How Human Uses AI Agents (and Where Humans Still Make the Call)

Ask a marketing agency whether it uses AI, and the answer is probably yes. The more useful question is: What is AI doing, and who's responsible for the result?

For the brands we work with, the answer to that question matters. An agency promising more efficiency can sound appealing until “efficiency” starts to mean less attention, less expertise, or a client relationship managed by a bot.

That's not how we're approaching it at Human. We're changing how the work itself gets done. Instead of adding AI to the same chain of handoffs, waiting, and repeated briefings, we're building workflows where AI agents investigate performance, surface hypotheses, and produce complete first versions of the work. Our strategists challenge that thinking, decide what moves forward, and own the result.

We're confident in the direction. What we're still working out is which approaches hold up across real client work, and we'll be clear about which is which. Here's what's working, what we're still testing, what people own, and what it means for the brands we work with.

Main Points

• AI agents can investigate, recommend, and produce complete work, while people set direction, challenge the thinking, and approve what moves forward.
• The bigger change is the workflow itself: fewer handoffs, less waiting, and one owner directing work that used to pass through several people.
• Output quality depends on what Human brings to the system: client context, brand voice, connected data, and expert review.
• When an experiment falls short, we treat it as a lesson about inputs and process, then change how we run it next time.
• For clients, the opportunity is ideas that move faster, insights that become connected campaigns, and room to try work that wouldn't have fit the budget before.

Why we're rethinking how agency work gets done¶

AI has changed the economics of getting a first draft, a data summary, or a basic analysis on the page. Those capabilities aren't exclusive to agencies; in-house marketing teams, founders, and freelancers can all access them.

That raises the bar for what a good agency should contribute. The value can't be “we can produce something you could also prompt a tool to produce.” It has to be knowing which problem to solve, what evidence to trust, what trade-offs to make, and how the work connects to growth.

Simply dropping an AI tool into an existing process doesn't fix the process. Our team has gotten faster at individual tasks, but the work still moved through the same handoffs: a strategist briefs a copywriter, the copywriter waits on data, a designer gets briefed again, and each step adds time and a little more drift from the original idea. Faster tasks inside a slow workflow only get you so far. So we're asking a different question: If we designed this job today, which steps need a specialist's judgment, and which could a well-directed system carry out?

We start with the outcome, not the technology. Then we decide where automation is useful, what context it needs, what access is appropriate, and where a person must review the result. The point is to redesign work thoughtfully, not automate it for the sake of saying we did. That's a bigger change than speeding up individual tasks. It changes who owns a piece of work from start to finish and how many times it has to be explained along the way.

Three ways AI shows up in our work¶

We think about AI in three practical modes. They overlap, but the distinction helps us assign responsibility.

AI assistant. A person works with a tool to research a topic, explore ideas, interrogate data, or develop an early draft. The person drives each step, checks the evidence and turns the material into usable work. Think of it as support at the desk, not an independent project owner.

Directed system. A person defines the task, supplies the right sources and standards, and has a system produce a specific deliverable, such as a first reporting deck, proposed alt text, or a set of page updates. The system executes a bounded job.

Autonomous agent. An agent can work through a multi-step job on its own: investigating why performance changed, developing hypotheses, recommending priorities, or producing a complete creative package across formats. People define the scope and the standard.

More autonomy doesn't mean more authority. Across all three modes, we think of the work as a sandwich: people provide the strategic inputs and context; AI takes on a growing share of the analysis and production in the middle; people challenge it, apply judgment, and approve what ships.

What our agents are doing today¶

Some of these workflows are already part of our day-to-day work, and some are still being proven out. We'll say which is which. None of them is a promise that every task on every account follows the same process. Every one runs on the model above: people set the direction and inputs, and someone on our team reviews the work before it reaches a client or a live site. Where that review involves something specific, we say what it is.

Reporting that starts with data, not copy and paste. Monthly reporting has traditionally involved pulling numbers from sources such as Google Analytics, Google Search Console, and Semrush, moving them into slides and checking that every figure landed correctly. We're using an agent to support that collection and deck-update work. The time that saves goes into explaining performance, which we come back to below.

AI visibility checks that answer specific questions. Brands want to know how they appear in AI-generated answers, not just where they rank in traditional search. We now use an agent to run a defined set of priority questions through AI tools and record how a brand shows up in the answers. That makes a manual research task more consistent and repeatable. No single response is a universal measure of AI visibility, though, so a strategist chooses which questions matter, interprets the patterns, and decides what follow-up work is justified.

Alt text audits and updates. Finding missing image alt text across a website can involve a long sequence of page checks and edits in the content management system (CMS). An agent can identify gaps, draft descriptions from the image and surrounding content, and make updates in a single pass. Accessibility and accuracy require more than filling an empty field; descriptions must match the image's actual purpose and context.

Image metadata for content migrations. In a separate content workflow, an agent has helped add descriptive metadata to images and icons inside a CMS so that later migration work can identify appropriate visuals more easily. Instead of relying on filenames or a person's memory of an asset library, the migration can use better descriptions.

FAQ updates from approved copy. Publishing FAQ revisions across multiple pages is a good example of a bounded execution task. One workflow we've built takes already-approved copy from a Google Doc and places it into the appropriate reusable FAQ fields in Shopify. The strategist owns the questions and answers; the agent handles the repetitive placement.

Outreach research. Prospecting for relevant sites takes research: assessing fit, finding contacts, and tailoring an introduction. We're testing an agent on Human's own outreach first, having it gather potential opportunities and prepare draft messages for review. An agent can't automate an authentic relationship into existence, and whether recipients respond is a separate question from whether the technology can send a message.

A complete example: one brief, one connected holiday campaign¶

Task-level examples only tell part of the story. The bigger shift shows up when you follow one piece of work from start to finish, so here's the model we're building toward for a holiday campaign.

In the traditional version, a strategist writes a brief, then the ads, landing page, and email teams each get briefed separately. Each group interprets the idea a little differently, waits on the others for assets or data, and the pieces come back together late, with someone reconciling messaging that drifted along the way.

In the version we're building, everything starts with one strategic brief grounded in the client's customer insights, offer, and brand voice. One owner directs the production of the ads, the landing page, and the email sequence from that same brief, so every piece carries the same idea from the start. Specialists in paid media, design, and email then review how everything works together: whether the ad's promise matches the landing page, whether the emails build on the same offer, and whether the whole campaign sounds like the brand.

The result is less coordination and more cohesive work. To be clear about where we are: producing individual pieces like reports, creative, and site updates with agents is already part of how we work. Running a full campaign this way, end to end, is what we're testing now, and we'll judge it on whether it performs better than the old way, not on how fast it came together.

Where people stay in charge¶

An agent can investigate a drop in performance, propose explanations, and recommend what to fix first. What it can't do is own the decision. That takes context, judgment, and accountability.

Our strategists still decide what to prioritize, which conclusions the evidence supports, and which recommendations make sense for a client's goals, timing, and budget. They connect findings across channels, challenge an implausible number, and distinguish a temporary fluctuation from a meaningful signal.

Consider reporting. Pulling the data is only the opening step. The valuable questions come afterward: Why did this happen? Is the change real? What else should we investigate? What is the client going to ask, and what should we do differently next month? Agents can now draft answers to several of those questions, too. Our job is to test those answers against what we know about the client and decide which ones hold up.

The same principle applies to content and creative. Good work requires brand understanding, factual checks, technical knowledge, and taste. A polished-looking draft can still be generic, misleading, or wrong for a particular audience. Our standard is the client's standard, not the speed at which a tool produced a result.

The human part also includes the relationship. Listening to a client, explaining a trade-off, handling a difficult decision, and owning the consequences are not production tasks to be delegated to an agent.

Why the system matters more than the tool¶

Clients can open the same AI tools we use. What they can't download is everything that makes those tools useful on their account.

That starts with understanding the client: their customers, margins, seasonality, and what they've tried before. It includes having their brand voice and messaging documented well enough that an agent can work from it, and connecting the right data so recommendations rest on real performance instead of guesses. It means building repeatable workflows that get better each time they run, and feeding what we learned from previous work into the next project.

Setting that system up, directing it, improving it, and owning the result is the work. A well-built system with weak judgment produces confident mistakes. Strong judgment without a system means good ideas wait in line. We bring both.

What hasn't worked yet¶

Some AI experiments are impressive as demonstrations and underwhelming as finished work.

In one test for Human's own social channels, the team supplied substantial context and used AI to develop research, creative directions, copy, and multiple concepts in a relatively short session. The volume and speed looked promising. The output itself still felt too generic to publish without meaningful refinement.

That result matters. A useful tool can accelerate ideation without producing a final deliverable. It may give us options to critique, not options ready to schedule. If the concepts lack a distinct point of view, if the copy could belong to any company, or if the creative doesn't meet our expectations, a person has to say so.

We evaluate AI-assisted work the way we evaluate any work: Does it solve the brief? Is it accurate? Does it meet the brand's standards? Is it worth the time and cost to refine? Sometimes the answer will be to adjust the inputs, narrow the task, or stop using AI for that part of the process. Experimentation only builds trust if we're willing to report what didn't work.

What we learned is that context alone wasn't enough. The prompt was detailed, but nobody had set a point of view before AI started producing, so the concepts landed on what any company might say. Next time, a strategist will define the angle and the audience tension first, and the system will draw on examples of our past posts that performed, so AI develops a specific idea instead of generating options from scratch. It's the same lesson behind the rest of this post: a clear owner and the right inputs change the output more than a better model does.

Preparing our site and our clients for AI agents¶

Agents aren't only changing how we work. They're changing how customers find and buy from brands. More shoppers now ask an AI assistant what to buy and trust the answer, which means a brand's site, product data, and reputation need to be readable by machines as well as people.

We started with our own site. Through WebMCP, Human exposes its services, case studies, team directory, and a fit assessment as tools an AI agent can call directly, so an assistant pulls live answers from the site instead of guessing from a stale summary. Our site is built on Sanity, so each tool draws from the same structured content that powers the pages people see, with no separate version for agents to fall out of sync. Building it ourselves is how we practice what we tell clients to prepare for.

Further Reading: How I Built WebMCP on Human’s Website With Sanity

For ecommerce brands, that preparation is agentic commerce: making sure AI systems can understand, trust, and accurately recommend their products. The work spans product data, structured content, and trust signals like reviews and third-party citations. It also means deciding which products should be eligible for AI recommendations, so AI-driven demand supports margin instead of eroding it.

Measurement has to keep up, too. Profit Compass, the profit intelligence software we built in-house, tracks the sessions and revenue that come from ChatGPT, Gemini, Perplexity, and Copilot alongside every other channel. It also flags anomalies, like spend spiking beyond plan, so issues surface quickly. A strategist still decides what those signals mean and what to do next.

What this means for clients, and where agency work is going¶

For clients, this should show up in ways they can feel. A good idea doesn't have to wait for five people to find time before it becomes something real. A customer insight from a review or a sales call can turn into a connected campaign across ads, landing pages, and email instead of a single post. And work that wouldn't have fit the budget before, like testing more creative angles or building a landing page for a smaller audience segment, becomes possible. We still have to prove these benefits account by account, and we won't promise turnaround or results we haven't measured. But that's the opportunity we're building toward.

The same model applies beyond campaigns. In an early test of AI-assisted website prototyping, we collapsed a long, handoff-heavy process into a few phases: shared context, an AI-assisted brand direction and prototype, then a joint review with strategists in the room. A full concept came together in about an hour. That was a test, not yet a production method.

Roles at Human are already shifting away from doing each production step by hand and toward defining the problem, supplying context, directing systems, validating work, and owning outcomes. That shift makes expertise more important, because a weak brief, missing brand context, or an unchecked assumption can now travel further and faster.

Our aim is straightforward: use technology where it helps us do better work, while keeping the experience of working with Human unmistakably human.

Frequently asked questions¶

The next chapter is still Human¶

The future of agency work isn't about choosing between AI and people. It's about using each for what it does well.

For ecommerce and B2B brands, Human aims to be a growth partner that combines smarter systems with experienced strategists who ask better questions, make deliberate decisions, and stand behind the work. AI expands what our team can investigate, create, and test for every client. The judgment, accountability, and relationships that make the work valuable remain human.