We think our customers should own the harness they build inside Gigi, and they should own everything that harness learns about their business. Our job is to give them the tools to build it, and then to put that ownership in writing.

There's an argument running through enterprise AI right now about how much a company should hand over to the AI labs. The loudest version says you should own your own model. Alex Karp made that case on CNBC a few weeks back, arguing that enterprises need to "own the weights."

Across our customer base, we've learned that no agency or brand has the technical wherewithal to own the weights of a model. The weights are the model itself, and owning them comes with ongoing maintenance costs, capex included, to keep them current. As Jamin Ball put it last month, "a weight file is a melting ice cube that quickly becomes outdated." We're also increasingly learning that as models get good enough and commoditize, owning a model is far less compelling to most enterprises than owning the inputs to the harness and the intelligence created from those inputs.

There are many parts of the harness that Gigi uniquely offers, and that's our IP. The proprietary ML models we build, our agentic skills, our benchmarks and evals are a few of them.

But Gigi has been successful because we let our customers customize Gigi the way they would a team member. They build a knowledge base. They write agentic operating procedures. They give Gigi context on the brands they manage. Enterprise agencies bring collaboration signals and their own proprietary tech. Most of these are manual inputs, and they're what make each customer's Gigi personal to them. It's how you'd onboard a new hire. You don't own the person. You own the training, the standards, and the institutional knowledge you put into them.

Once the harness starts using what you brought to get better at your business, you have a loop. Satya Nadella described this in June as "building a learning loop on top of models where human capital and token capital compound," and he called that loop the new IP of the firm. He's right. We want our customers to build this loop and own the IP that comes out of it.

Over the past year we built the harness and laid the foundation for compounding loops. Only now can our customers start building new IP with Gigi. We're starting to do this in three ways.

Intelligent brand personas. Instead of asking customers to fill out a context form once during onboarding and never touch it again, Gigi builds her own working model of each brand and rebuilds it every 24 hours from DSP, AMC, retail and search signal. The persona captures performance personality, seasonality fingerprint, audience affinity, and the efficiency floor and ceiling a brand operates within. By month three it holds more context on a brand's media performance than any human on the account team.

Intelligent task loops. Today most agentic optimization stops at execution, as if the change were the outcome. Task loops evaluate every optimization against what happened after it ran: a clean baseline, an accounting of other changes in the window, the primary KPI plus the side effects, and a verdict of improving, plateauing, hurting or inconclusive with a recommended next move. The optimization has to earn the right to keep running.

Customer-specific models. For the past year, many of our enterprise customers have asked us to contract that we will never train custom models on their inputs or their clients' data. We've agreed every time. When the best model for every task was a frontier model, there was no reason to ask for anything else.

That paradigm is shifting. Open source models have reached the point where, for the right task, a model post-trained on a single enterprise's actual workload can beat a frontier model on performance, price and latency at the same time. Enterprises should be running toward this, on one condition: they own the loop from post-training onward. A model shaped by your decisions, owned by you, without standing up the machine learning team that would normally be required to build one. This is the furthest out of the three, and it's what we're working to enable in the coming months.

Our first version of advertiser context was a form filled out once at onboarding, which quickly became a stale media brief. Ball's point applies to all three. Owning a finished AI artifact is worth far less than owning the machinery that keeps producing better ones, because the artifact starts aging the day you get it. A brand persona from last quarter is a document. The system that rebuilds it every night is a strategic asset.

It's easy to talk about this. We wanted to get ahead of it, so we've amended our MSA with enterprise customers.

First, a functional and legal decision we've been proud to commit to from the start: containerization of customer inputs and performance data. Gigi has never used one customer's data to train or improve the agent for another. Customer data isn't commingled. We laid this foundation early, and it's part of the reason enterprise GTM has worked for us.

Now we've added a new defined term. Learned Context is everything Gigi derives from Customer Data and customer inputs in the course of providing the Services.

Learned Context belongs to each individual customer. Like Customer Data, it's containerized. It is not commingled, and it is not anonymized and rolled out to everyone else as a best practice.

Because our customers own their Learned Context, we export all of it, on request during the term and automatically within 30 days of expiration or termination, in readable Markdown and in machine-readable form, at no cost, before anything is deleted. If a customer leaves, they leave with the accumulated intelligence from collaborating with Gigi, in a format a human can read and another system can ingest. Our customers own the loop.

Learned Context takes time to build. Every day our customers collaborate with Gigi is a day they reinforce their new IP and contribute to the loop.

Every day an enterprise decides now isn't the right time is a missed day of building the new IP of the firm. Every month an enterprise runs on legacy workflows is a month of judgment that gets exercised and then evaporates. Nothing may be broken in their business. But a Gigi customer who started in Q2 2026 walks into 2027 compounding off six months of human capital acceleration.

Model capability will keep improving for everyone at the same time. There's no alpha in relying on model advancements. You get alpha by building the harness and owning the loop.