rather than produce one, paste both versions and ask: "Compare the source and the translation sentence by sentence. List any differences in meaning, numbers, names, tone or terminology. Don't rewrite it; just list the issues."

5. Data analysis in plain words

Many enterprise tools can now read tables and spreadsheets, and some can run calculations behind the scenes (as of 2026, often labelled "analysis" or "code interpreter"). This is useful for questions like "Which district received the fewest kits per household?" or "What changed between last month and this month?"

What goes wrong: the AI misreads a column, skips rows, adds up a range as if it were a number, mixes units, or, if it isn't actually calculating, produces a number that looks computed but was predicted like any other word (Class 2). The answer arrives in the same confident tone either way.

Rules of thumb:

• Tell it what the columns mean and what the units are.
• Ask it to show its working: the rows it used and the calculation, not just the result.
• Check at least one number yourself with a calculator or your spreadsheet. If the tool offers a real calculation feature, prefer it for anything beyond simple arithmetic.
• Ask for the "so what" separately from the numbers. Interpreting a trend is a judgment; make sure the numbers underneath are right first.

Here is a table of [what it shows]. Columns: [column] = [meaning, unit]; [column] = [meaning, unit].

1. Restate the table's totals for each column so I can check you
   have read it correctly.
2. Answer: [your question].
3. Show your working: list the rows you used and each calculation.
4. Point out any data problems: missing values, ranges instead of
   numbers, mixed units, figures that look like errors.

Do not use any data that is not in the table.

Table:
"""
[paste table]
"""

Step 1, asking it to restate the totals, is a cheap check. If the totals it reads don't match yours, stop there; everything after will be wrong.

Know the Term: Structured output

Structured output means asking the AI to answer in a fixed, predictable shape instead of free-flowing text: a table with named columns, a form with set headings, or a data format such as JSON (a plain-text format that computer programs can read, made of labelled fields like "district": "Aramu"). You describe the shape, meaning the fields, their order and what to write when something is missing, and the AI fills it in. Software that uses AI goes a step further: it can require the model to follow an exact template and reject any answer that does not fit.

Analogy: handing someone a printed form to fill in instead of a blank sheet of paper. Every answer lands in the same box, every time.

Why you care: structured answers are faster to check, easy to compare across reports and ready to paste into a spreadsheet or another system. An empty box also shows you straight away what information is missing.

6. Meeting notes

AI can turn rough notes or a transcript into a clean record with decisions and action points in a few seconds. Some tools can record and transcribe meetings directly.

Before you start: if you plan to record or transcribe, everyone in the meeting must know and agree, following your organization's rules. Sensitive discussions (staff matters, security, negotiations) may not be suitable for AI tools at all.

What goes wrong: actions are assigned to the wrong person; tentative ideas become decisions ("we could try the northern route" → "the team will use the northern route"); disagreement is smoothed over; speakers are misidentified in transcripts.

Turn my rough notes below into meeting minutes.

Format:
- Meeting, date, participants (as listed in my notes only)
- Decisions: only things clearly agreed
- Action points: table with Action | Owner | Deadline. If the owner or
  deadline is not in my notes, write "not stated".
- Open questions / disagreements: anything discussed but not settled

Rules: do not turn suggestions into decisions. If you are unsure whether something was agreed, put it under open questions.

Notes:
"""
[paste notes]
"""

Send the draft minutes to participants for correction before they become the record. That review step is where misattributed actions get caught.

7. Brainstorming and critiquing your own drafts

Two of the most underrated uses of AI don't involve producing final text at all.

Brainstorming. The AI is good at generating many options quickly: titles, angles for a campaign, risks you haven't considered, questions a donor might ask. Quantity is its strength; you supply the judgment about which ideas are good.

I am [task and context]. Give me 10 [ideas / risks / questions], as varied as possible, one line each. Then mark the three you think are strongest and explain why in one sentence each.

Critiquing your own drafts. Ask the AI to be the tough reviewer your draft will eventually meet:

Act as a critical reviewer: [e.g. a donor's programme officer / our head of office / a reader with no background]. Review my draft below.

List, in order of importance:
1. Claims that are not supported by evidence in the draft
2. Places where the reader will be confused or need more context
3. Anything that could be read as overstated or insensitive
4. The single change that would most improve it

Do not rewrite the draft. Just give me the list.

Draft:
"""
[paste your draft]
"""

"Do not rewrite" keeps the thinking with you. The AI points; you decide what to fix and how. This is often more useful than asking it to "improve" the text, which tends to flatten your voice into generic prose.

8. The SitRep workflow, start to finish

Now put it all together on the running case study. A situation report (SitRep) turns messy field information into a short, structured, reliable update for decision-makers. It's a good test of AI at work because it needs speed and accuracy.

Here's the workflow, broken down by who does what:

Step	Who	What happens
1. Prepare	Human	Collect the field notes, remove names and sensitive details, confirm the tool is approved for this material
2. Extract	AI	Pull every fact into a table with sources, units and confidence
3. Check facts	Human	Compare the table with the notes; fix errors; decide on conflicting figures
4. Draft	AI	Write the SitRep from the checked table, in the house format
5. Self-review	AI	Compare the draft with the checked table; list anything added, changed or missing
6. Review and sign off	Human	Read the draft as the reader will; check tone and sensitivity; approve

Notice the design: the AI drafts from the checked table, not from the raw notes. Once you've fixed the facts in step 3, the draft can only be as wrong as the table, and step 5 catches drift.

Step 2 · Extract.

You are helping prepare a humanitarian situation report. From the field notes below, extract every fact into a table:
Topic | Fact | Figure and unit | Location | Date | Source | Count, estimate or second-hand?

Rules: use only the notes; write "not stated" for anything missing; keep households and people separate; flag any ambiguous date. After the table, list the gaps a SitRep reader would want filled.

Field notes:
"""
[paste anonymised notes]
"""

Step 3 · Check facts (you). Go down the table with the notes beside it. Correct anything wrong in the chat: "Row 4: the source says 14 cases, not 41. Row 7: '04/05' is 4 May, confirmed by phone."

Step 4 · Draft.

Using ONLY the corrected table above, write the SitRep in this format:
Headline (one line) / Situation overview / Affected population / Priority needs / Response so far / Gaps and unknowns. Under 300 words. Keep every "estimate" and "not stated" label visible in the text. Reader: the country director and partner organizations.

Step 5 · Self-review.

Compare your SitRep with the corrected table, line by line. List:
1. Any figure, name, date or claim in the SitRep that is not in the
   table
2. Any label ("estimate", "not stated") that was dropped
3. Any important row in the table that the SitRep leaves out
Do not rewrite the SitRep.

Step 6 · Review and sign off (you). Read it as the country director would. Is the most urgent need at the top? Is anything worded in a way that could cause harm if forwarded (locations of vulnerable groups, blame on a named partner)? Would you put your name on it? Only then does it go out.

Save it as a project

If your tool supports projects, store the SitRep format, two approved past SitReps and the standing rules ("use only the notes", "not stated", "keep units") there. Next week, steps 2–5 start from the same brief for everyone on the team (Class 4, Section 3).

9. What to delegate, what to keep

Not every task, or every part of a task, should go to AI. Use this table as a starting point, and adjust it to your organization's rules.

Task or step	Delegate to AI?	Why
First draft of routine email, update or talking points	Yes, then edit	Saves the most time; low risk once you check facts and tone
Changing length, tone, format or structure of your own text	Yes	Low risk; the facts are already yours
Summarizing a long document for your own understanding	Yes, with spot-checks	Good for orientation; check before quoting it
Brainstorming options, risks or questions	Yes	Quantity is its strength; you choose
Critiquing your draft	Yes	A second pair of eyes; you decide what to change
Translating for understanding or a first draft	Yes, with checks	Check terms, numbers and names; official texts go to professional translators
Calculations and data analysis	Partly	Let it help, but verify numbers yourself or with a real calculation tool
Deciding which figure is correct when sources conflict	Keep	A judgment that needs context, contacts and accountability
Final wording of anything published or sent in the organization's name	Keep	You are accountable; check tone, sensitivity and politics
Decisions about people: hiring, performance, eligibility for assistance	Keep	High stakes, risk of bias, and often personal data
Anything involving data the tool isn't approved for	Keep (don't paste it)	Data protection comes before convenience
Sensitive negotiations, security information, protection cases	Keep	Risk of harm if leaked or mishandled; usually not approved for AI tools

A quick test for anything not on the list: "If the AI gets this wrong and I don't notice, how bad is it?" If the answer is "mildly embarrassing", delegate and check. If it's "someone could be harmed, misled or unfairly treated", keep it, or delegate only a small, checkable part.

Prompt Lab

Work in pairs using your organization's approved AI chat tool. The exercises use the fictional field note and distribution table below from the SitRep Assistant case study.

Field note (fictional)

FIELD NOTE - Aramu district, Northern Veloria - sent Thurs eve From: field team, Veloria Relief Coalition

heavy rain since the weekend, Tula river burst near Kessa bridge. we got to Aramu town + Lomi and Tavet villages today. could NOT reach Dorra, road cut after the bridge, no phone signal there either.

approx 1,200 households displaced in total (our estimate + figures from village chiefs). most are staying at Aramu primary school, which is full - maybe 600 people there. more families in the community hall but we didn't manage to count them.

water is the big problem. 2 of the 3 boreholes in Lomi are flooded and people are drinking from the river. Tavet health post still open but running low on ORS, nurse reports 14 cases of acute diarrhoea since Monday, mostly children under 5. no deaths reported to us.

VRC is distributing food at the school, enough stock for about 3 days. next truck expected 04/05 if the Kessa road is passable. urgent: water purification, hygiene kits, ORS, shelter materials for the hall.

Distribution table (fictional)

Location     | Households (est.) | Hygiene kits given | Water purif. sachets Aramu town   | 640               | 410                | 12,000 Lomi         | 330               | 120                | 3,500 Tavet        | 230               | 200                | 4,000 Dorra        | not visited       | 0                  | 0 TOTAL        | 1,200             | 730                | 19,500

Exercise 1: A donor email with honest placeholders

Task: Draft a short update email to a donor from the field note, using square-bracket placeholders instead of invented details.

Draft an email from me (programme officer, Veloria Relief Coalition) to our contact at the Northern Fund, a regular donor we work with closely.

Purpose: a brief update on the Aramu floods and a heads-up that we may request a reallocation for water purification. Tone: warm, professional, factual. Length: under 180 words.

Use only the field note below. For anything the email needs that is not in the note (a date, an amount, a name), insert a placeholder in
[square brackets] instead of inventing it. Do not name individual
staff.

Field note:
"""
[paste the field note]
"""

What to look for:

• Did it use placeholders (e.g. "[amount]", "[date]") rather than inventing a figure for the reallocation?
• Did "approx 1,200 households" stay an estimate, or become a fact?
• Is the tone right for a donor you know well? Ask for a more formal version and compare.

Exercise 2: Summarize, then find what was lost

Task: Ask for a very short summary, then ask the AI to audit its own summary against the source.

Summarize this field note in two sentences for a senior manager.

Field note:
"""
[paste the field note]
"""

Then, in the same chat:

Now compare your two-sentence summary with the original note. List every caveat, estimate label, uncertainty, gap or figure that the summary dropped or changed. For each, say whether a senior manager would need to know it.

What to look for:

• Which caveats disappeared: "approx", "our estimate", the uncounted community hall, Dorra not reached, the ambiguous "04/05"?
• Did the short summary turn "maybe 600 people" into a firm number, or mix households and people?
• Now use the summary prompt from Section 3 instead. Does its "Caveats" and "Not covered" section keep what the two-sentence version lost?

Exercise 3: Checking a translation

Task: Translate the needs section of the field note into French (or another language one of you reads), then check the translation with a second chat.

Translate the text below from English into French for a partner
agency's coordination meeting, neutral professional register.

Fixed terms:
- oral rehydration salts (ORS) → sels de réhydratation orale (SRO)
- hygiene kits → kits d'hygiène
- household → ménage

Keep all numbers, names and place names exactly as in the source.
After the translation, list any term you were unsure about.

Text:
"""
water is the big problem. 2 of the 3 boreholes in Lomi are flooded
and people are drinking from the river. Tavet health post still open
but running low on ORS, nurse reports 14 cases of acute diarrhoea
since Monday, mostly children under 5. urgent: water purification,
hygiene kits, ORS, shelter materials for the hall.
"""

Then open a new chat and paste both the English source and the French translation:

Compare the English source and the French translation below sentence by sentence. List any differences in meaning, numbers, names, tone or terminology. Do not rewrite the translation.

What to look for:

• Were the fixed terms used every time (SRO, ménage, kits d'hygiène)?
• Did "14 cases", "2 of the 3" and "under 5" survive unchanged?
• Did the translation tidy up the note's informal style? Is that acceptable for this audience?
• If one of you reads French well: do you agree with the checker, or did it miss something?

Exercise 4: Analysing the table, and checking the numbers

Task: Ask a simple analysis question and verify the result yourself.

The table below shows households and distributions by location in Aramu district. Households are estimates.

1. Restate the total for each numeric column so I can check you have
   read it correctly.
2. For each location, calculate hygiene kits per household and water
   purification sachets per household.
3. Which visited location is least covered for hygiene kits?
4. Show your working, and point out any data problems.

Use only the table.

Table:
"""
[paste the distribution table]
"""

What to look for:

• Check the totals yourself: do 640 + 330 + 230 make 1,200, and 410 + 120 + 200 make 730? Did the AI check this, or just copy the TOTAL row?
• Check Lomi's kits per household with a calculator (120 ÷ 330 ≈ 0.36). Is the AI's figure right?
• Did it handle Dorra sensibly ("not visited" is not zero households) instead of dividing by zero or treating Dorra as fully covered?
• Did it mention that the household figures are estimates, so the ratios are approximate?

Exercise 5: The full SitRep workflow

Task: Run steps 2–6 from Section 8 on the field note. One partner operates the AI; the other does the human steps (checking the table in step 3 and the sign-off in step 6). Use the prompts from Section 8 in order, in one chat.

You are helping prepare a humanitarian situation report. From the field notes below, extract every fact into a table:
Topic | Fact | Figure and unit | Location | Date | Source | Count, estimate or second-hand?

Rules: use only the notes; write "not stated" for anything missing; keep households and people separate; flag any ambiguous date. After the table, list the gaps a SitRep reader would want filled.

Field notes:
"""
[paste the field note]
"""

After checking the table, send your corrections (for this exercise, assume the team confirmed by phone that "04/05" means 4 May), then use the draft and self-review prompts from Section 8.

What to look for:

• In step 3, did you find anything to correct in the table? Even one error justifies the step.
• Does the draft keep "estimate" and "not stated" labels visible, and is water near the top?
• Did the self-review in step 5 find anything? Check its claims yourself: it can be wrong too.
• At sign-off: would you send this to a country director? What did you change by hand, and how long did the whole workflow take compared with writing from scratch?

Check Your Understanding

1 You ask the AI to draft a donor email from your field notes. The notes don't say when the next distribution will happen. What is the best instruction to include?

1. "Make the email sound confident; donors don't like uncertainty."
2. "Estimate a realistic date based on typical humanitarian timelines."
3. "If anything needed is missing, put a placeholder in [square brackets] instead of inventing it."
4. No instruction is needed; the AI never adds information that isn't in the notes.

2 The AI summarizes a field note as "1,200 households have been displaced and 600 people are at the school." The note said "approx 1,200 households (our estimate)" and "maybe 600 people". What has the summary lost?

1. The uncertainty: both figures were estimates, and the summary presents them as facts.
2. Nothing important; rounding and simplifying is what a summary is for.
3. The figures are wrong; the AI must have misread the numbers.
4. The summary should have converted households into people for consistency.

3 You need a French version of a short partner briefing. The AI's translation reads smoothly. What is the most useful next step?

1. Send it; smooth, natural French means the translation is accurate.
2. Ask the same chat, "Is this translation correct?" and accept its answer.
3. Translate it into five more languages so it can be compared across versions.
4. Check numbers, names and key terms, then use a fresh chat or a colleague who reads French to compare it with the source.

4 You paste a distribution table and ask for kits per household in each village. The AI gives neat figures. What should you do before using them in a SitRep?

1. Nothing; computers are good at arithmetic, so the numbers will be correct.
2. Check that it read the table correctly (totals) and recalculate at least one figure yourself.
3. Ask it to be more careful and send the same question again.
4. Round the figures, since small errors don't matter in a SitRep.

5 In the SitRep workflow, why does the AI write the draft from the checked fact table rather than directly from the raw field notes?

1. Because the AI can't read raw field notes.
2. Because tables use fewer tokens, which saves money.
3. Because a human has corrected the facts in the table, so the draft starts from verified information and drift is easier to spot.
4. Because the country director prefers tables to prose.

6 Two reports give different figures for displaced households. Which part of the SitRep workflow should you keep rather than delegate to the AI?

1. Deciding which figure to use and how to present the disagreement.
2. Formatting the table of figures.
3. Listing every figure with its source and date.
4. Suggesting alternative wordings for the headline.

7 A colleague wants to use AI to decide which households are eligible for cash assistance, by pasting the full beneficiary list with names and circumstances. What is the best advice?

1. Go ahead, as long as the prompt says "be fair and unbiased".
2. Go ahead, but only if the tool is an enterprise version.
3. Use AI for the decision, then spot-check a few households by hand.
4. Keep this with people: it is a high-stakes decision about individuals and involves personal data the tool may not be approved for.

Exit Ticket

Before you leave

Name one task from your own work that you will now do with AI. Which step will you delegate, which step will you keep, and how will you check the result before it leaves your hands?

Key terms

• Delegate vs keep: deciding which steps of a task AI can do (with checking) and which must stay with a person.
• Placeholder: a marker in [square brackets] where information is missing, used instead of letting the AI invent it.
• Back-translation: translating a translation back into the original language to check that the meaning survived.
• Terminology list (glossary): fixed translations for key terms that the AI must use.
• Situation report (SitRep): a short, structured update on a crisis for decision-makers.
• Self-review: asking the AI to compare its output with the source and list differences, without rewriting.
• Human sign-off: the final review and approval by an accountable person before anything is sent or published.