The reality of AI agents in business operations

Companies face a basic math problem: growth requires more work than people have time to do. The standard response has been workflow automation. But workflows follow scripts. They are rigid, and break when reality doesn't match the script.

AI agents work differently. They pursue goals, and thus have the "freedom" to work around scripts, greatly surpassing what was possible with workflow automation.

What Makes an Agent

Traditional automation follows "if-then" rules. When a customer emails, send response template #3. When inventory drops, reorder 100 units. Simple. Rigid. Often, not what you would like, just what is possible with the technology you have at your disposal.

An AI agent reads the situation and decides what to do. Given access to company systems and a clear objective, it:

Reads data from multiple sources.

Makes decisions based on both given instructions and patterns it infers.

Executes actions across systems.

Adjusts when things go wrong by reeavaluating the path of actions it has followed to achieve the goal. It reflects upon its work and can decide to redo it in a better way. 

Think of the difference this way: automation is bus 133. It runs the same route, stops at the same stations and follows the same schedule, executing the same loop on repeat. An agent is a taxi driver who sees traffic building on the usual route, takes side streets and drops you at a better entrance.

The Business Case

The numbers tell the story. On August 1st 2024, Andy Jassy (Amazon CEO) shared an exciting finding regarding the real and quantifiable impact that the Amazon Q Developer agent for code transformation offers IT and developer teams of any size. Amazon has migrated tens of thousands of production applications from Java 8 or 11 to Java 17 with assistance from Amazon Q Developer. This represents a savings of over 4,500 years of development work for over 1,000 developers (when compared to manual upgrades) and performance improvements worth $260 million dollars in annual cost savings.

The question is not *if* your business will adopt AI agents, the relevant questions are around *when* and *how*.

Where Agents Work Today

I'll share a few (although decisevely not all) business areas where agents are already having impressive results in SMEs.

Email Management: A 30-person consulting firm deployed an agent to sort incoming emails, draft routine responses, and flag urgent items. Partners now save 90 minutes daily on email triage. The key to success: they configured the agent to escalate any email from key accounts or containing specific emotional cues, ensuring important context isn't missed.

Sales Follow-up: An 130-employee software company uses an agent to track prospect interactions and send personalized follow-ups. Conversion rates jumped +22%. They maintain human judgment by having reps review the agent's "interest scoring" weekly, allowing them to catch any misread signals before opportunities turn sour. Having "digests" based on CRM notes + emails of your previous interactions and key decisions that were made, remain open, were not followed-up on, before calls is something reps have found very useful.

Meeting Documentation: A 150-person manufacturing firm's agent attends video calls, captures decisions, and assigns action items. Nothing falls through cracks anymore. They built in a verification step: technical leads review summaries within 24 hours, ensuring specifications are captured correctly before work begins.

Investment Materials Generation: A public-credit hedge fund uses a suite of AI agents to write out the initial investment materials on an investment opportunity they are analyzing. When an analyst identifies a potential investment, the agent pulls data from financial statements, rating agency reports, and market comparisons. It drafts a 1, 3 and10-page memos with the specific information (and specific depth of information) the firm wants, based on the type of document it is, ensuring a continuation of their existing process and an operationalization of their investment know-how that allows them to cover 60% more opportunities at the top of their funnel. 

Client Research: A 60-person B2B services firm deployed an agent to research prospects before calls. It compiles company news, salient posts in social media related to your product/pain it solves by the person reps will speak with, and potential pain points into one-page briefings. Sales reps enter meetings better prepared, leading to 35% more qualified opportunities.

The Deployment Trap

Most articles promise easy deployment through "no-code platforms." This is misleading. 

Here's what actually happens:

You map your process. You realize your process makes no sense. You fix the process. You configure the agent. The agent fails because your data is messier than you thought. You clean the data. The agent works but makes decisions you didn't expect. You add rules. The agent becomes as rigid as old automation. You start over.

A manufacturer learned this the hard way. They deployed an agent to optimize supply chain orders. It worked perfectly, reducing costs 15%. Then a key supplier had a quality issue. The agent kept ordering from them because the metrics looked fine. By the time humans noticed, they had warehouses full of defective parts.

A More Honest Framework

If you're going to deploy agents, expect this sequence:

Phase 1: Process Archaeology: Document what actually happens, not what the manual says. The invoice process you think takes 5 steps probably takes 11, with 6 unofficial workarounds. Find them all.

Phase 2: Power Struggles: Decide who controls the agent. IT wants governance. Business units want speed and reliability. Vendors want lock-in. These conflicts must be resolved before deployment, not after.

Phase 3: Capability Building: Your best people need new skills. Not coding, but agent supervision and really habituation to using these tools. They must learn to set objectives, interpret agent decisions, and apply critical thought appropiately. Budget 3 months minimum.

Phase 4: Failure Planning: Agents can fail in ways you can't predict. Build manual overrides, where a human in the loop can easily intervene. Create fallback processes. 

Strategic choices to be made

Agents change power dynamics. Technical teams gain influence. Business units may be reticent to adopt the technology or change their workflows. Vendors accumulate control. Consider:

Buy vs Build: Building gives control but requires time and expertise. Buying means faster deployment but vendor dependence. Most companies need both, which few vendors support.

Competitive Dynamics: When everyone has agents, efficiency becomes table stakes. Advantage comes from using agents differently, becoming deeply atuned with how they work and continually improving/tweaking them to experiment, not just having them.

Organizational Learning: Companies learn through repetition and variation. Agents reduce both. You need new mechanisms for building expertise.

Moving Forward

Start small. Pick a process that's painful but not very risky. Expect 3 months from planning to stable operation. Budget for failure.

More importantly, be honest about the trade-offs. Agents can process invoices faster than humans. They can't tell you why a vendor relationship feels wrong. They can screen resumes efficiently. They can't spot the brilliant misfit who becomes your best hire.

The question isn't whether to deploy agents. They're coming regardless. The question is whether you'll do it with clear eyes, and the right information in your hands about what you're gaining and what you're losing.

The companies succeeding with agents share one trait: they stopped pretending this was just another software deployment. They recognized it as organizational redesign. Take it as such.