Three AI Agent Projects, Three Lessons for Enterprises

Two AI stories trended this week. Microsoft staff wrote that being pushed to use AI everywhere has made work harder, not easier. And a day-long Microsoft 365 outage showed how fragile AI tools become when the systems under them fail.

Both point to the same thing: agents only help when they're built into how the work actually runs. That's easy to say in the abstract. It's harder, and more useful, to talk about specific operations: what the work looked like before, what we built, where it got difficult and what we learned.

So this post is three of those stories. They cover very different businesses, and each one taught us something we now bring to every project.

1. DataBeat: an analyst's morning is a coordination problem

The work before

DataBeat monitors programmatic ad yield for publishers. Every morning, their analysts opened dashboards across multiple ad exchanges and multiple publisher accounts, and tried to answer one question: what moved overnight, and why?

CPM down on one exchange. Fill rate dropping for one publisher. A spike in one country. Each answer meant pulling data, comparing it against yesterday, against last week, against the same weekday, then writing it up for someone senior. That routine took four to six hours, every single day.

What we built

The first instinct, which we hear a lot, is "let's give the analysts a chatbot over the data". But a chatbot still needs someone to ask the right questions in the right order every morning. The routine itself stays manual.

Instead we treated the morning routine as what it really is: a team of specialists handing work to each other. Built on the first version of our SDK, it became more than a dozen agents, each with one job:

• one understands the question being asked
• one writes the SQL to get the data
• one detects anomalies against several baselines: previous day, seven-day average and day of week
• one explains the likely root cause
• one writes the narrative for executives
• one dispatches alerts to Slack, Jira or email depending on severity

A manager agent coordinates them and a summariser pulls the result together.

What changed

Daily yield analysis went from hours to minutes. Across the analyst workflows, time spent fell by 93 to 99 percent. We validated the anomaly detection on more than 175,000 rows of real revenue data before trusting it, and it runs in production across multiple publishers and exchanges, each kept separate.

The lesson

Anomaly maths is well understood. What made the routine repeatable was deciding which agent owns which decision, with people on the decisions that matter. An analyst's morning is a coordination problem, not one model call.

2. Wyra: a single looping agent breaks down in outbound

The work before

Wyra helps AWS Partners find and win customers. Before a lead ever reaches a CRM, someone has to watch the AWS ecosystem for signals: new certifications, competencies, marketplace listings, funding events. Then decide who's worth contacting, run the outreach and hand over only the qualified conversations.

That's a lot of manual prospecting, and it's exactly the kind of work that falls apart when the team gets busy.

What we built

We built Gil, an agent trained on AWS partner programs, playbooks and buyer personas. Gil reads ecosystem signals, scores leads by certification and partner tier, runs outbound sequences and makes voice calls in an AWS partner context. Like DataBeat, it was built on the first version of our SDK.

The important design choice was who owns what. Gil runs the sequences. Wyra's team, who are ex-AWS sales leaders, own the playbooks. Gil doesn't get to invent the strategy. It executes a strategy people wrote, and people can see and change it.

What changed

Partners can run targeted outreach at scale without manual prospecting, and pipeline arrives in the CRM already enriched. We didn't take that on faith: sequences, call quality and qualification were tested against human benchmarks before it went live.

The lesson

Outbound looks simple, "find people, message them". In reality it's dozens of signals, many steps and several systems. The obvious design is one agent that loops until the job is "done". That kind of agent drifts: it gets repetitive, chases weak signals and loses track of where each prospect is.

What makes it work isn't a better model. It's a defined role, people owning the playbook and a framework running the sequence step by step. Less freedom for the agent means more dependability for the business.

3. Mivi: a consumer AI product never stops shipping

The work before

Mivi makes earbuds. Their AI Buds add something new: avatars the owner wakes by saying "Hi Mivi". Each request has to be understood, routed to the right avatar, answered in the right tone and spoken back fast, all through a pair of earbuds.

What we built

There are five avatars, each with its own knowledge and way of talking: a Chef, an Interviewer, a Guru, a Therapist and a News Reporter. Language models understand intent and adapt in real time, while the earbuds handle audio and noise cancellation. Mivi runs on the LeafMesh SDK today.

What changed

The product launched with avatars that hold human-like conversations instead of just playing music. Owners get personalised interactions, and the avatars keep improving after launch.

The lesson

With a consumer product, launch is the start, not the finish. Five personalities on one voice interface means behaviour has to stay consistent while it keeps changing. Mivi's product team refining each personality, with the SDK running the agents underneath, is what kept quality steady as the avatars grew.

What the three have in common

These are three very different businesses: ad tech, B2B sales and consumer electronics. But when we laid the projects side by side, the same four things decided whether each one worked:

1. One job per agent. Specialists with one clear job are easier to trust, check and fix than one agent doing everything.
2. People own the judgement. Playbooks, personalities, thresholds and final calls stayed with humans who understood the business.
3. Coordination is the product. The manager, the hand-offs and the order of steps mattered more than any single model call.
4. Prove it before you trust it. Real data, human benchmarks and a clear "this is good enough" line before going live.

None of this is glamorous, and it doesn't make for a flashy demo. But it's what separates an agent that impresses a room from an agent that's still running, quietly and correctly, six months later.

If you're weighing up where to put your first agents, we're happy to walk through one of your processes with you. It's usually obvious within an hour where agents will help and where they'll just add noise.