AI agents for business and the workflows they reshape

1. Observe

Before an AI agent can reason, it must perceive. This phase is about collecting relevant signals from the environment: user interactions, application programming interface (API) responses, data streams, and sensor input.

Where automation stops at preprogrammed triggers, agents continuously learn what matters and what doesn’t.

📋 Real-world example: Amazon Alexa’s ambient learning update introduced contextual recall. Now Alexa notices patterns in user routines like adjusting thermostats or lights. Doing so lets it suggest similar actions next time. That’s smart observation in action.

“Think of strong observation as intelligent contextual awareness. Agents that see well now make better decisions later.”

{{Anna Demianenko}} 

2. Plan 

Once an agent understands its environment, it shifts into planning mode.

This is where LLMs or small language models (SLMs) act like executive brains. They weigh options and sequence actions toward a goal.

📋 Real-world example: GitHub Copilot doesn’t simply autocomplete code. It interprets developer intent, reviews context from multiple files, and generates a plan. It’s constantly planning and replanning as the developer writes.

3. Act 

The action phase is where planning meets production.

The agent takes its internal strategy and interfaces with the real world through APIs, CRMs, databases, or software tools. This is also where adaptability separates smart systems from fragile ones.

📋 Real-world example: AutoGPT showcased how agents could autonomously write, test, and iterate on code. Given a goal (“build a simple website”), AutoGPT planned the process, executed tasks through the web, corrected errors, and completed the job. And a continuous feedback loop is what powered the process.

1. Design around the workflow

Agents fail when they’re built in isolation. Map your process first. Then, take the time to locate points of friction and design agents to remove them. 

❌ Avoid: Treating agents as isolated apps.

✅ Do: Reconstruct the full workflow and place the working agent where it adds leverage.

2. Onboard and assess like an employee 

Define the agent’s “job” and KPIs. Run structured tests comparing AI outputs to expert judgment and retrain from real usage data.

❌ Avoid: One-time QA or demo testing.

✅ Do: Continuous performance monitoring tied to business metrics.

3. Build observability

Track every input and decision. Neglecting this step means errors compound without you even noticing. Use dashboards and explainability logs to identify logic failures as soon as they surface.

❌Avoid: Measuring only outcomes.

✅ Do: Monitor reasoning at every step.

4. Use agents where reasoning matters

Agents excel in judgment-based tasks. Think user research, trend forecasting, and performance analysis. For rule-based work, simple automation or ML is usually cheaper and safer.

❌ Avoid: Over-automating predictable tasks.

✅ Do: Apply agents where adaptability adds value.

5. Keep humans in the loop

When switching some processes to agentic AI, humans must stay in the feedback loops to validate and guide agents. To ensure efficiency of the human-agent interaction, design interfaces that make AI reasoning visible.

❌ Avoid: “Full autonomy” mindset. 

✅ Do: Build human-agent collaboration with clear oversight points.

1. Finance

The finance industry is built on information asymmetry, meaning whoever processes data faster wins. And that’s exactly what Accern.Rhea set out to fix.

We helped Accern design an agentic research platform that does more than summarize data. In particular, it:

• observes real-time market signals,
• plans research pathways through dynamic filtering,
• acts by generating ready-to-send reports and trend summaries.

With a hybrid GUI/prompt interface, users can interact with the system through chat-like language commands or by using visual controls within the interface. The agent switches between datasets, retrieves filings, and builds institutional-grade reports while learning from user edits.

🏅 The outcome: Analysts cut hours of manual research, VC teams gained faster due diligence cycles, and the product helped push the client from Series B to acquisition, raising over $40 million in the process.

2. Media and news 

In a world drowning in opinion, curating truth is the next frontier. For VTNews.ai, we designed an AI newsroom that functions much like a network of collaborating agents.

One agent scans 130k+ sources in real time. Another compares coverage from left, right, and center-leaning outlets. A third predicts prospective trajectories of news narratives. Together, they provide readers with a 360° perspective alongside bias analytics and event timelines.

🏅 The outcome: 90% of users reported escaping their “information bubbles”. Within the first month, the platform attracted 85k new users.

3. Legal

Legal teams face an avalanche of new case law every month. Our redesign of eDiscovery Assistant turned what used to be a static database into an intelligent legal research engine.

It now observes new rulings, categorizes them by relevance and jurisdiction, and acts by surfacing the most pertinent cases within seconds. The interface standardizes complex workflows by pulling up rules, checklists, and related precedents in just three clicks. What it achieves is cutting research time by 75%.

🏅 The outcome: One law firm reported recovering $400K in damages after using the redesigned system to uncover precedents others missed.

4. Fashion and retail

Fashion commerce has long been shackled by high-cost, low-speed content creation. Mannequin literally flips the model.

Our team designed a sales and marketing platform for a fashion-tech pioneer that replaces expensive model photoshoots with AI-generated humans wearing real garments.

🏅 The outcome: The platform plans and generates localized assets by adapting poses, backgrounds, and ethnic appearances to match regional markets. For global retailers, that means instant production of campaign visuals fine-tuned for Tokyo or Paris.

5. Creator tools

Short-form video runs on volume, yet a single clip still takes timelines, cuts, and music choices, and that work stops many creators before they post. For Elva, a mobile AI video editor, we designed an agent that does the editing while the user stays in the director’s chair.

Elva works as an agentic editor. In particular, it:

• observes the footage, detecting scenes and mood, and coaches users while they film with real-time framing hints in the viewfinder;
• plans the edit from a spoken request, asking clarifying questions when intent is unclear;
• acts by assembling a finished, social-ready clip with rhythm cuts, music, and mood matching, then presents it as a draft for the user to approve.

The key UX decision was giving the agent a face. An expressive blob persona shows what the model is doing at every step: listening, thinking, editing, or done. Users always know where the work stands. Elva learns each user’s preferences over time. Premium features appear inside the editing flow, next to free alternatives, at the moment their value is clear.

🏅 The outcome: Elva won the Webby for Best Visual UI in AI.