Of the twelve systems below, only two are actually AI agents. We built all twelve.
Most examples of AI agents published online are hypothetical. An agent that books your meetings, an agent that chases your invoices, an agent that qualifies your leads. No client, no number, no date it went live. The twelve below are systems we built and still run, each with a written case study behind it and a named step it took over: resume screening at a recruiting agency, property research for a real estate investor, quoting at a construction company, calendar policy at an investment fund. Some are genuine AI agents. Most are deterministic automation on n8n, the workflow tool we build on, some with a language model doing one job inside and some with no model anywhere, and calling those agents would be dishonest.
Thirteen years, 1,584 projects shipped, 200+ n8n workflows and 89 AI agents currently running in production. That last number uses the industry’s loose definition of “agent”, and by the stricter test in the next section, a good share of those 89 are automations with a model in the middle. We are stating that up front because the same slippage is what makes buyers feel cheated three months into a build.
Is it an AI agent, or automation with a model inside it?
An AI agent decides which tool to call based on the situation in front of it. An automation follows a path drawn in advance, even when a language model writes text somewhere along that path. That is the entire test, and it takes ten seconds to apply. Run it on your own process with the interactive version of this test.
Gartner named the problem. In its June 2025 press release, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, it estimated that only around 130 of the thousands of vendors describing themselves as agentic AI companies have genuine agentic capability, and it called what the others are doing “agent washing”. If that ratio is even roughly right, the base rate on any given “AI agent” pitch is poor.
The distinction is not academic, because the two things cost different amounts to build and to run. An automation has a fixed path, so it has a fixed price per run and a small number of failure modes you can enumerate on one page. An agent chooses at runtime, so it needs rules for what each tool it calls is allowed to do, and a place for a wrong answer to land before it reaches a customer. That is more engineering up front and more cost per run afterward. It also buys something automation cannot: the ability to handle inputs nobody anticipated when the system was drawn.
Neither one is better. Automation is the correct answer more often than the market admits. Of the twelve below, eight are automation, two are genuine agents, one is genuinely borderline and we say so where it appears, and one is not a system of either kind, it is the tool layer that agents call. Every one is labeled where it appears.
What do AI agents look like in intake and triage?
In intake, the job is deciding what an arriving thing is and where it goes, which sounds like agent work and usually is not. The path is known in advance. What varies is the content, not the sequence, so a model is doing classification inside a fixed pipeline rather than choosing a strategy.
AI Resume Screening for a Growing Recruiting Agency cut first-pass screening from 8 hours to 5 minutes per 100 resumes. Every application takes the same route: parse, score against the role, write the result where recruiters already work. A recruiter still decides who gets a call. That review point is not a courtesy, it is what makes a wrong score survivable. Automation, by our test, and it should be.
Automated Lead Follow-Up for a Real Estate Agency handles the part of the funnel that leaks fastest, which is the hours between an inquiry arriving and someone responding to it. No outcome metric is published on that one, so we are not going to attach a number to it here. Also automation.
The pattern in intake is that the expensive problem is response time and consistency, not reasoning. When the sequence is stable, giving a model authority over the sequence adds latency and a new class of failure for no gain.
Which examples of AI agents do research and extraction?
Research is where genuine agent behavior starts to earn its cost, because the next step depends on what the previous step found. A property with no permit history needs a different lookup than one with three open violations.
Automated Property Research for a Real Estate Investor took a manual 30 to 60 minute investigation down to 2 to 5 minutes per property. We call this one borderline. There is a defined set of sources, which makes it look like a pipeline, but what gets pulled and how far the system digs depends on what comes back early. If a page in these examples deserves an asterisk, it is this one, and pretending otherwise would fail the test we just set.
AI Estimation Tool for a Construction Company moved quoting from 20 minutes to 3 minutes. Estimators still own the number that goes to the client. The model reads the inputs and drafts; the human signs. Automation with a review gate, and the gate is why it was allowed near pricing at all.
AI Access to Job and Customer Data for a Field Service Team is a genuine agent. Somebody asks a question in plain language about a job or a customer, and the system decides which lookups answer it. Nobody drew that path in advance because the questions are not knowable in advance. No metric is published for this one.
MCP Server for an AI-Native PM Workflow is the counterexample that clarifies the category. An MCP server (Model Context Protocol, the standard way a model is given access to an outside system) is not an agent. It is the tool layer an agent calls, exposing project data to a model in a way that does not break every time the schema moves. Plenty of what gets sold as agent development is actually this, and it is valuable, and it is not an agent.
What are examples of AI agents in sales and follow-up?
Sales is the one area where the agentic version is often worth the extra cost, because a conversation genuinely branches and the branch cannot be drawn ahead of time.
Multichannel AI Sales Agent for a B2B Agency runs across 4 channels. Same prospect, same context, different medium, and the system decides what to do next based on what the person actually said rather than which step of a sequence they are on. That is an agent by any reasonable definition. If you are budgeting for a system like this one, budget the most engineering time for making this pattern safe, it is the one where a bad call reaches a prospect directly.
Automated Lead Generation for a 5-Person Sales Team brought cost per lead from $43 to $12. The sequence is fixed. Sourcing, enrichment, qualification, delivery into the pipeline the team already works from. A model writes and scores inside it. That is automation, and the $12 is a direct result of removing hand research rather than of any reasoning ability.
Worth naming the difference in what you buy. The lead generation system pays back as arithmetic you can check in a spreadsheet. The sales agent pays back in conversations that would otherwise have gone cold at 9pm on a Saturday, which is harder to measure and easier to overpromise.
Where is there no AI in the system at all?
Some of the most valuable systems we run contain no model anywhere, and they are usually the prerequisite for the ones that do. When two systems disagree about what a customer is, a model asked to reason across them produces confident nonsense.
Connecting Five Disconnected Systems for a Construction Company cut the time to catch an invoice mismatch between systems from up to 3 months to within 2 hours. Not faster reporting. A mismatch that used to sit unnoticed until a customer called now gets flagged the moment it happens.
Real-Time CRM Sync for a B2B Services Company has run with zero discrepancies in 6 months. That number is the whole product. A sync that is right 98% of the time is worse than no sync, because people learn to distrust it and go back to checking both systems by hand.
Neither of these is an AI agent and neither was ever going to be. We include them because a buyer shopping for agents frequently has this problem instead, and spending the budget on a model that reasons over broken data is a reliable way to produce a pilot that works in the demo and dies in the quarter, which is why AI pilots stall before they ship.
What do AI agents look like in enforcement and monitoring?
Enforcement systems check whether a rule was followed and act when it was not. They are deterministic by design, because the value comes from the check being identical every single time.
Automated Calendar SOP Enforcement for a US Investment Fund brought manual review to zero. Somebody used to read the calendar against the policy. Now the policy reads the calendar. There is no judgment in that job once the rule is written down, and introducing one would mean occasionally getting the rule wrong, which is the only outcome nobody would accept.
Operations Infrastructure for a Professional Services Business saves 48 hours a week. That is more than a full-time role, recovered from handoffs between systems that used to require a person to carry information across. Mostly deterministic, with models used narrowly where text needed writing.
Both are the least glamorous examples on this page, and the 48 hours a week is the largest recovery of human time anywhere in the set. That is not a coincidence, and it is the argument against buying the word “agent”.
What do the working examples have in common?
Twelve systems here. Fifteen are written up in total across our case studies, and that fuller set adds healthcare and insurance operations to the construction, recruiting, real estate, field service and financial work covered above.
Three things, and every system on this page has all three. The ones that fail are usually missing one of them, and you can check for all three before anyone writes code.
They replace a step somebody was measurably doing by hand, or close a gap somebody was measurably exposed to. Eight hours of screening. Twenty minutes per quote. Up to three months before a mismatched invoice was caught. The step existed, it had a duration, and someone could tell you what that duration was before the project started. When nobody can state the baseline, there is no way to prove value afterward, and the project joins the large population of pilots that run perfectly while showing nothing.
They write into a system of record, not into a document. Output that lands in a PDF nobody opens has no effect on the business. Output that lands in the CRM, the applicant tracking system, the ERP or the job board changes what the next person sees when they open the tool they already have open. The clearest tell of a project heading nowhere is that nobody can name the field the answer gets written into.
They have a review point wherever a wrong answer is expensive. Recruiters approve candidates. Estimators approve prices. Sales sees the conversation. This is not caution for its own sake, it is what makes a probabilistic component acceptable in a business process at all, and it is a large part of what production-grade automation means in practice.
When is an AI agent the wrong choice?
Five situations, and they cover most of what we decline to build as agents.
The rule is deterministic. If the policy is written down and has no edge cases requiring judgment, a model adds latency, adds cost per run, and adds the possibility of getting a settled question wrong. The calendar enforcement system is the clean example: the SOP existed as text, so it became code. An agent version would have been more impressive in a demo and worse at the job.
The data is too inconsistent to reason over. Contradictory records do not produce uncertainty in a language model, they produce confident wrong answers, which is the most expensive failure mode there is because it looks like success. The construction integration and the CRM sync are both what this problem looks like when it gets solved properly. Fix what the systems disagree about first. The agent conversation is cheaper afterward and sometimes turns out to be unnecessary.
The process changes faster than the maintenance can keep up. An agent carries a maintenance surface that automation does not: prompts, tool contracts, an evaluation set that has to be kept honest. If the underlying process is rewritten every quarter, that surface rots on the same schedule, and the annual maintenance can quietly exceed the annual saving. Count the hours saved per year against the hours the system will need, before, not after.
The step is not worth a system at all. This is the one most agencies will not tell you, and our own site leads with it: find the $2,000 problem before selling anybody a $20,000 one. If an off-the-shelf tool already does it, or an operations person could build it in an afternoon now that the tooling has moved, the correct recommendation is that nobody gets hired.
A wrong answer is unrecoverable and no review step is possible. This is the hard limit. One more from our case studies, outside the twelve above: Rebuilding and Securing an Online Prescription Business involved 30,000 records secured, and there is no AI agent anywhere in it. The problem was integrity and access control on regulated data, where a plausible-sounding wrong answer is not a customer service issue. That engagement was engineering, not agents, and the honest version of AI agent development includes saying so when it applies.
How should you decide which one you need?
Run the test on a proposal, not just on an idea. If a vendor pitched you three AI agents last quarter, take their descriptions and ask of each one whether the system chooses its next action or follows a path someone drew. Two of the three will usually turn out to be automation, which is fine, and which should be reflected in what you are being charged.
Then take the step you want automated and ask whether you could write down every decision it contains in advance. If yes, you want automation, and you should pay automation prices and expect automation reliability. If the next action genuinely depends on what the last one found, in ways nobody can enumerate, you want an agent and you should budget for the guardrails and the review point that make it safe to run.
Then ask the harder question, which is what the output changes. Name the field it writes into and the person whose next move is different because of it. Systems that can answer that tend to still be running a year later. Systems that cannot tend to be working perfectly and helping nobody, which is a more common outcome than failure and much harder to notice.
Line up the numbers from the twelve above and a pattern falls out that we left implicit until now. The 48 hours a week saved on operations infrastructure, the invoice-mismatch detection time cut from up to 3 months to within 2 hours, the zero discrepancies held for 6 months on the CRM sync, the drop from $43 to $12 in cost per lead, the 8 hours to 5 minutes on resume screening, the 20 minutes to 3 minutes on construction estimating: every one of those numbers belongs to a system with little or no AI in it.
The two genuine agents on this page, the multichannel sales agent and the field-service data agent, both ship without a published outcome metric, and that is not an oversight. A system that chooses its own path at runtime does not hand you one clean before-and-after the way a fixed pipeline does, so its return shows up as retained accounts and answered questions rather than as a number you can put in a headline.
The more agentic the system, the harder its return is to measure. That is not a claim about the market in general. It is what our own twelve examples show when you put them side by side.
If you are still deciding whether to build any of this internally or bring in help, that ground is covered in what an AI automation agency actually does. And if the honest answer after the arithmetic is that the step costs less than the system that would replace it, the right decision is to keep doing it by hand and spend the budget on the one that does not.
All outcome figures are from our own published case studies, verified August 2026. Clients are described by role and industry rather than named. Gartner’s agent-washing estimate is from its June 2025 press release predicting that over 40% of agentic AI projects will be canceled by the end of 2027.