A practical guide to using AI while keeping control of your knowledge

Why I’m coming back to this

An organization can publish more than ever and still struggle to be understood.

AI helps us write descriptions, newsletters, and campaigns. But it also describes us. Being found matters. AI raises the stakes because it can put together its own account of our work.

In a survey published in June 2026, Pew Research Center found that six in ten US adults said they’d read AI summaries at the top of search results. Someone may read an AI’s account of your work before they ever visit your website. Pew Research Center

A play and a community workshop might both explore belonging. An AI could get their dates and locations right while missing why the organization brought them together.

Being included in an answer doesn’t mean being understood.

At Culture Creates, our work on Artsdata keeps bringing me back to what sits behind a record: who supplied it, what it connects to, and what gets lost when it moves.

People know why a programme exists, how a relationship developed, and what they learned along the way. Sharing that knowledge can help an organization become more cohesive: colleagues can understand how their work fits together and build on what others know.

I think part of marketing’s future lies here. How we want to be known needs to grow out of what we actually know, do, and learn. Marketing can help develop and express that understanding.

Reaching people still matters. So does helping them understand what your work means and why it matters.

When we use an AI tool ourselves, we can choose what information to give it and check its answers. When we publish information, other people and AI services may use it. We can make our own explanations available, but we can’t make every outside AI use them or represent them accurately.

I wrote the first Common Sense AI Playbook eleven months ago. It offered one sequence, starting with getting your information straight. This version makes an important distinction: the preparation needs to fit the job. Translating a document and working across an organization’s knowledge need different foundations.

The aim is to use AI while keeping the knowledge, judgment, and choices that make the work yours. The examples here come from cultural organizations, but the questions apply to any organization.

Start with three questions

What would you like to make possible?

Start with something you can point to.

Maybe you want to prepare an event listing more easily. Find something in the archive. Help a colleague understand why a programme was put together the way it was.

Pick one useful starting point. Then work out what information and help you need.

What happens if the AI gets it wrong?

A small task can still matter a great deal.

A translation of a public statement needs careful checking. Trying out ideas for an internal meeting leaves more room to experiment.

Decide who will review the result and what they’ll check. Be clear about where it’s going: an internal conversation, everyday work, or the public.

What can it see, and what can it do?

Set this out before uploading files or connecting a tool.

Is the information public? Internal? Restricted? Are there conditions on using an artist’s contribution, someone’s personal information, or material you have licensed?

Then separate three permissions:

  • Read and help internally. Find information an authorized staff member is allowed to use and help them work with it.
  • Publish or communicate. Answer the public, send a message, or publish a listing.
  • Change something or act. Update a record, make a booking, or start a process.

Permission to do one doesn’t grant permission to do the others. Answering a visitor’s question is already a form of public communication, even if the assistant can’t change a record.

Use a tool your organization has approved for the information and task. If you don’t have an AI policy yet, start with a limited trial that the person responsible has agreed to: one tool, specific information, clear limits, and someone checking the results.

Choose your starting point

Route A: help with one task. You have material you want to translate, summarize, or use in a draft.

Route B: a task you repeat. You want to improve something your team does regularly, such as preparing listings or reports.

Route C: knowledge you want to develop, connect, and keep. You want people to understand how their work fits together, build on what colleagues know, and keep the records and explanations behind it.

These routes can overlap. You don’t have to work through all of them.

If your main concern is what people and outside systems can learn about you, go to “Put your knowledge to work publicly.” That applies even if you aren’t using AI internally.

Route A: get help with one task

You have a 500-word workshop description that’s already approved. You need 100 words for a newsletter.

Give the AI that description. Ask it to shorten it without adding information. Then compare the result with the original.

Did it keep the meaning? Did it drop something important? Did it invent anything?

Give the tool what it needs for the task. Leave unrelated material out.

The review should fit the use. Check a summary against its source. Have a translation checked by someone who understands both languages. Put a public statement through your usual approval process.

Try this on material you know well. Count the time you spend checking and correcting it, too.

You’ll know it’s useful when the result does the job and is worth the work it took to get there. Keep the approved version where your team normally stores that material.

That can be enough. A task like this doesn’t require you to reorganize the whole organization.

Route B: make a repeated task work better

Think about the event listings your team prepares every week.

Each needs a date, location, description, biography, and approved image. Where does each piece come from? Who checks it? What happens when the date changes?

Write down the current steps before adding AI.

Some of the work may be straightforward. Ordinary software can copy an approved date from a schedule. AI might help shorten a description. A person still needs to decide whether that description says what it should.

Work through a few questions:

  • Which steps follow clear rules?
  • Where do you need someone’s judgment?
  • What can happen automatically?
  • Where should the process stop for approval?
  • If something is wrong, where do you fix it?

Some AI tools can work across applications and carry out several steps. These are often called agents.

You can give an agent the information it needs to prepare a draft without giving it the ability to publish, change the schedule, or send anything.

Ask the person setting it up to show you those limits. What can it read? What can it change? How do you stop it? How do you find out what it has done?

Put the limits in the tools themselves, as well as in the instructions you give the AI.

Try a few real cases, including one with a changed date or missing image. Decide in advance what would make the change worthwhile. Count checking, corrections, and upkeep as part of the work.

You’ll know it’s useful when people can follow what happened, deal with an exception, and do the job with a worthwhile improvement in time or quality.

If the recurring problem is that people have to reconstruct knowledge from several departments every time, the next route may help.

Route C: build a second brain your organization can keep

By a second brain, I mean the knowledge your organization keeps connected and usable: its records, explanations, sources, and the relationships between them.

That knowledge can be useful with or without AI. AI can help you explore it and ask questions about it.

Choosing to do this is an organizational decision. What do we want to understand better? Which knowledge matters to our work? What do we want to be known for?

A curator and a colleague in marketing might want to explain why a play connects with an archived interview about belonging. The records exist. The reasons for the connection may only emerge when they talk.

Your people know things your systems don’t.

As that knowledge develops, education colleagues or community partners may contribute other readings. Fundraising, often called development in nonprofits, can bring knowledge from its relationships with supporters and use the shared understanding to explain what support would make possible.

This is how the work can support cohesion. People can see how their contributions fit together while retaining different perspectives and the ability to question how their knowledge is used.

People build the shared understanding. The second brain gives it somewhere to live and develop.

You can begin with notes and records in tools you already use, linked so people can follow the connections. AI can help people explore them. But ask:

If the AI service you use shut down tomorrow, what would your organization still have?

Would you keep the records, connections, explanations, and corrections people had built? Could another tool use them? Or would you be back at the beginning?

1. Start with a question people actually have

For example:

“Which upcoming performances connect to material in our archive about belonging, and why?”

Try answering it with what you already have. Notice what you can’t find, what someone has to explain, and what people keep having to work out again.

What would an answer help you decide or do differently?

Talk to the people who know the material. Ask what a person reading the records would miss. Ask why they see a connection and what it’s based on.

Start small. One programme or a handful of works is enough for a first exercise.

Identify what you might include and who would use it.

2. Agree on access, then put the first collection together

Public event details, internal notes, and private records need different treatment. Decide which people and tools may receive each part before copying it.

Talk with contributors about how their explanations can be used, how they’ll be credited, and how they can correct them or ask for future use to stop. Be clear about who handles those requests.

Let people explain things in their own words. Some knowledge belongs in a conversation or a relationship and shouldn’t be captured.

Keeping the system doesn’t give an organization unrestricted authority over everyone else’s knowledge.

This is part of making data sovereignty practical. The organization needs choices about its technology and information. The people and communities contributing knowledge also need their authority and conditions respected.

Remember that a copy is another place to update and protect. And once something is published openly, you may not be able to retrieve every copy someone else has made.

For a first collection, choose material everyone in the agreed group is allowed to read. A shared folder and a spreadsheet index may be enough. The index can link each work or event to its documents, an explanation, its source, and the person responsible.

Check the sharing settings on both the folder and the index. A link or short explanation can reveal information too. A note in a spreadsheet can record an agreed credit or limit on reuse; it cannot stop someone opening the file.

Begin with selected documents and exports from existing systems. Give someone time and responsibility to look after them, and set a date to see whether the collection is helping.

3. Keep the sources attached

Where did this come from? When was it collected? Who looks after it?

Keep that information with the record.

Different sources may be responsible for different things. The schedule gives you the performance time. A curator explains why the work was chosen. One doesn’t replace the other.

Label exports with the date they were taken. A copy of next month’s schedule can go out of date quickly. An archived interview remains a source from a particular time.

If two records give different performance times, find out which is right and correct the error.

4. Agree on a few shared rules

At some point, a list of documents may become difficult to use. The same artist appears under different names. A work has several performances. Different teams keep rebuilding the same connections.

A small shared model helps here. It’s a few definitions and rules everyone uses when connecting information.

For a cultural organization, you might start with people, organizations, places, events, and creative works.

Keep the rules simple and apply them consistently:

  • A work and a performance of that work are different things. Keeping them separate helps you find both the work’s history and its upcoming dates.
  • A recording can be linked to the performance it documents.
  • An organization can operate a venue. Each needs its own information.
  • A person can buy a ticket, perform in a show, and make a donation. These are different relationships with the same person.

Before linking personal records, check that they describe the same person and that you’re allowed to create and use that connection for the purpose you have in mind. The fact that a performer is also a donor may itself be private. Ticketing and donation details can stay in their existing systems, but access rules need to cover the connections too.

The shared model gives teams common reference points. They can add the descriptions and relationships they need.

What do we need to agree on to connect, and what must we remain free to describe differently?

Someone needs to look after the definitions and help resolve new cases. AI can suggest matches and connections. Check uncertain ones before treating them as part of what you know.

If you need different access for different people, frequent updates from other systems, or connections you can’t maintain reliably by hand, ask whoever manages your document or content system what it can support. That might be your administrator, provider, or a developer. Ask them to show how sources, relationships, and permissions will be kept. Your existing tools may already do what you need.

5. Keep the voice behind the connection

“Belonging” is a label. It doesn’t tell you why someone connected a work to it.

A curator might connect a play about returning home to belonging because “the characters return to a familiar place, but no longer feel accepted there.” Keep that explanation in a note linked from the play’s entry in your spreadsheet. Include the curator’s name, the source and date, and how the explanation may be used.

Another curator might see exclusion or changing identity in the same work. Keep that reading too, with its explanation.

This is where cultural voice and agency become practical. Artists, curators, and communities get to express what they see and why it matters.

AI can suggest a plausible connection without knowing why you programmed the work or what an artist was trying to express. If we leave those explanations to machines because it’s convenient, their interpretations may become the ones people encounter.

Our contribution has to go further than approving what AI suggests. We need to make our own knowledge and perspectives available.

Keep clear what a source records, what a person interprets, and what AI has suggested. If you keep an AI suggestion, record what it was based on, that it came from AI, and who reviewed it.

When you use an explanation in a campaign or resource, keep it traceable to its author and source, even if it isn’t quoted. Agree on public credit with the contributor and check that it appears. If you’ve connected an AI tool, test whether its answers preserve the source and agreed credit too.

6. Set the limits before the AI answers

If you decide to connect an AI tool to the collection, set its limits before anyone starts using it.

The AI should only receive information the person using it is allowed to see, and that the AI service is allowed to process.

Check what the AI can actually retrieve. Staff access doesn’t automatically mean an AI provider may receive the material.

For a public event assistant, that might mean approved dates, descriptions, and venue details. Private contracts stay out of reach.

Giving it the contracts and saying “don’t reveal these” doesn’t reliably protect the information. Keep material it shouldn’t use outside its reach.

Decide what it can retrieve before it starts answering.

You can connect an assistant to a selected collection without giving it access to ticketing, scheduling, or other operational systems. Publishing, changing records, and sending messages need their own permissions.

Have the person setting it up show you how this works. If people have different access levels, test each one. Check that private information stays unavailable, including information revealed by a connection between records.

7. Try it with someone who didn’t build it

Go back to the question you started with.

Can someone else find the relevant material? Understand why it’s connected? See whose explanation they’re reading and where it came from?

Does it help them make a decision, contribute something, or understand how their work relates to someone else’s?

If you’re using AI, can they tell the difference between a recorded interpretation and a new suggestion from the tool?

Try an outdated record, a question the collection can’t answer, and a request for private material. If you’ve connected an assistant, check that it doesn’t turn an empty search into “nothing exists.”

You have a useful first collection when another person can answer the question, spot what’s missing, and correct a mistake.

You also need to know that the time it takes to keep it useful is time the organization can actually provide.

8. Check that you can take the knowledge with you

Choose a small sample: a record, a connection, and someone’s explanation.

Use your system’s export tools, or ask your administrator, provider, or developer for help. Choose a documented format another tool can read. Then open the sample in that other tool.

Can you still tell who or what the record is about? Follow the connection? Read the explanation? Find its author and source? Understand the definitions and corrections?

Test the access restrictions in the new tool before anyone uses the transferred information.

A downloadable file isn’t enough if the things that made the knowledge useful have disappeared.

Being able to move your knowledge is one part of data sovereignty. Who runs the services, where the information is processed, and the terms of use matter too.

An optional way to guide AI: MCP

Once you have useful knowledge, you may want an AI tool to work with it. You can provide both the information and guidance on how to use it.

The Model Context Protocol, or MCP, offers a standard way to do this with compatible AI tools. Through an MCP connection, you can make selected records available, explain your terms and relationships, and provide ways to search or retrieve information. Your organization supplies the definitions and guidance. MCP guide

For example, you could explain the difference between a work and a performance before an AI searches for upcoming dates. You could also point it to the source of a curator’s explanation.

There are two ways this might help you.

Use an existing connection. Culture Creates developed an MCP interface for Artsdata. A connected AI can look up how Artsdata organizes its information before searching its cultural records. Check whether your approved AI application supports MCP, then use that application’s instructions for adding a service. Your administrator or provider can use the Artsdata MCP technical documentation for setup details.

Make your own knowledge available. Ask your software provider or a developer whether an MCP connection would help with the uses you have in mind. You could let staff search approved information and explore why a curator linked two works, without giving the assistant the ability to publish or change a booking.

Guidance doesn’t guarantee that an AI will interpret everything correctly. The service still needs to enforce what it can access and do, and you need to check its answers. This route serves AI applications that connect to it; it doesn’t make every outside AI consult your knowledge.

MCP is an option when it helps. For a one-off task, supplying selected documents to an approved AI tool may be enough.

Put your knowledge to work publicly

Knowing something inside your organization doesn’t mean people outside it can find or understand it.

Your team may correct a performance time in the internal schedule while an old time remains on a partner’s website. Staff have the right information, but someone searching online may still get the wrong answer.

Choose which parts of that shared understanding you want to make public and who they need to reach.

Start with one activity. Look at its name, dates, location, participants, and description on your website and the partner channels you use.

Do they agree? If the date changes, where do you fix it? How does the correction reach the other places?

Make approved facts available in fields software can read, alongside descriptions and explanations for people. This is often called structured data. Ask your website provider to show you how it’s produced and updated.

Research commissioned by CAPACOA for Artsdata found an association between more complete structured event information and higher page views from Google Search. That’s evidence for paying attention to public event data. It doesn’t tell us whether AI answers preserve a curator’s interpretation. CAPACOA research summary

Shared references also help different services recognize the same artist, venue, or event. Artsdata provides those references. You can use them inside your organization and contribute selected public information while keeping your own descriptions and perspectives.

Think back to the belonging example. The event details could feed your website and a partner’s calendar. The curator’s explanation, with their agreement on use and credit, could appear on a programme page and inform an education resource, marketing campaign, or fundraising appeal.

For marketing, try the programme description with someone unfamiliar with it. Can they explain the connection you’re trying to communicate? Their response can show what came across and what needs more context.

Outside AI systems may still use other sources or miss your explanation. You can maintain a clear account of your work without controlling every account others produce.

You should be able to show where your public information comes from, correct it at the source, and check what your chosen channels display.

Footlight CMS is free to use and can work alongside your existing website. It was developed by Culture Creates as part of the Artsdata public infrastructure.

Footlight helps cultural organizations maintain information about their events and connect it to the people, organizations and places involved. It’s designed for sharing public information through Artsdata for others to find and reuse, so use descriptions and images approved for open sharing.

Start with a few upcoming events. The Footlight user guide shows how it works. Contact Culture Creates if you’d like to get started.

Put these choices in your AI policy

Write down the tools people can use, the information they can share, the actions AI can take, and who checks the results.

Include conditions on contributors’ knowledge and on public communication. Say who handles corrections.

Find out what providers can do with the information they receive, how long they keep it, and what your organization can retain or export. Bring in the people responsible for privacy and contracts when personal or restricted information is involved.

A useful policy helps someone make a decision. Can I do this? With this information? Who needs to be involved?

Keep it readable. Update it as you learn.

Keep what people learn

Set up work with AI so that useful, reviewed knowledge has somewhere to go.

If someone finds an important connection, corrects an error, or develops an explanation worth keeping, make it easy to review and save it with its sources and origin.

For a one-off task, saving the approved result in the usual place may be enough. For recurring knowledge work, decide who reviews contributions, where they belong, and who keeps them current.

Saving every conversation gives you another pile of material to manage. Keep what is useful, with the context that makes it understandable.

Build this into work you’re already doing. When a programme is approved, record the reasoning. When a mistake is corrected, update the source. Allow time for it, and check whether the benefit is worth the effort before expanding.

Use AI where it helps

Once you’ve checked and saved a relationship between records, software can retrieve it. You don’t need to ask AI to reconstruct it every time.

AI can also help you work out a repeatable process, with ordinary software running the parts that follow clear rules. That avoids asking a model to do the same work again and again. If you’re assessing energy savings, look for evidence across the whole process, including setup and ongoing use.

What progress can look like

One task done more reliably. Less time reconstructing context. Colleagues understanding each other’s work. Clearer public information. Knowledge you can keep when a provider changes.

You should be able to explain what AI can see, what it can do, how you check its work, and what your organization keeps.

Start with something you want to make possible. Build the foundation it needs. Keep the knowledge and judgment that make the work yours.