How do I add our own policies?

Upload your AI policy or start from the Flow template, tick the commitments the assessment should ask about, and press one button. In Know Your AI, ticked rules on approved tools and human review become everyday situations, you can accept a situation for your other rules, and Flow writes questions for the rest.

Only a company admin can do this. A manager can read results but cannot open Policy.

If your company has written a way of working, an AI policy, or a playbook, you can hand those documents to Flow. Flow reads them and lists the commitments it found, meaning the rules your policy actually makes. You tick the rules the assessment should ask about, then press one button. It saves the list and writes the questions for the rules you ticked. In Know Your AI, a ticked rule on approved tools or human review becomes an everyday situation that asks people what they should do, and gets no written question. For a rule on another topic, you can accept a situation Flow drafts. Until you do, Flow writes a question for it.

If you have no policy document, use the Flow template. It is five commitments drawn from the EU AI Act’s Article 4 duty. Take it, untick anything that is not true of your company, and save. Nothing you upload or take shapes anything until you save it.

The round’s questions stay shut until you save. They are written from the rules you tick, so there is nothing on that page until those rules are stored. Open it before you save and it says so, with a button back to Policy.

Policy has two tabs, numbered in the order the work happens: 1. Documents and 2. Commitments. Commitments are read out of documents, so documents come first. The numbers are there on every visit. They name the order, not a set of steps to finish, and you can work in any order you like.

Each tab says in one line what the assessment does with it, and carries a count: how many documents you have added, and how many rules are ticked out of how many there are. The count is the table of commitments only. The two always-asked rules described below are not in it, and the assessment asks about them on top of whatever you tick.

The tab you are on is part of the address, so you can send a colleague a link that opens the one you mean. With no documents yet the page opens on Documents; once you have one it opens on Commitments.

In this guide:

Step 1: Add a document, or take the Flow template
Step 2: Check the commitments
Step 3: Tick the rules to ask about
Letting Flow suggest which rules to ask about
Step 4: Save, and get your questions
Where the saved list is used
What we remove before reading
Limits

Step 1: Add a document, or take the Flow template

Open Policy on the rail. With no documents yet, the page opens on 1. Documents, where the large drop box says Drop your AI policy here. The heading names the tab you are on.
Drag your files onto the box, or press Choose files. Flow reads PDF, Word, PowerPoint, and plain text or Markdown files. You can pick several at once.
If your policy is in another format, save it as a PDF or a Word file first, then add that.

While you have no documents, a line under the drop box reads No policy yet? Press Start from the Flow template and its five commitments land on your Commitments tab, all ticked. Untick the ones that are not true of your company, and edit any wording you want to change. From there they behave like any other commitment.

The line shows only while you have no documents. Once you upload one, your own rules are the list.

You can also start from the Commitments tab while the list is empty: it offers Add a document, and Write one yourself for a rule you would rather type.

A read takes up to a minute, and Flow reads one file at a time however many you pick. Every file you picked appears at once. The one being read has a moving bar under its name and says what is happening to it: Reading the file, then Removing names and emails, then Finding the rules in it, then Checking every quote is really in it. The files behind it say Waiting its turn with their place in the line, and the line gets shorter as each one lands.

If one file cannot be read, it says so and keeps its Try that file again. The rest of the queue carries on without it.

Under the drop box, Flow says how much room is left before you pick anything, so you can see there are two slots free rather than finding out afterwards. If you pick more files than will fit, or one over 10 MB, Flow names the files it did not add and adds the rest.

Once every file has been read, Flow moves you to the Commitments tab, highlights the new rows for a moment, and says how many commitments came out of which document. It waits for the whole queue: it will not carry you off the list of files while some of them are still going. If any file failed, you stay on Documents instead, with the failed one in front of you and its Try that file again beside it. If the title box is empty, the file’s name becomes the title.

One read gives two results, so the document’s line on the Documents tab reports both: “14 commitments · 4 of 5 ground rules”. The five are your escalation route and the four kinds of material, which together make up the Always asked card at the top of your commitments. A file that decides none of them says nothing new for the ground rules, which is worth knowing when you add a second policy: those answers are one set for the whole company.

If the read finishes but finds nothing, the document is still stored and its line reads Not read yet.

Each document on the Documents tab has a ⋯ menu with two items. Read it again re-reads the same document, which is what to press if the first read came back with nothing. Remove deletes it, and asks you to confirm first: the confirm names the document and how many commitments on your list came from it. A rule a second document also states stays on the list.

Your tools’ use policies

A tool whose use policy is In force on Tools shows up here too, as a document named after the tool, such as “Cursor use policy”. It is on the list as soon as the policy is in force. Flow reads it the same way as a file you upload, which takes about a minute, and its rules wait on Commitments for you to approve them. It is not read for your ground rules, because it is about one tool. It does not count toward your five uploads. Flow reads up to 20 tools’ use policies. If more are in force, the Documents tab says how many were left out. When one of the 20 leaves, a waiting one is added the next time you save any tool. Its ⋯ menu has Read it again only. To change or remove it, use Edit it on the tool page under its name: change the words there and Flow reads it again, or put the policy back to Being written and it leaves this list.

The line under each document name counts the commitments on your list that came from it. A rule that two documents both state is counted under both, so those numbers can add up to more than the total in the heading. When that happens, a line under the heading says how many rules were found in more than one document.

Step 2: Check the commitments

A commitment is something the document holds people to: a rule, a data tier and what it permits, an approval step, an escalation route, or something a person is expected to own. Background and ambition are left out, because nobody can be held to them.

Two rules are always asked. They sit in a card above the table, Always asked:

Where a wrong answer goes. One line: a person or a channel. Write who to tell when AI gets something wrong.
What may go into a tool. Four kinds of material, each marked Allowed or Not allowed: customer contracts, personal data about other people, your own notes and drafts, and material that is already public.

They have no tick box, because every assessment asks about both. In Know Your AI they are the data and escalation situations (see Step 3). If a document you added decided one of them, the answer is already filled in, with the sentence it came from underneath. If you took the Flow template, its defaults are in the boxes and each row reads Template default. Change anything that is wrong. Your answer always wins over what a document said.

You can leave them unfinished. The card says which kinds of material you have not ruled on yet, and saving still stores your commitments. The assessment then asks about the parts you did answer and leaves the rest alone. In Know Your AI, the data situation needs your answer on personal data about other people, and the escalation situation needs the route.

The Commitments tab is a table with one row per commitment, merged across every document you added. Each row is numbered, and carries its Asked box, the rule, and the documents it came from. The number is the row’s place on screen, so a filtered list still starts at 1.

The Source column lists every document that states the rule, one title per line. Press a title and that document’s own sentences open in a small panel over the table, with the document’s name above them. Press outside it to close it. Where one document says the same thing in two places, the title carries a count, “2 quotes”. The quotes let you check the rule against your document without opening the file. They are copied word for word, and any rule whose quote is not really in the document is thrown away rather than shown to you.

This list is yours to fix:

A saved rule cannot be edited. Its wording is what your document says, and your people are asked about exactly that. If one is wrong, delete it and add your own.
Delete, in the row’s ⋯ menu, removes a line you do not want. It is the only thing in that menu.
Add a commitment, at the foot of the table, gives you an empty row to write a commitment the read missed. A rule you type yourself needs no quote, and its source reads Yours.

On a long list, use the box above the table to filter by wording, and the three buttons beside it to narrow it further (All, Asked, Not asked). To shortlist one kind of rule, press the Topic heading on the table’s own Topic column and pick one. The three work together, and the counts on the buttons always describe what the table is showing. Filtering only changes what is on screen. It stores nothing and removes nothing.

Every commitment has a topic, found by the same read that found the rule: Approved tools, Data tiers, Escalation, Human review, Retention, or Ownership. A rule that fits none of them reads Other, and so does every rule you write yourself. The six are the same for every company, so a rule under Data tiers means the same thing here as anywhere else.

Your documents are one tab away, on Documents.

Step 3: Tick the rules to ask about

Each row has one box, under Asked. A rule read out of a document starts clear; the Flow template’s five start ticked, because they are what the duty asks of everyone. The box is the whole state: a ticked row is asked about, a clear one is not.

In Know Your AI, some rules become situations. Know Your AI asks a few everyday situations, one per topic, and each one ends What should you do? One option is your rule, written as something to do. The others are mistakes people really make, plus I’m not sure, which is always last. Each person gets the options in their own order, and which mistakes are offered can change from one round to the next. Once someone answers, they see your rule and where it comes from, for example “From AI policy”. It never says right or wrong. The first answer is the one that counts: after someone has seen your rule, they cannot go back and change it.

Approved tools and Human review: the situation asks about the first rule you ticked on that topic. With nothing ticked on the topic, there is no situation.
Data and Escalation: these come from your Always asked answers, and people see that answer after they answer the situation. The data situation is asked once you have said whether personal data about other people may go into a tool; the one about an email with a price in it also needs your answer on customer contracts. The escalation situation is asked once you have written the route. A ticked rule on data or escalation gets a written question of its own, because it may say more than your Always asked answer does.
Retention, Ownership and Other: there is no ready-made situation, so you key one yourself. On the round’s Questions page, under Situations for your other rules, press Draft a situation and Flow drafts one from the rule. Read it, fix anything that is wrong, and press Accept. You can also Write it yourself, or mark the rule Not askable. Nothing is asked until someone at your company accepts it, and until then the rule keeps its written question. A round asks up to three of these, in the order of your list. An accepted situation stays with the rule and is asked in every later round. If you reword the rule, its situation is not asked again until you accept one for the new wording. The Flow template’s rule on keeping chats comes with a situation Velory wrote. Edit it, or mark it Not askable if it does not fit.
A second ticked rule on approved tools or human review is listed on the round’s Questions page under Not asked this round.

Each answer is stored as one of three things, together with the rule it was about: the person picked your rule, they picked another option, or they were not sure. For approved tools, human review and a situation you accepted, the rule is the one you ticked.

Once a rule has an accepted situation, its written question is not asked, so nobody answers the same rule twice. The question’s row on the Questions page says Asked as a situation this round.

Check each situation before you send. A situation is an everyday case with an answer we wrote from your rule, and your rule may mean something else. On the round’s Questions page, open a situation to see the answer that follows your rule and the rule people see after they answer. If the answer is not what your rule asks, tick Doesn’t match our rule. That situation is then not asked in this round. For approved tools and human review, Flow writes a question for the rule you ticked instead, so the rule is still asked. A data or escalation situation is listed under Not asked this round. The box can be changed until you press Send. The next round starts with every situation asked again, because its situations are different ones.

A round on the earlier version of Know Your AI has no situations. It asks the data and escalation questions from your Always asked answers, and no question about a ticked rule.

To tick several at once, narrow the table first (by topic, by wording, or with the Asked buttons) and press the action beside those controls. It names what it will do and to how many rows, for example Tick all 6 shown, and it acts on exactly the rows in front of you. Nothing you have filtered away moves. Once every shown row is ticked the same action reads Untick all 6 shown.

What you tick is also what Flow writes questions from. Tick three rules and you get up to three questions. A rule nobody could answer about from their own work gets none, and Flow names the ones it left out. In Know Your AI, a ticked rule on approved tools or human review gets no question, because its situation already asks about it, and neither does a rule with a situation you accepted. Rules on every other topic still get one, and so does the rule behind a situation you marked as not matching. Tick nothing and nothing is written: the button at the top just says Save.

Leave a box clear and Flow writes no question from that rule. If you tick nothing, the assessment and the write-up are exactly what they were before you uploaded anything.

Untick a rule you had already ticked and its question is switched off. It stops being asked and keeps the wording you gave it, so ticking the rule again brings it back as you left it. Switched-off questions are not listed with the round’s questions. A line where they would be says how many there are.

Ticking or unticking counts as an edit, so the line under the heading says Edited since you saved. Press the button at the top to store it.

Letting Flow suggest which rules to ask about

A long policy leaves you guessing which of its rules matter most for your people. Flow can rank them for you. Press Suggest which to ask, the star button in the row of controls above the table.

It ranks your rules against what Flow already knows about your company: the description on your Company page, the roles your job titles map to and how much of your workforce holds each one, your teams and their sizes, and how often your own documents repeat a rule. A rule your documents state in three places is a rule you already lean on, and the ranking counts that.

Flow ticks the top five and shows them in a panel above the table, each with one or two sentences saying why it matters here. The rule is shown in the same words as the row, so you can find it in the table by reading it. Ticked rows carry a star next to the rule.

Nothing is saved. The boxes are ticked on screen only. Untick any you disagree with, tick anything it missed, and press Save when the list is right. It takes a few seconds to run.

Two buttons sit under the picks:

Suggest again runs the ranking again and asks how many rules you want first. Type any number from 1 up to the number of rules you have. Stars can move to other rules on a second run. Boxes you ticked or unticked yourself stay exactly as you left them, and so do rules that were already ticked when you last saved.
Clear suggestions removes the stars and the panel, and clears the boxes the suggestion itself ticked. Your own ticks stay.

The suggestion lasts for as long as you have the page open. Reload the page and the panel is gone; the ticks you saved are what remain.

When the button cannot run. Flow ranks rules against your people’s roles, so it needs them. It will not guess from your company name. The sentence next to the button says which of three things is missing:

Nobody has a job title. Add people with their job titles on People.
No job title has a role yet. Flow matches job titles to roles after each change to People, usually within a minute. Reload the page to check.
The job titles are waiting on Velory. Flow could not match them by itself, so Velory matches them for you. See what a title waiting on Velory means.

The button is also absent while you have fewer than three commitments, because there is nothing to rank.

The reasons name roles and teams, never a person, and never a headcount. Nothing about any individual is sent when the ranking runs: what goes out is your rule wording, your company description, your role names with a percentage each, and your team names with their sizes, with any team under five counted simply as small.

Step 4: Save, and get your questions

One button at the top of the screen does both halves. It reads Save and create the assessment when a rule you ticked has no question yet, and Save when there is nothing to write.

Press it. Flow stores the list and moves you to the round’s Questions page straight away, before the questions exist. A banner over the list says how many are coming, and there is an empty row for each one, with the commitment it will check. Writing them takes up to a minute.

When they land, the banner says how many were added and the rows fill in. They are ordinary questions of yours, to edit, reorder or delete. Each one carries the commitment it checks under its title. The banner also says they are not live until the round goes out, which is true of any edit here.

If a rule you ticked got no question, the banner lists it under the count. That happens when nobody could answer about that rule from their own work, for example who sits on a review board or how often it meets. You have two moves: reword the rule on Policy so it asks about something a person does, then save again, or write the question yourself with Add a question. The list is there while the banner is, so read it before you leave the page.

The writing carries on if you close the tab. Come back to the round’s questions later and the questions are there, or the banner is still counting. If the writing fails, the banner says so and offers Try again. Your policy is saved either way. Only the questions failed.

Come back and press it again and you will not get the same questions twice. A rule that already has a question is skipped, so wording you edited is never overwritten. Tick two more rules, press again, and you get two more questions.

Nothing you upload reaches a question or a write-up until you press it, which is the point: a person decides what counts as company policy, not a model.

Under the heading, Flow says when the list last changed. Before your first save it reads Not saved yet. After one it gives the day, and it says Edited since you saved whenever what is on screen is not what is stored. Save again to store the change.

The saved list has a length limit. If yours is over it, nothing is saved and Flow tells you how many of your rules fit. Trim the list and save again. Nothing is stored halfway, so you never end up having signed off on rules that were quietly dropped.

Saving with nothing ticked is allowed, and it counts as having done this step. It writes no questions, and the assessment and the write-up work the way they did before you added anything.

Once a round is out, changes go in the next round

Sending a round freezes its questions and the ticked rules it asks about. People are answering them, and rewording a question under an answer would change what that answer means. So after you press Send:

Questions cannot be edited for that round. Saving, publishing and deleting are all refused.
The situations keep the ticked rules you sent them with. Which rules are ticked is read from the list the round went out with, so ticking a rule after Send adds no situation for the people still answering. The situations for your other rules are kept the same way: a situation you accept, edit or mark Not askable after Send applies from the next round, and the people still answering meet the ones the round went out with. Your Always asked answers are read when each person starts, so a change to those reaches the people who start after it.
A Policy save still stores your list. Ticking or unticking a rule is saved the moment you press the button.
New questions go into your next round. If you have started one, they are written into it and the round that is out carries on unchanged. If you have not started one, nothing is written yet. Start the next round and the questions come with it, as below.

Flow says when a ticked rule has no question. A banner appears in the place where the gap matters:

On the round’s Questions page, before you send it. It reads “3 commitments ticked on Policy · not asked in this round yet”, names the rules under that, and gives you two buttons: Policy, to go and check the list, and Draft 3 questions, which writes them into this round. It behaves like a save on Policy: the same banner, the same empty rows filling in.
On the round’s Questions page, after you send it. The same count, worded for a frozen round: “3 commitments ticked since this round went out · they go in Round 3”. The round it names is the one you have open and have not sent, whichever that is, so opening an old round’s questions tells you where its leftovers went rather than naming a round that has closed. If every round you have is sent, it says no round is open for them yet: start the next one. There is no button either way, because nothing can be added to a round people are answering.
On Rounds, over a round you have not sent yet. When you saved your policy after the last round went out, the card carries one line: when the policy changed, how many new commitments have been drafted into this round, how many ticked rules still have no question, that the round before it keeps the questions it was sent with, and that nothing changes for anyone until you send. Review the 2 questions opens what has landed, and Draft 3 questions writes what has not. A round in both states offers both buttons.

Starting the next round brings your new rules with it. The new round is the last one copied question for question, so a rule you ticked after that round was written has no question in the copy either. Start Round 3 carries a box, ticked for you, reading Draft 2 new questions from your policy. Leave it ticked and the questions are writing by the time the new round’s Questions page opens. Untick it and the new round starts with the copied questions only. The Draft button on its Questions page is still there when you want them.

However you get them, a rule never gets a second question, and nothing reaches anyone until you press Send.

Each round keeps the policy it asked about. A round records your saved list, and the date you saved it, at the moment you sent it. Change your policy afterwards and the round that is already out carries on asking the old one.

Rounds shows you which. The current round’s card reads “Asked against the policy saved 3 Sep” under its question count, each past round carries the same date in its line, and the round’s own Questions page says it above the list. A round sent before you had written a policy down says so instead. So a report you read months later is not describing a policy written after the round finished.

Where the saved list is used

Once a list is saved it is used in the places below, and only the rules you ticked are used. The documents themselves never reach the assessment.

The situations in Know Your AI. A ticked rule on approved tools or human review becomes the situation for its topic, as Step 3 describes. After they answer, the person sees the sentence in your document that the rule came from, with the document’s name. A rule from a tool’s use policy names the tool instead. A rule you wrote yourself, or one from the Flow template, shows its own wording and your company’s name.
Writing questions. Saving writes a question for each ticked rule that has none yet, apart from any rule nobody could answer about from their own work. In Know Your AI, rules on approved tools and human review are left out, as Step 3 says. Each question keeps the rule it checks: the list says Checks: and the rule under the question’s title, and the editor says it in full, in your own words. The training question has no rule to name, so its Checks: line says “Training you have had”. A question you typed yourself says nothing there. Your people never see that line.
The wording of the four skills questions. Flow writes their options in the words of each person’s job, and the rules you ticked are part of what it reads so the options fit how your company works. They never decide which option is right.

Two things the list does not do:

It does not change anyone’s AI level. The AI check grades the same things for everyone, on what the person said about their own work. Whether they follow a company rule is not one of them. Someone who says a rule is ignored here has answered the question, and the answer is what they are graded on.
It does not carry anyone’s words to anyone else. Nothing an employee typed is read back to another employee, and no admin or manager screen carries a person’s words. See what does my manager see?.

A company that saved a list and ticked nothing on it has the assessment and the write-up it had before, unchanged.

Questions written from your rules go under Optional, below the required questions every Know Your AI asks. Edit or remove them like any question of your own. When Flow writes them it is shown the required questions and told not to write one that asks the same thing again.

What we remove before reading

Your policies are company documents. They still mention people, though, so before a document is stored, and before any of it reaches a model:

Email addresses become [email].
Phone numbers become [phone].
Names of people on your Flow roster become [a colleague].

This runs on the title as well as the text, so a file named after a colleague does not keep their name. We store only the cleaned text, never the original file. Nothing about your roster is sent with the document.

If a rule reads oddly because a name was taken out, delete it and add your own version. You can write “Ask your manager first” in place of a name.

Limits

Five documents per company that you upload. Remove one before you add another. Your tools’ use policies do not count toward the five.
Twenty tools’ use policies. Past that, the Documents tab says how many are not read.
10 MB per file.
About ten pages of text per document. A longer document is read in part and says so, so split it across documents or cut it down to the sections you care about.
Twenty-five commitments per document, and a limit on the saved list as a whole. If yours is over it, nothing is saved and Flow tells you how many of your rules fit.
Twenty-five questions written in one press. Tick more than that and the rest are written the next time you press, so press again to pick up what is left.

The cleaned text stays until you remove the document or the company account closes. See what happens to your data.