Fracas Digital • Aug 22, 2026 • 9 min read
AI Agents for Business: What They Actually Do (Beyond the Hype)
An AI agent is software that can plan, use tools, and complete multi-step work with minimal human input. That is different from a chatbot, which answers one question at a time, and different from an RPA bot, which follows a fixed script and breaks the moment an input looks unfamiliar. Businesses use agents for support, community management, internal knowledge lookup, and scheduling. The decision that actually matters is not which platform to buy. It is whether an off-the-shelf tool fits your workflow, or whether you need one built around it.
Almost every article ranking for this topic answers it as a shopping question: which of the twelve platforms should you subscribe to. That is a fair question if your workflow is generic enough to fit a template. Most business workflows are not, and nobody writing a platform comparison has actually run one of these things in production long enough to tell you what breaks.
We build and operate AI agents, we don't just sell you a login. Fracas designs, deploys, and runs agent stacks for crypto and Web3 teams, then reports the numbers. See what we build.
What an AI agent actually is
Three terms get used interchangeably and shouldn't be.
A chatbot answers a message and stops. It has no memory of what it did five minutes ago beyond the current conversation window, and it cannot take an action outside the chat.
An RPA bot (robotic process automation) follows a fixed script against structured data. Click here, copy this field, paste it there. It is fast and reliable, right up until the input format changes slightly, at which point it fails silently or loudly, but it fails.
An AI agent reads a goal, breaks it into steps, calls tools or APIs to execute those steps, and adjusts when something does not go to plan. It has a working memory across the task. Give it "resolve this support ticket" and it can look up the account, check the order status, draft a reply, and escalate if the answer needs a human sign-off. Give a chatbot that instruction and it can only talk about doing it.
McKinsey's State of AI report found that 62% of organisations are at least experimenting with AI agents, but only 23% are scaling an agentic system in even one business function. That gap between piloting and running is the real story, and it is where most of the guidance available online stops being useful.
Where businesses are actually using agents right now
Skip the platform list. Here is what agents are doing inside real operations, sorted by function.
Customer support triage. An agent reads an incoming ticket, classifies intent, checks account and order data, and either resolves it directly or routes it to a human with full context attached. The win is not replacing support staff. It is removing the fifteen minutes a human spends gathering context before they can even start answering.
Community and social operations. Discord and Telegram communities generate volume that does not fit office hours. An agent can moderate spam, score sentiment, flag anything that looks like a coordinated attack, and hand genuine questions to a human moderator with the relevant history attached. This is the function Fracas runs in-house, and it is covered in detail below.
Internal knowledge retrieval. Instead of a colleague pinging five people to find out where a policy document lives, an agent with access to the company's internal wiki, Slack history, and shared drives answers directly, citing the source document. This is the single fastest win for most mid-size teams, because the workflow is low-risk and the payoff is immediate.
Scheduling and operations admin. Booking calls across time zones, chasing meeting confirmations, reconciling calendar conflicts. Deliberately unglamorous work, and exactly the kind of repetitive, rules-plus-judgement task an agent handles better than a person, because it never gets bored on the fortieth reschedule of the week.
For a broader look at where the line sits between rule-based automation and agent-based reasoning, our guide to what AI automation actually is covers the underlying mechanics.
Platform or custom build: the decision the listicles skip
Every ranked "best AI agent platforms" article treats this as a menu. Pick platform 4, 7, or 9 depending on your budget, done. That framing only works if your workflow already matches what the platform was built to handle.
Off-the-shelf platforms are the right call when your workflow is common and well-documented: a support agent answering FAQs from a knowledge base, a scheduling assistant syncing two calendars. The platform vendor has already solved that problem for a thousand other customers, and you inherit the fix.
A custom build earns its cost when your workflow touches systems the platform was never designed for, when the judgement calls are specific to your business (what counts as a genuine escalation in your community, not a generic one), or when you need the agent to hold context across tools that don't talk to each other out of the box. We wrote about the specific build-versus-buy trade-offs, including the cost side, in our comparison of agency builds versus in-house agent development.
The tell that you have outgrown a platform: you are writing workarounds to make the platform's rigid triggers behave like judgement. At that point you are paying platform fees for something that increasingly needs its own purpose-built system anyway.
What a working deployment actually looks like
This is the part the buyer's-guide articles cannot write, because none of them operate one.
Fracas runs its own agent stack across Discord and Telegram for community management, alongside internal ops. We automated moderation tagging, sentiment scanning, role assignment, and governance notifications. The moderation queue went from 200+ daily items, most of it spam, down to 15 to 20 items that actually need a human judgement call. Response time to genuine member issues dropped from 4 hours to 20 minutes. Manual community workload fell by roughly 60% overall.
None of that happened on day one. The first two weeks were spent watching the agent make wrong calls on edge cases, tightening the escalation thresholds, and rewriting the sentiment prompts until false positives on sarcasm stopped clogging the human queue. That tuning period is the part every generic guide skips, and it's the part that decides whether a deployment actually works.
The full breakdown, including the specific workflows and the Discord-versus-Telegram split, is in our piece on AI agents for crypto community management.
Why most agent projects fail
Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027. Having built and run several of these, the failure pattern is consistent and rarely technical.
The scope was undefined. "Automate customer support" is not a brief. "Handle password reset requests and billing enquiries under £50, escalate everything else" is a brief. Teams that skip this step end up with an agent that either does too little to matter or makes calls nobody signed off on.
Nobody owned the tuning period. An agent needs two to four weeks of active monitoring after launch, where a human reviews its decisions and corrects the thresholds. Projects that treat launch as the finish line, rather than the start of a feedback loop, degrade fast because the agent never learns from its early mistakes.
The workflow needed judgement the agent wasn't built for. Strategic decisions, genuinely novel situations, anything requiring institutional context the agent was never given, still need a person. Agents that get pushed past that boundary produce confidently wrong output, and that is where trust in the whole project collapses.
The one action worth taking this week if you are evaluating agents for your business: write down the exact decision boundary before you touch a platform or a developer. What does the agent decide on its own, and what gets escalated. If you cannot answer that in two sentences, the project is not ready to scope yet.
FAQ
What is the difference between an AI agent and a chatbot?
A chatbot answers questions within a single conversation and cannot act outside it. An AI agent can plan multi-step work, call external tools or APIs, and carry context across the whole task, not just the current message.
Do small businesses actually need AI agents, or is this an enterprise thing?
Smaller teams often get more relative value, because the workflows tend to be simpler and off-the-shelf platforms fit without customisation. The knowledge-retrieval and scheduling use cases above are usually the fastest wins for a small team.
How long does it take to deploy an AI agent for a business workflow?
A single, well-scoped workflow on an existing platform can go live in one to three weeks. A custom-built, multi-step agent typically takes six to twelve weeks including the post-launch tuning period, which is not optional if you want it to actually work.
Should I buy an AI agent platform or have one built?
Buy when your workflow is common and matches what the platform was designed for. Build when your workflow touches systems the platform doesn't support natively, or the judgement calls are specific enough to your business that a generic template will not hold up.
Why do so many AI agent projects get cancelled?
Mostly undefined scope, no owner for the post-launch tuning period, and asking the agent to make judgement calls it was never equipped to make. None of those are platform problems. They are project-management problems that show up regardless of which tool you pick.
If you are trying to work out whether your workflow needs a platform subscription or a built agent, that is usually a thirty-minute conversation, not a months-long evaluation. Book a call and we'll tell you honestly which one it is.