An AI production assistant for film: what it answers, what it changes, and what it never does.

Quick answers

What is an AI production assistant for film? Software that reads the schedule, the cast and crew, the locations, the to-dos and the budget together, and answers in plain language. In Spring Pro it is Gaffer, a page on the production desk. It is not a PA, the entry-level crew job that shares the name, and its own name is the crew job of chief lighting technician.

What can it answer that one tool cannot? Questions that cross every page: are we ready for Day 1, what moves if a scene moves, what a day costs, and what is not priced yet. Each needs the schedule, the people, the places and the money at once. The sections below show each one on a real sample.

Can it change things? Yes, and its brief tells it to act only when you ask. That is an instruction to the model, not a lock. It has 15 tools that read and 23 that change the production, plus an undo, and each change is saved the way an edit on the page is saved. One "undo that" reverts everything it did in its last run.

What does it never do? It does not send messages, pay or move money, or change the words of your script, and on the production desk no tool for those exists. It drafts a call sheet or a message as text for a person to send. It is also told never to guess a number, a date, a name or an id, but that is a rule for the model, not a lock, and the checked answer below shows where it slipped on owners.

Can it be wrong? Yes. NIST calls confident, wrong output confabulation, and says it is a natural result of how these models are built. On the one real answer we checked, every date, count and amount matched the project. The owners it named were not recorded in the project: no to-do has one, so it named people from the crew list. Check numbers and names before you act on them.

Does it replace the 1st AD, the UPM or the line producer? No, and this guide is written on that footing. It does the reading and the drafting, and a person decides and sends. Guild agreements now speak to this, and the section on agreements below says what we read.

Why the useful questions cross every page

A production office does not ask one-document questions. "Are we ready for Day 1?" needs the schedule for the date and the warnings, the call sheet for the times, Locations for the status and the address, the crew list for the contacts, the to-dos for what is late, and the budget for what it costs. In most tools those live in different files or different products.

Someone in the production office has to hold all of it at once, whether that is a line producer, a UPM or a production coordinator. A recap by a Film Independent writer of a 2017 talk by line producer Angela Sostre says that once a script is ready a line producer is usually called on to create a realistic schedule and a rough line-by-line budget. The writer's own example is money saved on travel offsetting the cost of an extra half-day of shooting. The recap also says Sostre notes that actor availability can be an issue.

That is the work an assistant can take the reading off. The table shows where each fact behind the Day 1 question is kept in Spring Pro.

Gaffer's tools read all of these in one run. The Reports page and Gaffer also count risks with the same function in the app, so they share one definition of a risk.

The film we asked about

Midnight Diner is a short crime film of 7 scenes and 2 1/8 pages. Its schedule has four shoot days that carry scenes and two empty days after them. It has 44 people on its lists: 5 cast, 32 crew and 7 vendors. Its budget was locked on 4 October 2026.

One thing to know before the numbers. The schedule starts on Monday 5 October 2026, and the numbers and screens on this page were made on Monday 12 October 2026. Every day is already in the past, so a to-do due before 12 October counts as late. That is why the sample looks so far behind.

Every day lacks a crew call and every location is still at Scouting. The to-dos are not assigned to anyone: none of the 90 to-dos has an owner set. Keep that in mind for the checked answer below.

What it reads: 15 tools, none of which change anything

Gaffer works through tools, small functions that read or change the production. Fifteen of them only read. The table groups them by what they reach, from the tool descriptions in the code.

It does not see the storyboard pictures, only how many frames a scene has and whether one is picked. It does not see receipts, only whether one is attached. Weather, sunrise and turnaround are not among its tools, and the limits section says what it does about turnaround instead.

How a question runs, from the brief it works to

Gaffer follows a written brief in five steps. It reads first and is told never to guess a number, a date, a name or an id. It traces the ripple of the question through the story, the schedule, the people, the money and the risk. It acts only when asked, then reads back what changed and names any new problem the change made, such as a day over its pages or an actor on a new hold.

It leads with the answer, in one or two sentences, then the detail. For a decision it gives two or three options, each with what it costs and what it protects, then its pick. You choose, and it makes exactly that change.

Story comes into its reasoning. It can read each scene's brief and words, so a cheaper option can be weighed against what the scene must do. The production desk's Gaffer has no tool that writes the script or draws a frame, and the creative calls stay with the people who make the film.

A run is one question and everything Gaffer does to answer it, up to 60 rounds with the model. A round can call more than one tool, so this is not a count of what the Steps card shows. Near the end it is told to wrap up and say what is left. A busy or failed model call is tried again twice before it says so, and the Stop button ends the run.

The conversation is saved with the project, so it is there on your next visit. As a chat grows, old tool readings are shortened and Gaffer reads again when it needs them, so treat a number in an old message as old.

Question 1: are we ready for Day 1?

This is the question typed into the first screen on this page, and one of the four suggestions asks it about the next shoot day. For this question the brief says to read get_production for the risks and get_day for the list of what is not ready. get_production returns the film-level risks and counts, and get_day returns one shoot day as its call sheet.

The table is every check the day read makes, in the order the code runs them, with what it returned for Day 1 of Midnight Diner. The warnings in the last row are the schedule's own: more than one set, day and night mixed, or a day over its pages target.

The check does not list facilities (power, toilets, parking and the rest) as not ready, though the day read carries them for each location. It does not check turnaround, weather or permits unless a permit is a to-do.

The film-level risk list flags a location as needed this week only for days in the next 7 days. On 12 October, with every day already past, it lists one risk, "20 P1 to-dos are past due", owned by department heads. The day read does not depend on that window, which is why it still finds the unlocked Diner.

One real answer, checked against the project

We asked Gaffer that question on the real Midnight Diner project and saved the answer: "Are we ready for Day 1? Rank what is not ready and who owns each fix." It answered with nine ranked items. The table checks the main claims in the part of the answer the screen shows, against the project data. It is not every phrase.

The top of the answer is scrolled out of the capture, so we checked only what is visible. We captured it from a development copy of the app and did not record which model route produced it, so the live server may answer differently. Another run would word it differently.

Every date, count and amount matched. The assignments did not come from the project. The brief tells Gaffer to name the owner of each fix, and also to say a fact is not set when the desk does not hold it. With no owners recorded, it named people from the crew list instead of saying none were set.

It also ranked and grouped. Six of the seven priority 1 to-dos on those scenes are named in the list, and the seventh (call sheet for day 2 out by 18:00 the day before, due 9 October) does not appear in the nine items and sits only inside the count of 27. That is a reasonable summary, and it is also why you read the count against the page.

The fix is in your hands. Set an owner on each to-do, which the day read and the scene read both carry, and ask who owns a fix after that. Until then, read every name in an answer as a suggestion.

Question 2: what moves if a scene moves?

Lou works on Day 1 and Day 4 and waits through Days 2 and 3, two hold days. Whether a hold is paid depends on his deal, and the app's planned cost has no hold pay in it. His second scene is scene 5, the warehouse. The brief tells Gaffer to read the schedule and the day out of days, lay out options with their cost in days, holds, pages and money, and make a move only on your pick.

We ran that move through its tools on the sample, the same change as the right-click menu below makes. We did not capture Gaffer's written answer to it, so the table is what the tools return, not its wording.

The planned cost of the four days does not change, because the app's planned cost does not include hold pay. About $1,086 of cost moves from Day 4 to Day 2, and Lou's two holds go away, which saves whatever his deal pays for a hold. The price is a day with two sets and both day and night. That is a company move and a change of light, and the real decision is which problem you would rather pay for.

The same trade is worked in full in the day out of days guide and in how to make a shooting schedule. After a move, the tool returns day lines for the days it touched, with pages and warnings, and the brief tells Gaffer to read back the costs and the holds too.

Question 3: what does a day cost?

Every shoot day carries a planned cost, worked out by the app from the people on it and the items on its scenes. Gaffer reads it from the schedule and from the day read. On Midnight Diner, Day 1 is $26,247, the dearest, and the four shoot days together are $94,159.

The what_if tool prices a change without making it. Two examples on the sample, with the contingency of 10% the budget already carries:

The extra day is an average day, not the next one on the schedule, so it is a first estimate and not the price of a particular night. Both numbers come from the app's engine and not from the model's arithmetic, as long as Gaffer calls the tool, which its brief tells it to do before a big change.

For the arithmetic of a single overtime hour on one day, see film production scheduling. The budget side of the same questions is in how to budget a short film.

Question 4: what is not priced yet?

A budget is only as good as what is priced in it. In Spring Pro, breakdown items in the budget's categories and location fees that have no budget line are not priced. The get_budget_lines tool returns them as a to-price list, and save_budget_line prices one from it.

On Midnight Diner that is 66 items in 9 accounts. The tool returns only the first 60, so Gaffer sees 60 of the 66, and the 3 location fees are among the 6 it does not see.

None of the 66 was added to the script after the lock on 4 October, by the same test the Budget page uses to flag new script items. They are work not done yet, not scope that crept in. The change log is empty too, so no budget line or deal has been changed since the lock.

None of the 60 Gaffer sees comes with a suggested rate, because no budget line with the same description has been priced before.

The tool takes an account filter, so a question about one account at a time sees all of it, location fees included. A question about the whole budget sees the first 60.

What it can change, when you ask

Twenty-three tools change the production. Each makes the same write as the page it belongs to: Schedule, Call sheets, Crew, Locations, the to-dos on a scene and the Budget. A change from Gaffer is a normal edit saved the normal way, not a separate copy.

After the lock, a change to a budget line or to a person's deal is logged with who, what and how much. Changes to the budget's rules, funding and expenses are not in that log. Gaffer's brief tells it to act only when asked. That is an instruction to the model, and the rules table below says which rules the tools hold and which only the brief does.

One more tool, remember, saves a standing preference for the project, such as "always call cast an hour before crew". It is added to every question afterwards, and the step shows in the chat, so nothing is learned silently.

What it cannot change is as important as what it can. There is no tool that writes the script's lines, sends anything, or pays anyone, and the next sections show how each rule is held.

The one undo per run

If a run changed anything, the app keeps the project as it was before the run's first change. Say "undo that" and Gaffer calls undo_last_changes, which puts back what it changed. The brief tells Gaffer to end every change with that offer.

We ran it on the sample with the real tools: move scene 5 to Day 2, then ask for the undo. This is what we saw.

• The schedule came back exactly. Day 2 held scenes 2 and 7 again and Day 4 held 5 and 4.
• Another person's edit stayed. We typed an address for the Diner between the move and the undo, and the undo left it alone.
• An edit to the same field did not. When Gaffer set a call time and someone then changed that same call, the undo put the field back to what it was before the run, and their edit was lost. Look at the day before you undo.
• A read-only question did not use the undo up. A change, then a question that only read, then "undo that" still reverted the change.
• Only one run goes back. After an undo there is nothing left to undo, and the tool answers "Nothing of mine to undo."
• The lock can be undone. Locking the budget replaces the lock and clears the change log, and "undo that" restores the earlier lock.

Two limits from the code. The undo is held in the open production desk, not saved with the project, so a reload, a closed tab or leaving the desk to open a scene or the script loses it, although the chat itself is saved. We read that in the code and did not test it in the running app. And a stopped run still counts as a run, so what it had changed before you pressed Stop can be undone.

What it never does, and how each rule is held

Some rules are built into what the tools can do. Others are instructions to the model, and an instruction can be broken. The table says which is which. The tool rules were checked by running the app's own tool code on the sample.

Say lock only when you mean it. Locking an already locked budget replaces the lock and clears the change log, and nothing but the brief stops Gaffer from calling it.

Where it goes wrong

NIST's generative AI profile names one of its risks confabulation: confidently stated content that is wrong, which it says is a natural result of how these models are built. It also warns that people may lean on such systems too much, which it calls automation bias. Spring Pro's own screen carries the same warning: "Gaffer is AI and can make mistakes. Check important numbers."

Here is what we found in the code and on the sample, besides the owners with no record behind them.

• A capped list. get_budget_lines returns 60 unpriced items, and the sample has 66.
• An average day. what_if prices an extra day at the average planned day, not at the day you have in mind.
• Shot by date. The overview counts a day as shot once its date has passed, even when no scene is marked shot. On 12 October it reads all four days that hold scenes as shot and finds no next day. Keep the start date true when a shoot slips.
• Turnaround and meals from a rule of thumb. Its brief gives about 12 hours from wrap to the next call, never under 10, and a meal within 6 hours of call. That is not your agreement, and the app does not check turnaround. See turnaround and meal breaks.
• Old readings shorten in long chats. As a chat grows, older tool results are shortened and only the newest turns are kept, so Gaffer reads again when it needs a number. A run stops after 60 rounds with the model and keeps what it had changed so far.
• Everyone on the project can ask. Spring Pro has no per-person permissions yet, so treat Gaffer as able to read what anyone on the project can read, including deals, rates and contact details.
• Your data goes to a model. Each step sends the conversation and what the tools read, which can include names, deals, phone numbers, the script's text, to-dos and the budget, to an AI model through Spring Pro's server. The server function takes signed-in people only and keeps the model key off your browser. A development copy of the app can reach a model by another route. Read the provider's data terms and your own agreements before you use it on a confidential script.

AI, the crew and the agreements

Guild agreements have started to say who must do the work. The Directors Guild's June 2026 release on its new four-year agreement lists its 2023 AI safeguards as renewed, including that its members' work "must continue to be performed by a person". Members ratified it later that month.

The guild's December 2023 release on a tentative agreement for commercials is more specific. It says all duties of directors, assistant directors and unit production managers must be done by a person, and that generative AI does not count as one. It calls these the same protections the guild secured in its Basic Agreement earlier that year.

That agreement is for commercials. It still needed a members' vote when released, and we did not confirm the vote. It gives the term as ending on 30 November 2026, so look for what replaces it.

IATSE's memorandum of 3 July 2024 adds an article on artificial intelligence. A producer may require employees to use AI systems, and those employees stay covered by the agreement while they do. The producer will not require an employee's prompts in a way that displaces a covered employee, may require consent before an employee uses AI, and may reject its output. We read the memorandum, not the ratified text, so check the final agreement.

None of this is a ruling on a tool like Gaffer, and we are not lawyers. It shows where the line sits: the assistant reads and drafts, and a named person decides and sends. If your production is under an agreement, read it and any policy your producer has issued on AI use, including privacy and security, before you use any AI tool on crew data.

How people get these answers today, and what other AI helpers say about themselves

Here are four ways people get these answers, plus Spring Pro's. This is our view. The two vendor rows come from the vendors' own pages, read on 12 October 2026, and we did not test either product.

A preview and approval step is slower and safer on a large change. Act then undo is faster and shows its work afterwards. Which suits you depends on what a wrong change costs and on whether anyone else edits the same fields while you work.

The Filmustage page and the FinalBit post are marketing, so treat their claims as claims. For a comparison of one scheduling suite with Spring Pro, see the StudioBinder alternative.

From a short to a feature: what changes

The questions do not change with size. What grows is how many people ask them and what a wrong answer costs. These sizes are rough labels, not industry definitions.

Doing it in Spring Pro

Gaffer is a button in the production desk's sidebar, just below Setup. Open it, pick one of the four suggestions or type your own question, and press Enter. While it works you see a label that reads Thinking or names the tool it is running, and a Steps card that counts the tools as they run. The reply streams in as it is written.

The four suggestions are "How are we doing this week?", "What is most at risk right now?", "Are we ready for the next shoot day?" and "Where are we over budget and why?". There is one conversation per project, saved, so it is there next time. The pages it reads are described in shooting schedule software, call sheet software, production budget software and script breakdown software.

What it does not do yet: it cannot send, pay or edit your script, it has no turnaround, weather or permit checks of its own, it reads one unit and one script per project, it does not see storyboard pictures or receipts, and it has no preview step before a change. The undo lasts only while the production desk stays open.

The checklist

Copy this. Tick it before you rely on an answer or let it change the plan.

• Ask in one sentence, and name the day or the scene by number
• Set an owner on every to-do before you ask who owns a fix
• Keep the schedule's start date true, because a day counts as shot once its date passes
• Ask for options before a large change, then pick one
• Check each number it quotes against the page it came from: Schedule, Call sheets, Budget
• Check any turnaround, meal or overtime rule against your own agreement
• After a change, read the days it touched, and say "undo that" before you reload or leave the desk if it is wrong
• Read anything it drafts before you send it, and keep a copy of what you sent
• Read the provider's data terms and your agreement's AI terms before you use it on crew data

How these numbers were made

Every number about Midnight Diner was computed, not typed: from the sample project's own data and the app's cost engine, and by running Gaffer's own tools on the sample for the readings, the move and the undo, on copies held in memory. The desk those tools read is rebuilt in the page the way the production desk builds it. It leaves out storyboard frames, scene running times and the money tracker, and no number here uses them. The build stops if the code or the sample changes so that a sentence here is no longer true.

Midnight Diner is a demonstration production, not a real shoot. We captured one real Gaffer answer on 12 October 2026 from a development copy of the app and checked it by hand, and another run would word it differently. The descriptions of agreements and of other tools come from the sources below, read on 12 October 2026. The IATSE memorandum was read as text, the guild items are union releases and not the agreements themselves, and agreements change, so check yours.

Questions people ask

Sources

Every outside claim that names a source comes from one of these, read on the date shown. Standard conventions, such as the eighths rule and the usual specialist cases, are common practice. Figures from agreements and summaries change at each negotiation, so check the current text before you rely on a number.

• NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative AI Profile (July 2024, a US government standards body). Checked 2026-10-12.
• Directors Guild of America, board approves the tentative new agreement (12 June 2026, a union release). Checked 2026-10-12.
• Directors Guild of America, membership ratifies the new contract (26 June 2026, a union release). Checked 2026-10-12.
• Directors Guild of America, tentative agreement on a new commercials contract with the AICP, sent to members for ratification (8 December 2023, a union release). Checked 2026-10-12.
• IATSE Basic Agreement, memorandum of agreement of 3 July 2024 (a copy held by the Berkeley Labor Center). Checked 2026-10-12.
• Filmustage, AI Agent for Film and TV Pre-Production (a vendor page). Checked 2026-10-12.
• FinalBit, AI and Agentic Scheduling: The Future of Feature Film Production (a vendor blog post, first published 26 August 2025 and marked updated 6 July 2026). Checked 2026-10-12.
• Film Independent, a recap by a staff writer of a talk by line producer Angela Sostre on building a better budget (16 August 2017). Checked 2026-10-12.

Keep reading

• Film production scheduling: the whole picture.
• Scenes become strips. Strips become shoot days.
• How to make a shooting schedule for a film, with a real sample.
• How to make a film call sheet, with a real filled-in example.