Judgement as infrastructure, part three: the last video should make the next video smarter

The promise of AI isn’t that every production starts faster. It’s that fewer productions should have to start from zero.

A company commissions a customer video: an agency, a crew, lighting, a location, customer approvals, schedules and several rounds of editing. The interview produces an hour of material. A three-minute film comes out of it, maybe with some social cuts alongside it. The project closes.

Six months later the next customer shoot begins, and someone opens a blank document and types the same four questions: tell us about your business, what challenges were you facing, why did you choose us, what difference has the product made.

The organisation has done this before, possibly dozens of times, but the brief behaves as though it hasn’t. That is the part I am increasingly interested in.

Part one of this series was about preserving recurring editorial judgement rather than repeatedly applying the same corrections at the end of a workflow. Part two was about the danger of turning every previous decision into a permanent rule.

Video makes the argument much easier to see because the cost of forgetting is so visible. A production does not only create a film. It creates a second, less obvious output: learning about how the next production should be better — and most organisations keep the first while throwing away the second.

---

Every production creates a hidden second output

An interview guide produces evidence about which questions worked, and a transcript shows where the customer became specific and where they reverted to safe language. An edit reveals which proof points were missing. A producer discovers which visual explanation would have made the story clearer. A customer correction exposes a claim that should have been verified before filming. A stakeholder review reveals that the brief did not contain enough context for the agency to understand the real story.

None of this is especially glamorous, but it is valuable, and the organisation has paid to learn it. Yet most of those lessons remain scattered across transcripts, email threads, comments, people’s memory and project folders. When the next production starts, the team might reuse the previous brief as a template, but that is not the same as carrying the learning forward. Templates preserve structure. What I am describing is preserving judgement.

There is a difference between remembering that the previous interview had eight questions and remembering that the broad opening questions repeatedly produced polished but unusable answers, while asking for a specific moment, consequence or before-and-after comparison produced the material the editor actually used. That second thing is the useful knowledge.

AI makes the learning layer practical

Before generative AI, preserving this properly would have been expensive. Someone would need to read previous transcripts, compare briefs, inspect changes between edits, remember stakeholder feedback and turn it all into a useful set of lessons. On a large production programme, that quickly becomes another research project.

AI changes the economics. It can compare the source material cheaply. It can surface candidate patterns across transcripts, briefs and revisions, asking which questions repeatedly produced generic answers, which evidence had to be chased afterwards, which claims were corrected and which briefing gaps kept appearing.

But that does not mean the model gets to decide what becomes a rule. One unusual customer is not a pattern. One stakeholder preference is not an editorial standard. One successful edit may simply have been a good producer making the best of weak source material. This is where the argument from part two becomes practical: the AI can surface the candidate lesson, but the editor still decides whether it deserves to survive.

That produces a controlled loop:

previous transcripts + briefs + edits + review evidence → candidate lessons → human validates what is reusable → new story and evidence gaps → stronger brief → human interview and production → new evidence and review → learning log → next production starts stronger

This is much more interesting than asking ChatGPT for ten interview questions. Question generation is cheap; brief intelligence is the valuable part.

A better brief should know what the organisation still needs to learn

The normal interview guide begins with questions. I think a stronger one should begin one step earlier: what are we actually trying to find out, what do we already know from research and previous customers and previous interviews, what evidence is still missing, what assumptions are we testing rather than quietly treating as true, and what did the last few productions teach us about how to get useful material from this kind of conversation. Only then do the questions become interesting.

Take something generic:

How has technology helped your business?

It gives the customer an enormous space in which to say almost anything, which usually means they say something safe. A more evidence-led route might be:

Before you changed the way you worked, what was the task most likely to keep you working late?

And the brief can prepare the interviewer for where the answer might go:

• Can you remember a particular week when that happened?
• Roughly how much time did it take?
• What was the consequence for you or the team?
• What specifically changed?
• How did you know the change had worked?
• What do you do with that time now?

The point is not to turn the interviewer into a machine following a decision tree. Quite the opposite: the aim is to make the human interviewer better prepared to listen. A good interviewer should still abandon the planned route when the customer says something unexpected. The system’s job is to make sure the interviewer enters the room knowing more about what has already been learned, what remains unknown and where useful evidence may be hiding.

GEO gives the system another useful input

I have already written about this from the other direction in Are you getting the GEO value from all that video?. That piece asks whether organisations are getting enough useful first-hand knowledge out of expensive customer and event shoots. If visibility analysis shows that an audience is asking an important question and the organisation lacks a strong answer, that gap can be brought into pre-production and the interview can be designed to capture better evidence.

This adds another loop: GEO or audience analysis can tell you that a question matters and the current evidence is weak, while the production-learning system can tell you what previous interviews already taught about investigating that kind of question well. Together:

audience / GEO question → what evidence is missing? → what have previous interviews already taught us? → what should we investigate this time? → human interview → new first-hand evidence → video + transcript + supporting knowledge where useful → audience response, discoverability and production learning → next brief

There is an important boundary here. GEO should not turn customer interviews into keyword theatre. A customer should not be made to answer search queries on camera because somebody wants an AI citation. Visibility analysis identifies the evidence gap; editorial judgement translates that into a natural human question capable of producing credible first-hand material. That is a much healthier relationship between search intelligence and editorial work.

Agency briefing may be the clearest business case

This becomes particularly useful when an external agency is involved. An agency can be excellent at production and still know almost nothing about the editorial history behind the brief it receives.

It may know the campaign objective, deliverables, customer name and product message. It probably does not know that the last four customer interviews produced weak material whenever the conversation started with product benefits. It may not know which claims repeatedly needed verification, which customer language proved unusually useful, which story assumptions collapsed in the edit or which footage internal editors repeatedly wished had been captured. So the agency has to rediscover some of the organisation’s knowledge itself, which is wasteful.

Imagine instead briefing the next production with:

• the audience question being investigated
• what the organisation already knows
• the evidence still missing
• the strongest and weakest question patterns from previous interviews
• known proof requirements
• recurring production traps
• relevant GEO questions where useful
• examples that demonstrate the editorial standard
• previous lessons that have survived human review

The agency still owns the production craft. The client simply stops making the agency start with amnesia — which is the kind of AI use I find much more credible than trying to automate the producer.

This is not really a video idea

Video is the easiest example because the inputs and costs are obvious, but the principle is much broader: a customer story should improve the next customer-story brief, a research programme should improve the questions in the next research programme, a webinar should improve the next moderator guide, an executive interview should improve the next interview, an event should improve the next event brief, and a report should leave behind more than the finished report.

This changes what the Learn stage means inside the Editorial Intelligence Framework. Learning is not only analytics after publication. It has at least two forms: external learning — what did the audience, customer, market or search environment tell us — and production learning — what did making the work teach us about how to investigate, brief and execute the next one. Both should feed the next cycle, and that includes performance, but also missed evidence, review patterns, production friction, collaborator feedback, recurring human judgement and the things the team discovered too late to fix cheaply.

The result is not simply a content operation that produces faster. It is one that remembers.

Fewer blank documents

I don’t think organisations need another grand AI transformation programme to test this.

Take one repeatable production type. Gather a small set of previous transcripts, briefs, edits and review comments. Ask the system to surface candidate lessons. Have an editor reject most of them. Keep the few that appear genuinely reusable. Put those into the next brief and see whether the resulting interview produces stronger evidence or avoids a problem that previously emerged too late. Then learn from that too — if it works, repeat it; if it doesn’t, stop.

The point is not to build a machine that knows how every video should be made. It is to stop paying for the same lessons over and over again.

The last video should leave behind more than a video. It should leave the next brief smarter.