Building Better Organizations with AI: Assessment, Analysis and Capacity Development

Lesson 1

1.1 What AI Can and Cannot Do for Your Organisation

QUADRAEDGEBuilding Better Organizations with AI

Module 1 · Lesson 1.1

What AI Can and Cannot Do for Your Organisation

Before choosing a tool, get clear on the shape of the help. AI is fast at language and pattern work, and unreliable at truth, judgement and accountability.

AssessAnalyseDecideBuildGovern

Good atSummarising, drafting, structuring, spotting patterns across lots of text.

Bad atKnowing what is true, understanding your context, taking responsibility.

Your roleFraming the question, supplying context, checking the output, owning the decision.

Framing

A language engine, not an oracle

Today's general-purpose AI predicts plausible text. That makes it genuinely useful for work that is mostly language: turning forty interview transcripts into themes, drafting an assessment framework, restructuring a messy report, explaining a donor requirement in plain words. It also means the model will produce a confident, well-written answer when it has no idea — inventing statistics, citations and quotes that look exactly like real ones. Treat every factual claim as unverified until you check it.

Fit

Where it pays off first in small organisations

The highest-value early uses are usually unglamorous: cutting the time to synthesise qualitative data, producing first drafts of documents that a person then edits, translating between working languages, and making sense of survey free-text nobody has ever had time to read. The lowest-value uses are those where you cannot check the answer, or where the real problem is organisational rather than informational.

Limits

What it will not fix

AI will not resolve unclear strategy, weak governance, an absent theory of change or a team that does not trust its leadership. If your assessment finds those, the improvement work is human. Using AI to paper over them produces polished documents describing an organisation that does not exist.

Delegate to AI

Summarising long documents and transcripts
First drafts of frameworks, questions, reports
Finding themes across many open-ended answers
Reformatting, translating, tidying language
Explaining unfamiliar jargon or requirements

Keep with people

Deciding what the findings mean for your mission
Any judgement about individual staff performance
Facts, figures and citations going to a donor or board
Anything involving beneficiary or personal data
Accountability for the final decision

Key takeaway

AI is an accelerator for language work and a poor substitute for judgement. The value comes from pairing its speed with your context and your verification.

Practical Exercise

List five recurring tasks in your organisation. Mark each as 'delegate to AI', 'AI-assisted with human review' or 'keep with people'. Write one sentence justifying each placement.

Summary

AI is strong on language and pattern work and unreliable on truth and judgement. Early value comes from checkable, language-heavy tasks with a human owning the decision.

Resources for this lesson

Web link

Lesson content: 1.1 What AI Can and Cannot Do for Your Organisation

Open resource

Lesson 2

1.2 The Data You Already Have

QUADRAEDGEBuilding Better Organizations with AI

Module 1 · Lesson 1.2

The Data You Already Have

Most organisations sit on more evidence than they realise, in worse condition than they think. Assessment starts with an honest inventory.

AssessAnalyseDecideBuildGovern

InventoryReports, surveys, minutes, finance, HR, project data, feedback.

QualityComplete? Current? Consistent? Comparable over time?

PermissionWhat were people told when it was collected?

Sources

Take stock before you collect anything new

Annual and donor reports, project monitoring data, staff and partner surveys, board minutes, financial records, HR data, training records, beneficiary feedback, complaint logs and email threads all describe your organisation. Build a simple register: what exists, who owns it, what period it covers, where it lives and how sensitive it is. Most improvement priorities can be identified from this alone.

Quality

Four questions that decide what your analysis is worth

Is it complete enough to be representative, or only the projects that reported well? Is it current, or three years stale? Is it consistent — the same indicator defined the same way across teams? Is it comparable over time, or did the categories change? Weak answers do not stop the work; they set the confidence level you should attach to your findings, and you should state that level openly.

Permission

Consent, purpose and sensitivity

Data collected for one purpose cannot always be reused for another. Beneficiary data, health information, case files, complaint records and staff performance data carry legal and ethical obligations under GDPR or your national equivalent, plus donor conditions. Before any of it goes near an AI tool, decide whether it needs to be there at all — anonymised extracts and aggregated figures usually answer the question just as well.

Key takeaway

The condition of your data sets the ceiling on your analysis. Inventory it, judge its quality honestly, and keep personal data out of general-purpose tools.

Practical Exercise

Complete a data register for your organisation with at least eight sources. Flag which are sensitive, and identify two gaps you would need to fill before assessing performance credibly.

Summary

Assessment begins with an inventory of existing evidence, an honest quality judgement and a clear view of consent, sensitivity and what may lawfully be used.

Resources for this lesson

Web link

Lesson content: 1.2 The Data You Already Have

Open resource

Lesson 3

1.3 Choosing Tools You Can Actually Sustain

QUADRAEDGEBuilding Better Organizations with AI

Module 1 · Lesson 1.3

Choosing Tools You Can Actually Sustain

The right tool is the cheapest one that meets your data protection duties and that your team will still be using in six months.

AssessAnalyseDecideBuildGovern

CategoryAssistants, transcription, survey analysis, translation, spreadsheets.

TermsData retention, training on inputs, hosting location, exit options.

Total costLicences plus onboarding, admin time and the review you must not skip.

Selection

Start with the task, not the product

Name the task first, then look for the smallest tool that does it. A general assistant covers drafting, summarising and thematic analysis. Transcription tools turn interviews and meetings into text. Survey platforms increasingly bundle free-text analysis. Spreadsheet assistants handle cleaning and formulas. Resist buying a platform because it is described as an AI solution; buy capacity for a task you have.

Diligence

Six questions before you sign anything

Where is data stored and under whose jurisdiction? Are your inputs used to train the vendor's models, and can that be switched off? What is the retention period and can you delete on request? Who in your team gets admin control? What happens to your content if you stop paying? And can a small NGO reach a human being when something goes wrong? Vendors serving the non-profit sector often have discounted or free tiers — ask.

Sustainability

Adoption is the real cost

Most failed tool rollouts are not technical. Pick one or two tools, name one person as the internal point of contact, run a short hands-on session with real organisational documents rather than demo data, and revisit after a month. A tool used well by four people beats a licence pool nobody opens.

Key takeaway

Choose the smallest tool that fits a named task, check the data terms before the features, and budget for adoption rather than licences alone.

Practical Exercise

Pick one task from your Lesson 1.1 list. Compare two candidate tools against the six diligence questions and write a half-page recommendation with a defined pilot success measure.

Summary

Tool choice should follow from a named task, pass basic data-protection diligence, and be planned around adoption and sustainability rather than feature lists.

Resources for this lesson

Web link

Lesson content: 1.3 Choosing Tools You Can Actually Sustain

Open resource

Lesson 4

2.1 Designing a Simple Organisational Assessment

QUADRAEDGEBuilding Better Organizations with AI

Module 2 · Lesson 2.1

Designing a Simple Organisational Assessment

A usable assessment covers a handful of domains, uses a scale people understand, and produces evidence you can act on within a quarter.

AssessAnalyseDecideBuildGovern

DomainsSix to eight areas, not thirty. Cover strategy, delivery, systems, people.

ScaleA five-level maturity scale with described behaviours, not opinions.

EvidenceEvery rating tied to something you could show an auditor.

Design

Choose domains that match your decisions

For most NGOs and SMEs, six to eight domains are enough: strategy and governance; programme or service delivery; monitoring and learning; finance and compliance; people and skills; systems and data; partnerships and communications. Add a domain only if a real decision depends on it. If nobody would act differently based on the score, drop it.

Scoring

Describe levels by behaviour, not adjectives

'Level 3 — good' means nothing consistent across raters. 'Level 3 — a documented process exists, is followed by most teams, and is reviewed annually' means the same thing to everyone. Behavioural descriptions are what make a self-assessment defensible and repeatable next year.

Method

Where AI helps in the design

Use an assistant to draft level descriptors, to check your domains against a recognised framework, to generate interview and survey questions per domain, and to pressure-test whether two levels are actually distinguishable. Then have a human panel confirm the framework fits your organisation before anyone scores anything against it.

1Ad hoc. Depends on individuals. Nothing written down; results vary by who is available.

2Emerging. Some documentation exists but is inconsistently used and rarely updated.

3Established. A documented process exists, most teams follow it, and it is reviewed annually.

4Managed. The process is followed consistently, measured against agreed indicators, and adjusted on the evidence.

5Embedded. Improvement is routine, evidence-led, and survives staff turnover and leadership change.

Key takeaway

A good assessment is small, behaviourally described and evidence-backed. AI speeds up the drafting; people must confirm the framework fits.

Practical Exercise

Draft a maturity framework for two domains relevant to your organisation, with five behaviourally described levels each and the evidence required for levels 4 and 5.

Summary

Keep the assessment to six to eight decision-relevant domains, describe maturity levels by observable behaviour, and require evidence for high ratings.

Resources for this lesson

Web link

Lesson content: 2.1 Designing a Simple Organisational Assessment

Open resource

Lesson 5

2.2 Working with AI on Interviews, Surveys and Reports

QUADRAEDGEBuilding Better Organizations with AI

Module 2 · Lesson 2.2

Working with AI on Interviews, Surveys and Reports

Qualitative material is where AI saves the most time — and where unchecked output does the most quiet damage.

AssessAnalyseDecideBuildGovern

PrepareAnonymise, label sources, work in batches you can trace.

PromptGive role, material, task, output format and a rule about uncertainty.

VerifySample the quotes back to the transcript. Always.

Preparation

Get the material ready first

Strip names, locations and anything identifying, replacing them with codes you hold separately. Label each extract with its source so any finding can be traced back. Work in batches small enough that you can spot check them — twenty transcripts analysed in four batches of five is far more auditable than one bulk request.

Prompting

Five elements that make the difference

State the role and context, supply the material, define the task precisely, specify the output format, and instruct the model on what to do when it is unsure. That last element matters most: told to flag uncertainty and quote only verbatim, a model behaves far better than one told simply to 'analyse this feedback'. Ask for themes with supporting quotes and source labels, and ask explicitly for counter-evidence and minority views, which summarisation tends to flatten.

Verification

Sample, trace, and watch what got lost

Pick three quotes per theme at random and find them in the original. If any is paraphrased, reordered or absent, discard the batch and redo it. Then read two full transcripts yourself and ask what the thematic summary missed — usually the strongest single voice, the outlier and the thing nobody wanted to say plainly.

Key takeaway

Anonymise, batch, prompt precisely, demand verbatim quotes with sources, and always sample-check against the original before a finding travels.

Practical Exercise

Take ten anonymised open-ended responses from any survey you hold. Run a thematic analysis using the prompt pattern above, then verify every quote against the source and note what the summary lost.

Summary

Qualitative analysis is AI's strongest contribution to organisational assessment, provided material is anonymised, prompts are precise, and quotes are traced back to source.

Resources for this lesson

Web link

Lesson content: 2.2 Working with AI on Interviews, Surveys and Reports

Open resource

Lesson 6

2.3 From Findings to Priorities

QUADRAEDGEBuilding Better Organizations with AI

Module 2 · Lesson 2.3

From Findings to Priorities

Analysis produces a list. Prioritisation turns it into a decision — and that step belongs to people who carry the consequences.

AssessAnalyseDecideBuildGovern

TriangulateConfirm each finding across at least two independent sources.

WeighImpact on mission, feasibility, cost, and what blocks what.

Decide togetherA room of people, with the evidence on the wall.

Confidence

Triangulate before you act

A theme from staff interviews becomes a finding when finance data, project reports or partner feedback point the same way. Where sources disagree, that disagreement is itself a finding — often about communication or differing views of the same problem. Grade each finding as strong, indicative or uncertain, and let the grade drive how much you commit to it.

Bias

Interrogate the analysis, including the machine's

Ask who is missing from the data: field staff, volunteers, partners, people who left, beneficiaries who never complained. Ask whether the analysis simply reflects whoever wrote the most. Models also flatten minority views towards the average and can carry cultural bias in how they interpret indirect or non-English feedback — worth checking deliberately if your organisation works in several languages.

Prioritisation

Sequence, do not just rank

Score candidate priorities on mission impact and feasibility, then look for dependencies: some improvements unlock others, and doing them out of order wastes a year. Three to five priorities per cycle is realistic for a small organisation. Run the final choice as a facilitated session with the evidence visible — AI can prepare the pack, but the room decides.

Key takeaway

Findings become priorities through triangulation, honest bias checks and a human decision about sequence and consequence.

Practical Exercise

Take three findings from your own context. Grade each as strong, indicative or uncertain with reasons, then map them on impact against feasibility and identify one dependency between them.

Summary

Prioritisation requires triangulated findings, deliberate bias checks including the model's, and a facilitated human decision about sequence.

Resources for this lesson

Web link

Lesson content: 2.3 From Findings to Priorities

Open resource

Lesson 7

3.1 Turning Analysis into a Capacity Development Plan

QUADRAEDGEBuilding Better Organizations with AI

Module 3 · Lesson 3.1

Turning Analysis into a Capacity Development Plan

A plan that names an owner, a date and a way of knowing it worked is worth more than a perfect diagnosis with none of those.

AssessAnalyseDecideBuildGovern

SpecificEach action has an owner, a date and a first step.

ProportionateThree to five priorities, resourced honestly.

MeasurableYou defined success before you started.

Structure

What a usable plan contains

For each priority: the finding it responds to, the change you expect, the actions with named owners and dates, the resources required, the risks, and the indicator that will tell you it worked. Written that way, the plan doubles as a donor annex and a board dashboard, which is usually the difference between a plan that is monitored and one that is filed.

Modality

Choose the intervention that fits the gap

Not every gap is a training gap. A knowledge gap needs training or documented guidance; a process gap needs a redesigned workflow; a systems gap needs a tool; a capability gap may need recruitment or mentoring; and a culture or incentive gap needs leadership action that no course will substitute for. Diagnosing the gap type wrongly is the most common reason capacity development money disappears without effect.

Assistance

Where AI helps in the build

Drafting the plan structure from your findings, generating training outlines and facilitation guides, producing tailored templates and checklists, translating materials for multilingual teams, and drafting the monitoring questions. Keep the choice of intervention, the resourcing and the accountability with people.

Key takeaway

Capacity development succeeds when the intervention matches the gap type and every action carries an owner, a date and a measure.

Practical Exercise

Write one complete capacity development action from a real finding, including gap type, three actions with owners, resources, two risks and one indicator with its baseline.

Summary

Usable plans name the finding, the expected change, owners, dates, resources, risks and an indicator, with the intervention matched to the type of gap rather than defaulting to training.

Resources for this lesson

Web link

Lesson content: 3.1 Turning Analysis into a Capacity Development Plan

Open resource

Lesson 8

3.2 Ethics, Privacy and the People in the Data

QUADRAEDGEBuilding Better Organizations with AI

Module 3 · Lesson 3.2

Ethics, Privacy and the People in the Data

Organisational assessment is about people's work, and often about beneficiaries' lives. The safeguards are not paperwork; they are the reason the exercise is legitimate.

AssessAnalyseDecideBuildGovern

MinimiseUse the least personal data that answers the question.

Be transparentTell staff how AI is used and what happens to their input.

Never automateDecisions about individuals stay with people.

Legal

The duties that apply to you

Under GDPR and equivalent national law you need a lawful basis, a defined purpose, data minimisation, a retention limit and a route for people to access or object. Sending personal data to a tool hosted outside your jurisdiction is a transfer with its own conditions. Higher-risk processing may require a data protection impact assessment. Donor agreements often add stricter terms than the law does — read them, because they usually govern beneficiary data specifically.

Ethical

Beyond compliance

Staff who suspect an assessment is a restructuring in disguise will give you unusable data. Say what the exercise is for, who sees what, and what will not happen. Never use AI analysis to rate or rank individual employees. Take particular care with feedback from beneficiaries and from people in precarious positions, who often cannot afford to be candid — anonymity has to be real, not stated.

Fairness

Whose voice the model amplifies

Models perform unevenly across languages and can misread indirect, formal or culturally specific communication as vague or negative. If your organisation works in Montenegrin, Arabic, Kurdish or any language less represented in training data, sample-check the analysis in the original language, and consider having a bilingual colleague review themes before they become findings.

Key takeaway

Legitimacy comes from minimising personal data, telling people the truth about how AI is used, and keeping decisions about individuals firmly with people.

Practical Exercise

Draft a one-page participant information note for an organisational assessment covering purpose, data use, AI involvement, access, retention and the limits on how findings will be used.

Summary

Compliance, transparency and fairness are what make an AI-supported assessment legitimate. Personal data is minimised, AI use is disclosed, and individual decisions stay human.

Resources for this lesson

Web link

Lesson content: 3.2 Ethics, Privacy and the People in the Data

Open resource

Lesson 9

3.3 Governing AI Use and Proving It Worked

QUADRAEDGEBuilding Better Organizations with AI

Module 3 · Lesson 3.3

Governing AI Use and Proving It Worked

A one-page policy your team actually follows beats a twenty-page one nobody reads. Then measure whether any of this improved the organisation.

AssessAnalyseDecideBuildGovern

PolicyApproved tools, permitted data, mandatory review, disclosure.

HabitsVerification built into workflow, not left to individual conscience.

EvidenceReassess on the same framework and compare like with like.

Policy

What a one-page AI policy covers

Which tools are approved and who administers them. Which data may and may not be entered, stated concretely — no personal data, no beneficiary data, no unpublished financials. Where human review is mandatory before output leaves the organisation. When AI assistance must be disclosed, to donors, boards or partners. Who answers questions and how the policy gets reviewed. That is enough for most small organisations, and a policy people can hold in their head is one they will follow.

Practice

Make verification structural

Build the check into the workflow rather than relying on diligence: a reviewer field on any document with AI-assisted analysis, a rule that quotes are traced before a report is signed off, a shared folder of prompts that worked so the team improves collectively. Keep a light record of where AI was used in an assessment — a year later, when someone questions a finding, you will want it.

Learning

Close the loop

Reassess against the same maturity framework after twelve months and compare like with like. Track whether the priorities you chose were actually delivered, whether the indicators moved, and how much staff time the AI-supported workflow saved or cost. Publishing that honestly — including what did not work — is what distinguishes organisational learning from an annual documentation exercise.

Key takeaway

Govern with a short policy people remember, make verification part of the workflow, and reassess on the same framework to prove the improvement was real.

Practical Exercise

Draft your organisation's one-page AI use policy and define three indicators you will use in twelve months to judge whether AI-supported assessment improved decision-making.

Summary

A short, owned AI policy, verification built into workflow, and reassessment against the same framework turn AI use into demonstrable organisational improvement.

Resources for this lesson

Web link

Lesson content: 3.3 Governing AI Use and Proving It Worked

Open resource