To build a theory of change, define the long-term change you want, work backward through the outcomes needed to reach it, and explain how your activities could contribute. Record the assumptions under each connection, the evidence you have, and the evidence still needed. AI can organize source material and draft a diagram or narrative; your team must check the reasoning, consult the people affected, and approve what the model claims.

This practical lesson is for program leaders, evaluators, partners and funding teams building or reviewing one program’s causal explanation. Bring a program description, participant input, delivery records and relevant research. Leave with an evidence register, a short narrative, a review log and a diagram. These are drafts for review—not an automatic promise of a funder-ready submission.

This lesson begins the practical path after the orientation in Impact Measurement and Reporting. If you arrived from Case Intelligence, the same method helps define the outcomes your continuing records should test. For a definition and broader examples, see the Theory of Change guide.

The workflow: extract, write, review and visualize

1. Extract the evidence. Separate the program’s intended pathway from observations, assumptions and missing information.
2. Write the explanation. State why each early outcome could lead to the next, including necessary conditions outside the program.
3. Review the weak links. Check alternative explanations, definitions, missing voices and claims stronger than their evidence.
4. Visualize the approved version. Draw the pathway and label assumptions and evidence status. Keep the diagram connected to the register.

You can do every step manually. The prompts below help organize approved material, but they are not a substitute for facilitation or evaluation expertise. There is no universal four-hour completion time. A well-documented existing program may support a quick first draft; collecting missing participant input or testing a disputed assumption can take much longer.

Start with change, not a list of activities

A theory of change explains the proposed connections between what you do and what changes. A schedule of workshops is not yet a causal explanation. Ask what participants would need to experience or be able to do differently, and what other conditions would need to hold.

The Center for Theory of Change describes working backward from long-term goals, identifying connected outcomes, assumptions, interventions and indicators, then explaining the logic in a narrative. Use that reasoning before choosing the diagram’s layout. A pathway can branch or contain feedback; it need not fit one straight line.

Distinguish resources, activities, outputs and outcomes. Staff time and funding are resources. Coaching is an activity. Sessions delivered are outputs. Demonstrated skills or sustained employment are outcomes. Participants are the people experiencing change; do not classify the number of participants as the program’s resource input merely because it appears first in a spreadsheet.

The example we will build throughout this lesson

Fictional course exercise. A community training program aims to help people move into suitable, sustained employment. It offers instruction, practice, coaching and introductions to employers. Its proposed pathway is: accessible participation → relevant skills → suitable employment → sustained employment and improved economic security.

Use the same reporting exercise introduced in the related lesson: 80 starters; 60 people with known employment status 90 days after exit; 36 of those 60 employed. Twenty statuses remain unknown. In a purposively selected interview group, 12 of 18 people mention transport difficulties. These figures are teaching inputs, not a Sopact customer case study or proof that training produced the jobs.

The program does not yet have a comparable baseline skills assessment or evidence about employment beyond that checkpoint. Those gaps should remain visible. Do not turn “employed at day 90” into “continuously employed for 90 days.” They require different evidence.

Step 1: extract what the sources actually support

Build a register before asking AI to draw a polished diagram. Each entry should identify the source, date, population, definition, finding and limit. A number can be precisely quoted and still be poorly defined or irrelevant to the claim. Its presence does not make a causal link strong.

Use separate columns for source status and claim support. Source status might be available, incomplete or not supplied. Claim support describes what the record allows you to say. A reviewed attendance register may support an attendance count; it does not establish that attending caused an improvement in skills.

Task: organize the supplied program evidence for a Theory of Change.
INPUTS: [program description, participant input, dated records and research excerpts]
Use only material supplied or actually accessible. If you cannot read a URL, say so.
Return: pathway element | source ID/page/date | exact supporting passage or value |
population and period | observation or intended outcome | missing evidence |
assumption requiring review.
Keep resources, activities, outputs and outcomes distinct. Do not treat a number
as strong evidence merely because it is quoted. Do not infer causality from a
before/after result or attendance count. Mark proposed additions as proposals.
List conflicting definitions and unreadable sources for the human reviewer.

Review the output against the records. Check names, populations, dates and quotations. If a source says “36 employed,” do not add employer verification, a twelve-month window or a wage threshold that the source never supplied. Record corrections rather than silently treating the model’s wording as evidence.

Save the source snapshot and approved register version. Fixed instructions can improve consistency, but generative AI can still vary or make errors. For counts and rates, use an explicit calculation outside free-form text generation and check it. The NIST generative-AI risk profile identifies confidently incorrect output as a risk; source checking remains necessary.

Step 2: write the causal explanation without upgrading the evidence

Write the narrative in three parts: the problem and people affected; the proposed pathway; and what the evidence supports or leaves unresolved. Explain the connection between stages, rather than repeating the labels in a diagram.

For the fictional program, relevant skills could improve access to suitable work if employers need those skills, recruitment is accessible and participants can reach the workplace. Coaching alone cannot guarantee local vacancies, transport availability or suitable job conditions. These are assumptions or contextual conditions to investigate.

Write a draft Theory of Change narrative using the APPROVED REGISTER below.
INPUTS: [approved register, intended long-term change, stakeholder feedback]
Include: problem and affected people; proposed sequence of outcomes; activities;
assumption under each connection; evidence supporting each claim; unresolved gaps.
Distinguish observed results from intended outcomes and proposed explanations.
Cite register IDs for factual statements. Do not invent verification, dates,
benefits, participant views or research. Do not describe the draft as validated.
End with the questions the team must resolve before external use.

Example narrative, for review: “The program proposes that accessible training and coaching will help participants develop skills relevant to available work. Employer introductions may help participants find suitable opportunities, provided those opportunities exist and working arrangements are accessible. Of 80 starters, 60 have a known employment status 90 days after exit and 36 are employed. This describes employment at the checkpoint, not the program’s causal effect. Comparable skills assessments, better follow-up coverage and participant accounts are needed to examine the pathway. Sustained employment and economic security remain longer-term outcomes to investigate.”

This wording is less definitive than “our training creates lasting employment,” because the supplied evidence does not support that stronger statement. The useful next question is which part of the explanation the team most needs to test—not which sentence sounds most persuasive.

Define indicators that could test the pathway

An indicator needs more than a label such as “success.” Specify who is measured, what qualifies, the relevant time window and any target. The Center for Theory of Change’s indicator guidance distinguishes population, target, threshold and timeline. Set these for your context rather than copying an old example’s wage or success threshold.

A measurement plan also needs a source, collection owner and schedule. If you add a confidence question, explain what it measures and why it belongs in this pathway. One confidence item cannot replace a skills assessment, employment record or causal evaluation. Do not add extra sensitive questions merely because a framework has an empty box.

Step 3: review assumptions and alternative explanations

Invite program staff and people affected by the program to examine the draft. Use an evaluator or other specialist when the intended claim requires it. Ask what would make a link fail, which experiences the data exclude, and what could explain the observed outcomes without the program.

Review this draft against the APPROVED EVIDENCE REGISTER.
For each important claim, return: claim | source | supported / partly supported /
unsupported | alternative explanation | missing definition or evidence |
recommended correction | owner to resolve it.
Check denominators, time windows, unknown statuses and whether observations are
being described as causal effects. Check whose experience is missing and potential
negative outcomes. Do not invent a counterfactual or assume a participant quote
represents everyone. Prioritize fixes by how much they affect the decision.
Rewrite one unsupported sentence using only the supplied evidence.

These corrections should change the register and narrative together. A reviewer’s disagreement is useful evidence about an unclear definition or assumption; it is not a reason to hide the weaker result. Preserve unresolved issues, the person responsible and the planned resolution date.

Step 4: draw a diagram a reader can question

Use the approved pathway and register to draw the model. Keep outputs distinct from outcomes, place assumptions beside the relevant connections, and label evidence status in words. Color can help orientation, but a reader must be able to understand the model without distinguishing red from green.

Create a draft diagram from the APPROVED PATHWAY and REGISTER only.
Show resources and activities separately from outputs and outcome stages.
Use an accessible left-to-right or top-to-bottom layout. Label each proposed
connection and its key assumption. Mark observed evidence, proposed outcomes and
missing evidence with text, not color alone. Include the version and review date.
Add no figures or causal claims not approved in the register. Provide a text
alternative and a checklist of diagram entries to compare with the source.

This table is an accessible text version of the diagram. The arrows describe proposed relationships; they are not proof. A final graphic should carry the same labels and limits. Compare every figure and status with the register before sharing it.

Watch the AI-assisted workflow

Watch the 9-minute 50-second workflow demonstration. Treat any example grades or model output as prompts for review. A quoted figure does not establish evidence quality, and AI output should be checked against the actual sources. Browse the video library.

How to keep the model useful after the first workshop

Assign a version, review owner and date. Record what changed: an outcome definition, assumption, activity, source or interpretation. Keep the old version when reporting a previous period, so a reader can understand which model informed that report. Update because evidence or the program changed, not simply because a different AI response reads better.

For one weak connection, name a feasible learning question and the evidence needed. For example: “Are transport difficulties preventing participation, and whose experience are we missing?” Decide how to ask, who needs access to the answers and when the team will review them. A theory of change becomes useful when it influences a real decision.

Sopact Sense can help organize incoming forms, documents and interview evidence around the relevant program or contact and prepare source-linked analysis for review. Your team still approves the outcome logic and claims. Start with the manual register; consider a connected workflow when maintaining it across partners and reporting periods becomes difficult.

Choose which assumption to investigate first

Once the pathway is visible, a useful advanced task is deciding what deserves attention first. Do not grade an entire theory as reliable because its diagram contains quotations or numbers. Review the consequence of a mistaken assumption, the evidence currently available and whether the team can act on the answer.

Fictional example: The training program assumes relevant jobs are accessible to participants. Its records show employment at one checkpoint but do not establish job suitability or how opportunities were obtained. Another assumption concerns whether a minor workshop format preference affects enjoyment. If the decision is whether to expand the employment pathway, access to suitable work may deserve earlier investigation. This priority is a reasoned team decision, not an automatic score.

Ask staff and affected people to review the priority. A missing statement on a website means it was not available in that source; it does not establish that the organization never measured it. Seek existing evidence before adding a new collection burden.

Practice: select two assumptions and explain why one comes first. Write a source-backed statement of what is known and a separate proposal for the next inquiry. Save this with the model version. When evidence changes, update the explanation and retain what earlier reports used. One new survey item may inform a question; it does not validate the whole causal pathway.

Frequently asked questions

Can AI build a complete theory of change?

It can draft a pathway, narrative and diagram from supplied material. It cannot establish that assumptions are true or that an intervention caused an outcome. Treat its work as a proposal to inspect with staff, stakeholders and appropriate evaluation expertise.

What should I provide before using the prompts?

Supply the program description, intended change, affected groups, actual activities, dated source records and participant input. Include research relevant to the proposed links. Identify missing material and access restrictions instead of asking the model to fill gaps.

Is a quoted number strong evidence?

Not automatically. Check what the number measures, who it covers, the time window, its source and its relevance to the claim. A reliable count of sessions can support a delivery statement while saying little about outcomes or causality.

How is a theory of change different from a logic model?

A theory of change explains why and under what conditions change is expected. A logic model organizes resources, activities, outputs and outcomes into a practical structure. They can support each other; the next course lesson turns this pathway into a logic model.

How long should the narrative be?

Use enough space to explain the pathway, assumptions, evidence and limits clearly. A brief overview can link to a more detailed register. Follow a funder’s format where required, but do not remove material uncertainty just to meet an arbitrary two-page target.

What if there is no baseline?

Say so. Report what current records support and avoid presenting a current level as measured change. Consider whether appropriate historical evidence exists and plan future collection. Do not invent a baseline or assume a participant’s recollection is equivalent to a prior measurement.

Can the same prompts support a foundation reviewing partners?

Yes, with a separate register for each partner and its context. Apply consistent review questions while respecting different pathways and evidence. Do not rank unlike programs solely by how many boxes contain numbers or how polished the narratives look.

When is the model ready to share?

When the responsible team has checked sources, definitions, assumptions and the diagram, and unresolved issues are visible. Readiness depends on the audience and purpose. Neither a green score nor an AI review guarantees approval, validation or funding.

Practice and continue the course

Use one program description to complete the four artifacts. Ask another person to trace three claims from diagram to register to source. Revise any statement they cannot reproduce. Keep one unresolved assumption as the learning question for the next collection cycle.

Related practice: build a logic model from this pathway. Case Intelligence readers can continue to turn outcomes into a collection workflow. For presentation guidance, use How to Write an Impact Report and report examples.

By Sopact Academy · Revised September 12, 2026. All course-example figures are fictional. Method references: Center for Theory of Change and NIST, linked above.