The story should belong to the customer

One of the easiest mistakes to make with generative AI is to confuse the ability to produce language with ownership of the idea behind the language.

For FundRobin, that distinction matters because a funding application carries an organisation’s intent and commitments. It explains what the organisation wants to change, why the work matters and what it will do with the money.

AI can make that work much easier, but I do not think it should own the story.

The useful role is an agent around the organisation

The product direction I find most interesting is a system that does more of the hard preparation around the human decision, rather than trying to produce a finished application at the press of a button.

That begins with smart matching. An organisation should not have to manually read every funding opportunity and work out whether its mission, geography, activities and needs are relevant. AI can compare messy organisational context with messy funder language far more efficiently than a rules-only database.

But a match is more useful when the reasoning is visible. In the current FundRobin experience, the detailed view separates the funder-provided information from the AI’s rationale and retains the source link rather than blending them into one synthetic answer.

The product should help the user understand why an opportunity appears relevant: which parts of the organisation’s profile connect to the funder’s priorities, where the fit is strong, where it is uncertain and what evidence the recommendation used.

That turns the AI from a black-box ranker into something closer to a research partner.

The user can disagree with it.

That is a feature, not a failure.

Transparency changes the role of confidence

A model can sound very confident about a weak conclusion.

That is one reason I care about rationale and evidence in the interface. The purpose is not to expose a chain of hidden model reasoning. It is to expose the product-level evidence the user needs to make a judgement.

For a funding opportunity, that might include the relevant eligibility conditions, the pieces of organisational context that drove the match, the apparent strengths and the areas that need checking.

The user should be able to ask a simple question:

Why does the system think this is right for us?

If the product cannot answer that clearly, it is asking the organisation to trust the model rather than use the model.

That difference becomes more important as the AI takes on more of the workload.

Drafting should broaden the thinking, not replace it

The same principle applies to application drafting.

A single model can produce a polished response very quickly. Polish is useful, but it can create false confidence. The wording may be strong while the argument is generic, the evidence is weak or the framing reflects one model’s interpretation of the problem.

FundRobin currently uses multiple models together in a council-style proposal drafting approach: different model perspectives challenge and improve the response rather than treating the first fluent answer from one system as definitive.

The product intent is to reduce dependence on one model’s point of view and create a more rounded draft. I would not describe that as mathematically eliminating bias; that would be a much stronger claim than the design supports. The useful idea is simpler: different critiques can expose assumptions that one generation path misses.

That mirrors how good human work often improves. A draft becomes stronger when somebody asks a different question of it.

Does this actually answer the funder’s question?

Is the impact claim supported?

Is the language too generic?

What would a sceptical reviewer challenge?

Which part of the organisation’s evidence is missing?

AI can help create that challenge loop cheaply and repeatedly.

Collaboration matters because applications are rarely solo work

Another product lesson is that the person writing the application is often not the only person who holds the information.

Someone understands delivery. Someone else knows the finances. A senior leader may own the impact narrative. A colleague may have the strongest evidence from a previous programme. In many organisations, the application is already a collaborative process; the inefficiency is that the collaboration happens across documents, email, messaging and memory.

The collaboration layer in FundRobin matters because the current workflow is section-based. Contributors can leave notes against sections, Robin can summarise the discussion, suggest next steps or revise from the comments, and version history gives the team a path back when a change is not right.

The application becomes a shared piece of work: AI handles more of the synthesis while people contribute the knowledge and judgement the system does not have.

A useful system should make gaps visible. It should let a contributor see what question needs answering, what evidence already exists and where their input changes the draft.

The AI reduces coordination effort without pretending coordination no longer matters.

The impact story is not a generated asset

There is a harder boundary underneath all of this.

The organisation should decide what impact it wants to have.

That sounds obvious, but generative systems can blur the line because they are so good at producing persuasive language. If an organisation provides a thin description of a programme, a model can elaborate it into something emotionally compelling.

That does not mean the new language represents the organisation’s actual intent.

I want AI to help the customer articulate its story, not manufacture one for it.

The distinction is similar to a good adviser. An adviser can ask better questions, identify evidence, challenge logic, improve structure and help a team express what it means more clearly. The adviser should not quietly decide what the organisation believes.

For FundRobin, human review has a positive role. It preserves authorship where authorship matters, even as the models improve.

We should not submit the application for the customer

The clearest boundary for me is final submission.

A funding application can contain declarations, eligibility confirmations, financial information, terms and other commitments that the organisation needs to read and understand.

Even if an AI system becomes capable of navigating the application portal perfectly, I do not want capability to become silent authority.

The customer should make the final submission decision.

The system can do almost everything around that moment better:

• identify the opportunity;
• explain the match;
• assemble relevant organisational context;
• surface requirements;
• create and challenge a draft;
• coordinate collaborators;
• identify missing evidence;
• prepare the application for review.

Then the organisation reads what it is about to say and decides whether it is willing to say it.

That click is not wasted human effort. It is ownership.

Better automation should make the human moment smaller and stronger

Human-in-the-loop systems can be badly designed.

If the AI does poor work and the user has to rewrite everything, the product has not created much leverage.

If the AI hides its evidence and asks the user to approve a black box, the human checkpoint is theatre.

The design goal is different: automate enough preparation that the remaining human work is concentrated around judgement.

The user should not spend an hour finding the funder’s criteria if the system can retrieve them reliably. They should spend their time deciding whether their programme truly fits.

They should not spend an afternoon turning fragmented notes into a first draft if AI can do that. They should spend their attention making sure the draft expresses the impact they genuinely want to create.

They should not manually track every missing input from colleagues if the product can make those gaps visible. They should contribute the information the machine cannot know.

That is the version of human-in-the-loop I find compelling.

Not humans correcting a weak machine at every step.

AI doing more of the mechanical and synthesis work so that people can spend more of their attention on the story, judgement and commitments that should remain theirs.

The system can help the customer get there.

The destination should still belong to the customer.