Why building AI for enterprise is so damn hard

Leah MessengerLeah Messenger
17. Nov. 2026

The Optimizely product and engineering team’s tale of powering through the (almost) impossible, and how Virtual Teammates could be the answer to what your enterprise marketing team is looking for.

1) Personal agents serve one person: Enterprise AI has to work for a whole organization, and for the business itself.

2) Three things come first: Before it answers anyone, it needs organizational context, tools with the right permissions, and an identity of its own.

3) Same question, different answer: The right reply depends on who’s asking, so every message gets two checks: one on the way in, one on the way out.

4) Make the check its own step: Not a line in a prompt, because instructions can be bypassed.

 

Our VP of Product, Lauren Hammarstedt, handed her kids’ next birthday party over to an AI. And folks, it’s going pretty well.

The AI in question is Instinct, one of the new personal agents that run your errands for you. You text it or call it, it connects to your email, calendar, and messages, and it gets on with things without you even needing to open an app. Lauren uses it for party-planning, as well as about 100 other things a day.

Now, put that same assistant inside a marketing organization of 200 people. Lauren’s take: it would not work. The overall general take: yep, it still would not work. The reasons behind that say a lot about why building AI for enterprise-level is hard and oh-so-different to building an AI for individual-level.

Along the way, we’ll use our Virtual Teammates as the example of enterprise AI that works in a way a personal agent could never.

Playful birthday photo collage hero illustration.png

One user is easy, but two hundred? That’s a whole other problem

A personal agent has one boss. Add a calendar and it tells that person they’re double booked. It sees what they see, does what they ask, and answers them freely, because nobody else is asking.

A marketing organization on the other hand… ah. Over on this side of the field, we’ve got multiple campaign plans, confidential campaign plans, salary data, performance ratings, and executives who talk privately. Then add memory: an agent that remembers everything about everyone and answers whoever asks. This is all on top of benefiting the individual user too.

AI for an entire marketing organization is a whole other ballgame

Lauren Hammarstedt|VP, Product at Optimizely

What enterprise AI needs before it answers anyone

Before AI answers anything, three things have to be in place to make it worth your while as a company.

1

Organizational context: A personal agent only needs to know one person. Enterprise AI has to know how the whole organization works, including its agents, org-wide skills, context, and guardrails.

2

Tools, with the right permissions: It needs the systems and tools your team uses, and the right level of access to each — our agents and Virtual Teammates have a vast selection of connectors meaning you can work across tools in one platform.

3

An identity of its own: A unique user identity means you can limit what the AI can reach — this means anyone who chats with the AI gets that specific project’s data and nothing else. For example, Virtual Teammates have team permissions and individual owners. In addition to this, agents at the individual level cannot be directly accessed across the org to add another layer of protection.

Good to know: A Virtual Teammate builds their tiers around their manager (the person who set it up), their teammates, and everyone else in the organization. Managers sit at their own level, separate from principals. Every tier gets its own conversational guardrails.

Say a company runs a big Jira setup and doesn’t want an AI pulling data from one project to another. The fix is a Jira login for the AI with access to a single project

Alex Whitney|Principal Product Manager at Optimizely

The same question needs a different answer depending who asks

Your product launch is embargoed until Thursday. A teammate on the launch asks you for the date, and you tell them. A freelancer working on a different campaign asks the same question, and you tell them it’s coming soon. Same question, different answer.

You make calls like this dozens of times a week, and each one takes a split second. A colleague asks about someone’s salary. You answer one way if they work in HR and another way if they’re not.

Two things change the answer: who’s asking, and who or what the question is about.

AI built for the enterprise has to make that call itself. It needs to know who’s asking, how they relate to the information, and what it’s allowed to say to them. That means defining tiers of people, each with its own conversational guardrails: what can be discussed, with who, and what’s off the table.

In practice, that looks like blocked topics: Salary, personal finances, performance ratings, and confidential conversations among executives. A marketing team might add agency rates and contract terms, for example.

Given the speed everything is growing and developing, who knows what the future holds. We’ve got our eyes on a few ideas:

Answers that change with the audience: Ask about a campaign that’s not public yet and the right reply depends on whether they are working the campaign or not. Ask a freelancer for the total Q4 program budget and the right reply might be the budget for their own project.

Information that depends on the asker: A feature in alpha shouldn’t reach a pre-sales rep who talks to prospects every day. They get beta and generally available features only. In the same way, a marketer in EMEA who asks for the target account list should get EMEA’s list, not North America’s.

VTPersmissions.png

Every message needs two checks: one on the way in, one on the way out

Alas, knowing the rules isn’t enough. The AI has to apply them every time. One way to build that is two checks around the work.

1

Before it does anything, check who’s asking and whether to answer at all

2

Before it sends anything, check what this person is allowed to see, and redraft the reply to match.

Think of a mailroom editor who reads every letter coming in and going out. Incoming, they decide how the question should be handled. Outgoing, they redraft the answer for its reader.

Here’s an example of what’s happening when one of our Virtual Teammates says they are “thinking”:

1

Classify the sender: Manager, teammate, or someone else? A manager has privilege to fully access the teammates work and knowledge.

2

Screen the incoming message: For everyone else, a filter decides whether to reframe the question, and whether to answer or ignore it. If it won’t answer, it sends a short “sorry, I can’t reply to that” instead of going silent.

3

Do the work: If it’s answering, it goes into full response mode: researching, browsing the web, using tools, calling other agents. It does every step without sending a word.

4

Screen the outgoing reply: A second filter checks what this person is allowed to see and redrafts the reply to match.

5

Send: What arrives has been checked twice.

Yes, that tiny indicator of “thinking” is hiding a lot of work.

directory_VT.png

Why the check is its own step, not a line in a prompt

It’s tempting to write the rules into the AI’s instructions and call it done. The risk is that instructions can be bypassed. Make the check a separate step and the AI has to apply the rules before it can move forward.

Good to know: In a Virtual Teammate, each job is owned by an agent or bundle of agents. That makes it not deterministic the way code is, but close to it, because the agent does that one thing and nothing else. Each VT also has a self-evaluation job (agent) that it runs, which assesses its own performance after each run, scoring outcomes against expected results so every future run is better than the last.

A personal assistant works for you, enterprise AI works for you AND the business

A personal agent’s focus is the individual. It reads your email, watches your calendar, and does what you’ve asked, because its goal is to help you. Enterprise AI has a different job. Sure, there should be elements that benefit users at an individual level (hello, our Chief of Staff Virtual Teammate), but its core focus is the business. The priority is the goal the business gave it, the same way you have goals for your organization. Both have their place. One makes individuals more proactive. The other drives the work itself.

That changes what you expect from it. Ownership scales with the role, too. A high schooler shadowing you for a week waits for instructions. A 10-year specialist owns the outcome.

There’s a practical bonus. Agents, skills, tools, and connectors add up to a lot to learn, and plenty of people burn out before they find the best setup. Enterprise AI should hide all of that.

So, I’ve got a marketing analyst Virtual Teammate, and no, I do not want a daily update. I want a message when there’s something to say — which is exactly what I’d expect from a real human hire

Kevin Li|SVP, Product at Optimizely

Five questions  to ask before an AI teammate meets your whole team

1

Whose identity does it use when it connects to your tools, and can you limit it to one project?

2

Can it tell who’s asking, and does the answer change with the asker?

3

What can’t it talk about, and who decides?

4

Is the guardrail a separate step, or a line in a prompt?

5

Does anything check the reply before it goes out?

Instinct is great at its job. A marketing organization of 200 needs something with an identity, a rulebook, and a check on every reply. That’s what we built Virtual Teammates in Mark to be.