Logic Model Template: Build One With AI From a Client's Programme

A grant writer can use AI to build a logic model by pasting a client's programme description into a prompt that sorts it into inputs, activities, outputs and outcomes, then running a second prompt that checks each causal link and indicator for gaps. The client still confirms every row, and the funder's own wording decides the final labels.

Searching for a logic model template returns blank tables and fill-in diagrams. The table is the easy bit. The work is in turning a client's rambling description into rows that connect, spotting the places where the programme claims more than it does, and matching a funder's vocabulary. That is where AI helps, and where it quietly misleads if you let it write the story.

Follow me on Instagram@sagnikteaches

The five columns and what each one must contain

A logic model reads left to right: what you put in, what you do, what you produce, what changes in the short and medium term, and the long-term aim. Different funders rename the columns, but the logic is the same.

Connect on LinkedInSagnik Bhattacharya

Many templates add two more boxes: assumptions (what has to be true for the chain to work) and external factors (things outside the programme that could help or hinder). They are where most real weaknesses show up, and they are the part AI is best at surfacing.

Subscribe on YouTube@sagnikteaches

Gather the raw material before opening an AI tool

AI cannot create information the client has not given you. Before you start, collect:

1. The client's programme description, as written or as notes from a call.
2. The budget, with staffing and quantities.
3. Any past monitoring figures: attendance, completion, satisfaction.
4. The funder's guidance, template and scoring criteria.
5. Any existing evidence the client can cite, such as their own survey results.

If you are using a general assistant, strip identifying details and sensitive participant data first, and check the tool's privacy settings. ChatGPT's and Claude's business plans do not train on your content by default, and their personal plans have a model-training switch in privacy settings; confirm the current wording on the vendor's page. Participant stories and anything about named children or vulnerable people should not go in at all. For a comparison of tools, see the best AI app for grant writing, and for free-tier limits, free AI for grant writing.

Step one: ask AI to build the first draft table

Use one prompt, with a strict instruction not to invent. Here is a version that works with a typical programme write-up. The sample programme is an invented, generic after-school reading club run by a small charity.

You are helping a grant writer build a logic model from a client's programme notes.
Use ONLY the notes below. Do not invent activities, numbers, partners or results.

Produce a table with these columns: Inputs | Activities | Outputs | Short-term outcomes |
Medium-term outcomes | Long-term impact. Then list "Assumptions" and "External factors"
separately. After the table, list "Not in the notes" for anything you needed but could not
find (for example numbers, timings, evidence, indicators).

Write outcomes as changes in participants, not as things the programme does.

NOTES:
A small charity runs a weekly after-school reading club for children aged 7 to 10 at two
primary schools. Two part-time coordinators and ten trained volunteers run 30 sessions a
year in each school. Each session has 12 children reading in pairs with a volunteer. The
charity also gives each child a book to keep each term and sends a tips sheet to parents.
Budget: $38,000 a year, mostly staffing. Last year's records: 70% average attendance;
teachers' feedback form says children "seem more confident reading aloud".
Funder wants: improved reading confidence and enjoyment.

A realistic reply (illustrative) puts "two part-time coordinators, ten trained volunteers, two school rooms, $38,000 a year, books for each child" under inputs; "weekly paired reading sessions, termly book gifts, termly parent tips sheet" under activities; "60 sessions a year, 24 children per week reached, three books per child per year" under outputs; "children read aloud more willingly" as a short-term outcome; "children read more often at home" as medium-term; and "stronger enjoyment of reading" as impact. Under "Not in the notes" it lists: a measure of reading confidence, how children are selected, whether parents engage with the tips sheet, and evidence beyond the teachers' comment.

What to fix straight away: the 24 children per week figure comes from 12 children across two schools, which is right only if each school runs one session a week. The notes say 30 sessions a year in each school, so this is plausible but should be confirmed. Also "three books per child per year" is an inference from a termly gift; check whether there are three terms. These are small, but a logic model with a number the client cannot recognise loses trust when it reaches the funder.

Step two: gap-check the causal links

The second prompt is where AI adds value beyond formatting. Give it the draft table, ask it to act as a sceptical assessor, and keep the output to questions and flags.

Act as a sceptical grant assessor reading this logic model. Do not rewrite it.
For each arrow between columns, say whether the link is: supported, plausible but
unevidenced, or a leap. Then list:
1. Activities with no matching output.
2. Outputs that do not lead to any outcome.
3. Outcomes with no indicator or way to measure them.
4. Outcomes the programme cannot reasonably claim by itself in the stated time.
5. Assumptions that are hidden (things that must be true for the chain to work).
Quote the row text you are commenting on. Keep each point to one sentence.

On the reading club, a good reply flags: the parent tips sheet is an activity with no outcome attached (do parents change anything?); "children read aloud more willingly" has no indicator, only a teacher's impression; the jump from "read more at home" to "stronger enjoyment of reading" is a leap without a measure of enjoyment; and a hidden assumption that children can attend consistently, with transport home arranged. Each flag becomes a question for the client, not a sentence to paste.

The if-then read-through

You can also do a manual check: read the model aloud as a chain of "if... then...". If volunteers are trained and sessions run weekly, then children read aloud in a safe setting. If children read aloud in a safe setting, then their willingness to read aloud grows. If their willingness grows, then they read more often at home. Any "then" that sounds forced is where an assessor will stop. Ask the AI to propose what evidence would make a weak link believable, and put those on the client's to-do list.

Step three: make outcomes measurable

An outcome without an indicator is a wish. An indicator is how you would know the change happened, with a method and a rough target. Ask the AI to suggest several per outcome, then pick the ones the client can actually collect.

The last row is a useful one to show a client. AI will happily propose test scores because they sound rigorous, but a charity with two part-time coordinators cannot collect them, and an application that promises them sets up a failed report. Cross out indicators the client cannot gather.

Step four: adopt the funder's wording

Funders name things differently. One calls the left column "resources", another calls outcomes "results", a third wants "impact" labelled as "goals". Paste the guidance and ask AI to map your table to the funder's headings and flag any definitions that differ.

Here is a funder's logic model guidance and my draft table. Rename my columns to match
the funder's terms. List any term where the funder's definition differs from mine (for
example if "outputs" there includes participant numbers). Quote the funder's text for each
change. Do not change the content of any cell.

Say the funder says "an outcome is a change for participants that the funder could verify" and limits the application to three outcomes. The AI reply might propose merging the two short-term outcomes and moving the parent-engagement outcome to "other benefits", quoting the guidance. You still decide, but the quotation lets you justify the choice to the client. If the guidance includes a statement on AI use in applications, follow it, and see whether funders reject AI-assisted applications before you submit.

When the client gives you thin notes: interview questions

Often the programme description is two paragraphs and a budget. The "Not in the notes" list then becomes your interview agenda. Ask the AI to turn that list into questions, then edit them into a short call script. For the reading club, a useful set is:

1. How are children chosen to take part, and who decides?
2. What does a typical thirty-minute session look like, minute by minute?
3. What do volunteers record after a session, if anything?
4. What would a teacher say has changed for a child who has been coming all year?
5. What stops children attending, and what have you tried about it?
6. Which result are you proudest of, and what evidence do you have for it?

Type the answers up and feed them back into the build prompt as an addendum. The second version of the model is almost always better than the first, because the interview answers replace the AI's gaps with the client's facts.

Reviewing a logic model the client already has

Sometimes a client arrives with an old model written for a previous funder. The gap-check prompt works on it just as well, and it is often the more valuable use. Say the old model has this line under outcomes: "Increased community cohesion". The AI flags: no indicator, no link from any activity, and a timeframe longer than the grant. You take that to the client and ask what, specifically, they see changing between neighbours. Perhaps the answer is that volunteers from different streets now know each other, and that eight volunteers have stayed for more than two years. That becomes an outcome ("volunteers form lasting connections") with an indicator (volunteer retention and a short volunteer survey) in place of a phrase nobody could verify.

Before: "Increased community cohesion and wellbeing through reading."

After: "Volunteers report knowing at least three people in the project they did not know before joining (annual volunteer survey); at least 8 of 10 volunteers return for a second year (volunteer records)."

Both lines are illustrative, and the targets are the client's to set, not yours or the AI's. A target a client cannot meet is worse than no target.

A quick check on cost per output

A logic model quietly invites a funder to divide the budget by the outputs. Do it yourself before they do. The reading club's $38,000 divided by 60 sessions is about $633 a session, and divided by the 24 children in the room each week, if the same children attend all year, about $1,583 a child. Whether that looks reasonable depends on the funder and the programme, and it is a number the client should recognise before it goes into the narrative. If a figure looks high, the model may be undercounting outputs. Perhaps children drop in and out, so more than 24 are reached over the year, and the right output is "children reached" and "children attending at least ten sessions". Ask the AI to compute the ratios, then check the arithmetic yourself, because language models make small slips with division.

Turning the table into the narrative paragraph

Many applications ask for the logic model as a table and as a short written explanation. Once the table is signed off, ask the AI to write the paragraph from it, with the same no-invention rule, then read it against the table.

Write a 120-word explanation of this logic model for a funder, in plain English and
the present tense. Use only the table. Name the inputs, the two main
activities, the outputs with their numbers, and the outcomes with their indicators. Do not
add any claim that is not in the table. Do not use the words "transform" or "empower".

Compare the output line by line with the table. Typical corrections: a figure rounded up, a short-term outcome promoted to a long-term one, or a softened "contributes to" hardened into "will achieve". Keep the language of contribution for the long-term impact, and the client's own words wherever they are stronger than the model's.

The finished example

After two rounds with the client, the reading club's model looks like this (illustrative figures, client-confirmed):

Assumptions: children can attend regularly; volunteers stay for the year; schools keep the room available. External factors: teacher staffing, family circumstances, other reading support in the school. Compare that with the first draft: the vague outcomes became measurable, the unconnected tips sheet now links to the parent survey, and the test-score indicator is gone.

A second sector in brief: a community kitchen

To show the method is not specific to education, say a community kitchen serves 80 meals a week and wants funding to add a cooking class. The AI build step produces activities (weekly class, shared meal) and outputs (40 classes, 12 participants per class). The gap-check flags that the claimed outcome, "improved health", has no link from cooking skills to health in the notes, and the timeframe is far too short. The grant writer asks the client what they can observe: participants cook one new meal at home, and attendance at the shared meal holds. The outcomes become "participants report cooking a meal from scratch at home" and "participants stay connected through the shared meal", with a four-question survey. "Improved health" moves to the long-term impact box with the word "contributes".

A mistake that looks convincing: the confident, wrong outcome

Here is a failure worth recognising. Asked to complete a model with minimal notes, a language model may write: "Children's reading attainment improves by two grades within one year." It sounds like an outcome, it matches funder language, and it has no basis in the notes. If it reaches the application, the funder will expect it in the report. The cure is the "Not in the notes" instruction, which makes the model hand you its gaps instead of filling them, plus a rule that every number in the final table traces to a document the client supplied.

Time and cost

For a typical small programme, expect about 20 minutes to gather the documents, 15 minutes to run the build prompt and read the output, 20 minutes on the gap-check and a client call of about 30 minutes to answer the questions, then 15 minutes to match the funder's wording and finalise. Call it about two hours of your time compared with perhaps three to four hours of drafting from scratch. The saving is real but modest; the larger benefit is that the gap-check surfaces weaknesses before the assessor does. A free or entry-level plan is enough for this work, as the prompts are short. For pricing your time when drafting gets faster, grant writer rates when AI drafts faster sets out options.

Checks before you send the model to the funder

• Every number traces to a client document or is marked as an estimate the client confirmed.
• Every outcome describes a change in people and has an indicator and a collection method.
• No outcome is promised that the programme cannot measure or influence in the grant period.
• Column names and definitions match the funder's guidance.
• Assumptions and external factors are written down, not left out.
• The client has read the final table and signed it off by email.
• Any funder rule on AI disclosure is followed.

The logic model then feeds the rest of the application: the needs statement, the evaluation plan and the budget narrative should all use the same numbers. For prompts on the needs and outcomes sections, see ChatGPT prompts for grant writing, and for the larger process, how to use AI to write grant applications. If you want to set up a repeatable build and gap-check routine for every client, that is the kind of project my AI implementation consultation covers.

Questions grant writers ask about AI and logic models

What is the difference between a logic model and a theory of change?

A logic model is a table or diagram of inputs, activities, outputs and outcomes. A theory of change explains why each step should lead to the next, including assumptions. Many funders accept either, so check the guidance and use their word for it.

Can AI invent the outcomes for me?

It can propose them, and that is the risk. Outcomes must reflect what the client's programme can plausibly change and measure. Use AI to draft options from the client's own notes, then have the client confirm each one is realistic.

How many outcomes should a logic model have?

Fewer than most drafts. Two or three short-term outcomes, one or two medium-term and a single long-term aim is a readable pattern, but follow the funder's template if they give one. Each outcome needs at least one indicator and a way to collect it.

Should I put the AI-built table straight into the application?

No. Check every row against the programme documents, make sure the wording matches the funder's guidance, and ask the client to sign off. Also check the funder's guidance for any rule on AI use in applications, such as a disclosure requirement, and follow it.

What is the most common gap AI finds?

Activities with no matching output, and outcomes with no indicator. The next most common is a jump from output to long-term impact with no outcome in between, for example sessions delivered leading directly to better lives.

Further reads

• AI Grant Writing Tools Compared for Small Non-Profits — Tools compared for small non-profits.
• AI Grant Writing Software: What Each Costs a Freelance Grant Writer — What each tool costs a freelance grant writer.
• Grant Writing AI Agent: What It Can Draft in a Business Grant Bid — What an AI agent can and cannot draft in a bid.
• Can AI Write a Sports Club Grant Application? What Funders Expect — A worked sector example with funder expectations.
• AI Tool Discounts for Charities and Non-Profits: What to Claim — Discounts your non-profit clients can claim.
• AI Consultant for Nonprofits: What to Hire For and What to Pay — What non-profits pay for AI help.
• Grant Budget Narrative: Draft It With AI and Reconcile Every Figure — Line-by-line budget justification drafted with AI, plus a figures register that proves every number in the narrative matches the spreadsheet.
• Grant Report Template: Draft Funder Reports With AI From Project Data — A filled-in funder progress report drafted by AI from project data: outcomes table, narrative, lessons learned, and the evidence checks before sending.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: widely published logic-model guidance from foundations and public funders (inputs, activities, outputs, outcomes, impact), OpenAI's business-data page and Claude's pricing page (checked 4 November 2026).