Lesson 3 of 6
Build Context: The Organization Evidence Model
Context is the information around a data point that tells you what it means: who said it, when, in which program, compared with what, and measured against which standard. A bare answer can’t be read correctly on its own, by a person or by AI. Build context into one workflow first: give every person a persistent ID, record which survey wave each response belongs to, capture program details, write down your framework and rubric, keep related evidence next to each answer, and ask AI precise questions.
You will make: A written record of the context around one workflow: IDs, survey waves, program details, definitions, rubric and related evidence.
Builds on: Your source map and gap from lesson 1, and your tool checks from lesson 2.
Written for any team that collects data continuously (programs and training, customer and employee feedback, applications, grants and portfolios), especially a team of two or three with no dedicated data staff.
Watch first: why context never connects
In short: From application to exit, different teams collect data in different tools, so the context never connects. Open-ended answers go unread, and the rest sits in spreadsheets, surveys and people’s heads. This 90-second introduction shows the problem and what changes when context stays attached.
Chapters
Three things to notice, each a type of context explained below:
- Identity “Every touchpoint, one record, from application to exit.” Each new answer is read against that person’s history.
- Document Surveys, open text, interviews, case notes and even a 200-page report are analyzed together, instead of sitting in a folder.
- Instruction “Just ask for the insight.” Your question tells the AI what to compare, and it answers in your own terminology.
THE METHOD AT A GLANCE
- Choose one workflow and one decision question.
- Give every person or organization a persistent ID.
- Label every response with its survey wave.
- Record the program details that change what a number means.
- Write down your framework, definitions and rubric.
- Keep related evidence (notes, documents, open text) next to each answer.
- Ask AI precise questions, then review and update every quarter.
What is context in data collection?
In short: Context is the information around a data point that tells you what it means. A number or sentence can’t be interpreted by itself: you need to know who said it, when, in which program, compared with what, and measured against which standard.
Take one sentence from an exit survey: “I feel much more confident now.”
WITHOUT CONTEXT
Positive sentiment. Theme: confidence.
That’s all anyone, or any AI, can say.
WITH CONTEXT
This is Maria’s exit survey. At intake she rated her confidence 2 out of 5 and said she was “scared of interviews.” She attended 10 of 12 workforce sessions, and her coach noted she led a mock interview in week 8.
The same sentence is now evidence of an outcome, not just a mood.
Maria and her data are illustrative.
For AI this matters even more. A model only knows what it’s given at the moment it reads something. Give it a bare answer and you get a generic reading. Give it the surrounding facts and you get a reading that fits your program.
The nine types of context, in four groups
WHO AND WHEN
Identity. One persistent ID links every response, document and score to the same person or organization.
Time. Which survey wave a response came from (intake, mid-point, exit or follow-up), so you can show change.
WHERE
Program. Cohort, site, sessions attended and program design. A 70% completion rate means one thing in a 4-week bootcamp and another in a 12-month fellowship.
MEASURED AGAINST
Framework. Your theory of change, logic model, outcome indicators and definitions, or standards such as IRIS+ and the SDGs. It tells AI which outcomes to look for.
Rubric. Scoring criteria for essays, applications or pitch decks, so “strong leadership” means your definition, not a generic one.
SURROUNDING EVIDENCE
Relational. Other people’s input about the same person: coach notes, mentor feedback, employer checks.
Document. PDFs, interview transcripts and partner reports that would otherwise sit outside the dataset.
Mixed-method. Open-ended answers kept next to the ratings they explain: the words give the why behind the numbers.
Plus one you supply every time: instruction context. The plain-language question you ask, such as “Compare confidence at intake and exit, and flag anyone who declined.” It decides what the AI compares and what it ignores.
What isn’t context
Some things matter just as much but do a different job. Keep them, but don’t confuse them with context:
| Not context | What it is instead | Why it still matters |
|---|---|---|
| Source and citations | Traceability: where a value came from | Lets anyone check an answer, but doesn’t change what the answer means |
| Data cleaning and validation | Data quality | Makes a value correct, not interpretable |
| The response itself | The data | Context is what surrounds it |
How does AI use context?
In short: An AI assistant answers from its general training plus whatever you give it when you ask. It doesn’t know your participants’ history, your program design or your definitions. Where that context is missing, it fills the gap with a plausible guess.
The AI field calls supplying this information grounding or context engineering. The information a model can consider at once is its context window. What isn’t in it, the model can’t use. Here is the same question asked of the same illustrative data:
WITHOUT CONTEXT
Plausible, but wrong. It divided by the 25 people who replied and called them all “completers.”
WITH CONTEXT
Same data. Linked IDs, the survey wave and your definition of “completer” changed the answer.
How do you build context, step by step?
In short: Start with one workflow and one question. Then add each type of context your team controls. A shared sheet is enough to begin. The example follows a fictional three-person training team and one participant, Maria.
1. Choose one workflow and one decision question
EVIDENCE PLAN · ROW 1 (ILLUSTRATIVE)
2. Identity and time: one ID, every wave labeled
A persistent ID stays with one person across every form. Label each response with its wave, and the record becomes a timeline you can compare over time:
01 · INTAKE
Confidence 2/5 · “scared of interviews”
P-041702 · SESSIONS
Attended 10 of 12
P-041703 · WEEK 8
Coach: led a mock interview
P-041704 · EXIT
Confidence 4/5 · “much more confident now”
P-041705 · 30-DAY FOLLOW-UP
Used the skill: Yes
P-0417Maria’s record: one ID, five touchpoints (illustrative).
- Record the wave, not just a date. Intake, mid-point, exit and follow-up are what make a before-and-after comparison valid.
- Use the same idea for organizations. One ID per customer account, grantee or portfolio company links its applications, reports and check-ins.
- Respect consent. Link a survey to a person only when they agreed to that. Keep anonymous responses anonymous, and hold unmatched responses for review rather than guessing.
3. Program: record what changes the meaning of a number
Capture cohort, site, sessions attended and program design next to each record. Without them, results from very different settings get averaged together. A 70% completion rate is strong for a 12-month fellowship and weak for a 4-week bootcamp.
4. Framework and rubric: write down what you measure against
Start with definitions. Each field needs a meaning, allowed values and the mistake it prevents:
Illustrative definitions from four different uses
| Field | Used in | Definition | Watch out for |
|---|---|---|---|
| Used skill | Workforce training | Person says they used the named skill within 30 days of finishing | A missing reply is not a No |
| Jobs created | Grant portfolio | New paid positions filled at the grantee during the grant period, in full-time equivalents | Part-time roles counted differently unless the FTE rule is written down |
| Annual revenue | Accelerator / investor | Most recent fiscal-year revenue in USD, as reported by the founder | Mixing calendar and fiscal years |
| Issue resolved | Customer experience | Customer says their issue was fully resolved within 7 days of first contact | “Ticket closed” in the helpdesk is a different measure |
Then map each measure to what it reports against, and say how closely it matches:
Exact = same definition · Partial = overlaps, rule differs · Contextual = related goal, not a measurement of it
| Your measure | Reports against | Match |
|---|---|---|
| Completion | Contract or grant: “participants completing training” | Exact |
| Used skill within 30 days | Your logic model: short-term outcome “apply new skills” | Exact |
| Jobs created (FTE) | IRIS+ jobs metric (check the current version’s definition) | Partial |
| Issue resolved | Leadership KPI: first-contact resolution rate | Partial |
If you score anything (applications, essays, pitch decks), write each rubric criterion with an anchor. For example, “strong leadership = led a team or project and can name the result.” That way people and AI score against the same bar. For more depth, see the data dictionary guide and the Theory of Change exercise.
5. Surrounding evidence: keep it next to the answer
- Mixed-method: keep each open-ended answer on the same record as the rating it explains. Store your category in a separate field and keep the original words.
- Relational: store coach, mentor or manager notes against the same ID, labeled by who wrote them, so they can be compared with the person’s own account.
- Document: attach transcripts, PDFs and partner reports to the person or organization they describe, not to a shared folder.
How do you ask AI a good question?
In short: Your question is instruction context. Name what to compare, which group, which waves and what to flag, and ask the AI to cite the response behind each claim so you can check it.
VAGUE
“Summarize the survey results.”
PRECISE
“For the spring cohort, compare confidence at intake and exit. Flag anyone who declined, quote their own words, and cite the response for each.”
To draft your first evidence plan from documents you already have, share your survey questions, your last report and any external reporting requirements, then use this prompt:
AI is good at finding what your documents already say and where they contradict each other. Your team decides what the words mean and what the evidence can support. The prompt needs documents only, not participant responses. Check your data policy before sharing anything that identifies people.
How do you keep context current?
In short: Run a handover test every quarter and whenever someone changes roles, and log every definition change instead of quietly rewriting history.
HANDOVER TEST
Give the plan to a colleague. Without asking you, can they tell who each record belongs to, which wave it came from, which program it describes and which definition applies? If not, fix that before adding anything new.
When a definition changes, write down what changed, who approved it and from what date. Leave earlier reports as they were, with a note wherever comparisons are affected.
How does Sopact Sense use context?
In short: Most survey tools collect data without context. Sopact Sense keeps context attached from the moment data is collected, so AI can turn responses into evidence instead of just summarizing them.
It reads data at three levels, and each level uses different context:
| Layer | What it reads | Context it uses | Example |
|---|---|---|---|
| Cell | One answer or document | Rubric and framework | Score one application essay against a 5-criterion rubric |
| Row | One person across all their forms | Identity and time | Summarize Maria’s journey from intake to follow-up |
| AI Assistant | Any question across people, forms and waves | Every type of context, plus your instruction | Compare confidence at intake and exit by site, pull the themes behind any decline, and quote the responses |
Every answer is traced back to the response it came from. That isn’t context, it’s traceability, but it’s what lets your team check each claim before it’s published. People still approve definitions, review AI-generated themes and own the final report.
Go deeper: why clean data still gets ignored
This explainer shows how definitions, framework, audience and presentation affect whether a report is read and trusted.
Impact Reporting: Why Clean Data Still Gets Ignored · 6:07
Add this to your plan
Open your working evidence plan ↗
For your lesson 1 workflow, write one line for each type of context: how people are identified, which survey waves exist, which program details matter, your key definitions, any rubric, and where related evidence lives. Mark anything that is “not stated”.
Check your reasoning
For the training example: identity is the enrollment ID linked to the CRM contact; waves are intake, exit and a 30-day follow-up (day 25–35); program context is cohort and site; “used skill” is defined as self-reported use within 30 days; coach notes are stored against the enrollment ID.
Frequently asked questions
What is context in data collection?
Context is the information around a data point that tells you what it means: who said it, when, in which program, compared with what and against which standard. It includes identity, time, program, framework, rubric, relational, document and mixed-method context, plus the instruction you give AI. Without it, people and AI read the same answer differently.
Is the source of a data point part of its context?
Not strictly. Source and citations provide traceability: they tell you where a value came from so you can check it. Context tells you what the value means. You need both. A cited answer without context can still be misread, and an answer read with context but no citation can’t be checked.
Is a data dictionary the same as context?
A data dictionary is one part of framework context: it defines each field and its allowed values. Context also covers who a response belongs to, which wave it came from, the program around it and the evidence next to it. A dictionary alone can’t tell you that Maria’s confidence rose from 2 to 4.
What is a persistent ID?
It is one identifier that stays with a person, or with an organization such as a customer account, grantee or portfolio company, across every form and touchpoint. It links an application to later surveys without matching names by hand, and it lets each new answer be read against that record’s history. Use it only within the consent people gave.
Can Claude or ChatGPT build this context for us?
They can draft it. Give the assistant your forms, reports and reporting requirements, ask it to list each type of context using only those documents, and have it write “Not stated” for gaps. Your team still agrees the definitions and decides what the evidence can support.
Why do AI tools give different answers about the same data?
Usually because they are missing context. Without IDs, survey waves, program details and definitions, a model fills gaps with reasonable-sounding assumptions, and different tools or prompts assume differently. Giving every tool the same written context and a precise question makes answers more consistent and easier to check.
How often should we update our context?
Review it every quarter and whenever a form, program design or reporting requirement changes, or someone new takes over. Log each definition change with the reason, approver and effective date, and add a note wherever it affects comparisons with earlier reports.
Do we need special software to do this?
No. A shared sheet and written definitions are enough to start, and everything in this lesson works without Sopact. A dedicated platform helps once data arrives continuously from several forms, waves and cohorts, which is where most survey tools stop. At that point, keeping context attached by hand becomes slow and error-prone.