How is Gigi different than "company x's" AI agent?

I find myself having to answer this question a lot, and thought it would be helpful to outline an approach for any prospective buyer or reviewer of "agentic" technology in advertising.

That this question gets posed so often speaks to two things happening at once. First, there is an understandable gap in AI education in our industry. We've grown accustomed to assessing the capabilities of legacy SaaS tools. There's a comfort and understanding of how to distinguish feature sets and functionality of legacy SaaS providers that simply does not exist with agentic AI. Second, there is a massive disconnect between the marketing of agentic AI products in advertising and the actual value they provide. We're at a point in this cycle in which every technology provider has (rightfully) led their marketing efforts on their AI capabilities when in reality their AI bets are often co-pilots that sit on the bottom right of a legacy SaaS product UX, or aspirational capabilities of things that AI could potentially do in the distant future (see: Amazon's Ad Agent). Those two gaps compound. A buyer who has no framework for evaluating agentic AI walks into a market where everyone claims they can provide a panacea for all business problems.

I'm not writing this to prove why Gigi is different (we are). I'm writing this to allow any buyer to make better informed decisions on their agentic AI bets both now and in the future.

Consider this a helpful guide on how to assess AI agents within ad tech. Three questions: what tools can it call, what did they build around the model, and does it change the way my team works.

Understanding tool calling is the most basic and easiest aspect of agentic AI to assess as a potential buyer. In legacy SaaS, buyers would often create multi sheet RFPs that asked vendors to exhaustively list all of the features they offer. Tool calling capability is, at its simplest form, an outline of the things that an LLM can do.

For example, if I asked Claude to amend third-party pre-bid settings on a line item in a DSP, Claude would be unable to do this because Claude does not have an API endpoint that allows it to take this action. At Gigi, we've spent 20 months exhaustively integrating hundreds of API endpoints in Amazon's ad tech as tools that Gigi can call, so that Gigi can take action on behalf of our customers or use these tools to power its general intelligence. Some quick examples: read/write capabilities on line items, custom AMC metrics that we've built, AMC reports, inventory reports.

We've chosen to go a mile deep on the Amazon DSP, because that's what our customers demanded of us. We've seen other agentic AI companies go broad across media channels. Rather than exposing hundreds of tools on a single media channel, the way we have, they may choose to expose a consistent tool (i.e. budget management) across a number of media channels as the tools their AI agent has access to.

The power of agentic AI is that agents can take action in ways that legacy SaaS never could. The actions an agent can take are directly tied to the tools that it can call. If you want to understand what the agent you're hiring can do, find out what tools it can call.

The "harness" has quickly become the catch all phrase for all the sophistication an agentic AI company builds around the LLM itself.

The harness is everything wrapped around the model that turns raw intelligence into a worker. The LLM supplies reasoning. The harness decides what the model sees before it reasons, what it's allowed to touch, how long it can work before it stops, what it remembers between sessions, and what happens when it gets something wrong. Every AI company can buy the same models. What each one builds around them is the product.

While tool calling is mostly visible to the end user, much of the work in building a harness around an LLM are the invisible things that make an AI agent unique. When engineers say they prefer using Codex vs. Claude Code, oftentimes that preference is not rooted in the underlying intelligence of the frontier model, but rather a preference for the harness that OpenAI or Anthropic have built around their coding agents. The same is true in any vertical AI agent. The harness is all the work a company has done to create a unique and personalized product. The image below from investor Tomasz Tunguz is a helpful way to bucket all the aspects of a harness.

For the purposes of agents in ad tech, here are some helpful things to consider in understanding the harness of an AI agent:

Where is there automation? Where is there a human in the loop? Simon Poulton, EVP of Innovation & Growth at Tinuiti, has coined the phrase "agentic deference." The thought behind this phrase is that every enterprise, and every person at that enterprise, has different levels of comfort in deferring work to AI agents. A great agentic AI harness should offer the ability to meet each enterprise and end user where they are on their agentic deference journey, and provide critical controls to choose which actions can be automated or assisted with a human in the loop.

What context, workflows, and business rules can I bring to the agent? With agentic AI we no longer buy software, we hire software, and the expectation is that a buyer should be able to mold their AI agents to their specific business. This is something that we've leaned into heavily at Gigi, and it has allowed us to scale across enterprise agencies. As an example, our customers write agentic operating procedures, and Gigi follows them. Gigi builds and names orders, line items, audiences, and creatives in the exact manner unique to that agency, down to their naming conventions, without anyone on that team having to correct her afterward. The magic of AI is that we can customize and personify software in ways that legacy SaaS never could. We believe the ability to mold an AI agent to a specific business is the very reason that one should hire and deploy AI agents across an enterprise. Conversely, something like Amazon's Ad Agent may use AI to make the use of their DSP easier for beginners, but there's no ability to customize that agent for enterprises who already have defined ways of working within their own ad tech and a unique perspective on how to operate it and uniquely provide value to their clients.

Model selection and token costs. Gone are the days in which AI agents blindly default to frontier models from OpenAI and Anthropic for every task. The best AI companies now run a constellation of models, frontier, small, and open source, and route each task to the one that hits the right combination of speed, performance, and cost. Every one of those calls costs money, and for an AI-first company those token costs are the single largest input to gross margin. That gives a buyer a clean heuristic: check whether token costs are baked into the pricing model in some way. If they are, that company is spending tokens on your behalf to provide an AI-first product experience. If they aren't, that company is loosely using LLMs, and "agentic AI" in their marketing exposes surface-level usage at best.

Two agents can call the exact same tools and produce completely different work. The harness is the reason why.

Tool calling and harness education provide visceral ways for buyers to discern the potential impact and value of AI agents. But a more subjective way to view AI agents is understanding whether or not the AI agent fundamentally alters the way in which an enterprise works.

At Gigi, our promise to our customers is they can cut the operational costs of managing the media channels Gigi operates by 80%. We present demonstrable workflow efficiencies and back it up with data for each client that we onboard. Once deployed, the buttons team members used to press now get assigned as work to Gigi, and it's now the job of that team member to review, accept, and/or automate Gigi's work. P&L decisions are then made such that Gigi enables an enterprise to positively bend the operating model of the media channel.

Conversely, the lowest form of agentic work, and the most ubiquitous, is using AI agents to perform data analysis by asking questions to a chatbot. This obviously makes everyone more efficient, but headcount decisions can't be made by talking to an LLM. They can only be made when agents are able to take on work previously only possible by humans.

Changing the way one works is not just an economic assessment. The best AI products feel like magic. When assessing an AI agent, try to find magic. Did that company demonstrate magic to you in a demo? When your team first tries the product, do they feel the aha moment that we all felt when we experienced ChatGPT for the first time in late 2022? If so, lean in. Build alchemy together. If not, then why bother, this window of opportunity to redefine work is too short to experiment with AI fluff.

None of this requires you to become an AI engineer. Ask what tools it can call. Ask what they built around the model. Ask what work actually leaves your team's plate. Three questions. Most of what's being marketed as agentic AI today, including by the platforms themselves, can't get past the first one.