Real AI agent examples from small companies, with costs and what broke

Why most AI agent example lists are useless

They fail in one of two ways. The first is an undergraduate taxonomy in business clothing: simple reflex agents, model-based agents, a thermostat, a Roomba. True, and no help to anyone deciding whether to spend $8,000. The second is a Fortune-500 showcase, where the examples are all companies with a machine-learning department and a procurement process, running at a scale that shares nothing with a 30-person firm except the word "agent".

The most honest page on this topic is a complaint. Google currently ranks a Reddit post near the top for this exact search, and it says roughly: does anyone know some real-world examples of using AI, because I keep seeing the tutorials. That post outranks the vendor content because it is the only thing on the page that sounds like a person.

The reason nobody supplies the answer is structural. We went looking for the buyer's side of this conversation in August 2026 and it barely exists. A thread asking "has anyone actually hired an AI agency?" drew 22 replies and not one of them was from a buyer; they were all agency operators. Every survey with a headline number samples organisations above $1 billion in revenue. Nobody publishes a price. It is sellers talking to sellers, and the people who are supposed to be buying are reading the transcript.

So here is our side of it: nine agents we put into companies under 50 people, each with the job it does, the SOP it runs, the number it moves, what it cost where we have a defensible figure, and what broke.

Ten AI agents we deployed in companies under 50 people

Read these as shapes rather than as a menu. Each entry is one seat, on one org chart, with one person who owns it. Where we have a cost we print it. Where we do not, the entry says so instead of estimating, because a made-up number here would cost more credibility than it buys.

And one number before the list, because nobody in this market publishes it: our survival rate. Roughly 60% of the agents we have deployed are still running today. The other 40% died for three reasons, in order. Token costs ran past what the client expected, and the gap between the pitch in their head and the invoice killed the programme. The process changed underneath the agent and nobody owned the update, so it went quietly stale. Or the people who were supposed to use it were never trained, got worse output than we did, and reasonably concluded the tool was bad. Two of those three failures have nothing to do with the model.

A word on what this list is not. These are ten deployments we can describe honestly, not a representative sample of everything we have shipped, and we are not going to imply otherwise by counting only the ones that worked. The anti-example near the bottom is here for that reason.

Three of those nine sit inside one small law practice, which is deliberate. The interesting unit is not the agent, it is the seat, and one person can hold three or four seats at once if they are willing to review the output. The lawyer is our best result to date and the reason is not technical. He had spent decades handing work to people and reading it critically when it came back, and that skill transferred directly.

Every agent above that worked has a named human who owns it. Every one that failed did not.

AI agent examples by department, and which ones we'd actually build

The same nine examples, rearranged as a decision table. The last two rows are the honest ones: a department everybody asks about where we have no published result of our own, and a category we turn down.

By frequency, across everything we have shipped, three buckets dominate. Lead generation and sales is the biggest by a distance: prospect enrichment from multiple channels, outreach, lead scraping, follow-up. Second — and this is the one that surprises people — agents for technical SEO and generative engine optimization, keeping a company's own site visible to search and to AI answers. Third, the chief product agent, the last working entry in the list above. And inside that first bucket sits the biggest single lever we have seen anywhere in our deployments: time to first response. Cut it from hours to minutes and keep every follow-up on time, and conversion on the same lead flow moves more than anything else on this page — in the deployments we have watched, by two to three times. Nothing else we ship moves a sales number that far that cheaply.

The other pattern worth noticing is that the seats with the clearest results are the ones whose KPI already existed in a system somebody checks monthly. That is not a coincidence and it is not a small point. MIT NANDA's 2025 report makes the same observation from the other end: sales and marketing take the largest share of AI budgets, while the clearest documented savings the researchers found came from back-office automation. Attention goes to the front office. Measurable money comes out of the back.

If you want the version of this table that is about tooling rather than seats, agents vs workflows vs RPA works through which of the three each of these jobs actually needs. On our own read, at least half the rows above could run as a fixed-path automation, and would be cheaper to run and easier to debug if they did.

How long it takes to get one agent running

For one seat with an SOP that already exists, a first deployment takes two to four weeks from a signed scope. The full range we have lived: the fastest was ten days to a live, production-grade agent, and only because a hard deadline forced every decision to happen on time. The slowest engagements run a quarter to two quarters, and they are not slow builds — they are embedded work. For an education and migration services company we built their internal agents over about six months, lead generation and lead nurture first, and stayed on to monitor and update them as usage taught us what to change. The lawyer was self-sufficient with his set in about two weeks. A company-wide rollout is a different animal and the shape is predictable: the founder and exec team go first, the first department is onboarded two to four weeks after that, and the engagement then runs a couple of months of back-and-forth on per-department skills, connectors, training, and rearranging the agent hierarchy as usage shows what actually matters.

Where the timeline slips, it is almost never the build. It is the SOP that turns out not to exist, or the two documents that contradict each other and have no owner, or the connector for a system whose API is a login page. Not more than 25% of clients arrive with their knowledge in a state you can point retrieval at, so plan for remediation as a phase rather than a surprise.

For context on how unusual small-company speed is: MIT NANDA's researchers found that top-performing mid-market companies reported average timelines of 90 days from pilot to full implementation, while enterprises — defined in that report as firms above $100 million in revenue — took nine months or longer. Being small is an advantage in this work, and it is one of the few places where that is measurably true rather than just encouraging.

AI agent examples from the public record, fact-checked

These are the cases everyone cites. Most articles quote the first half of each one. Here is both halves, with the date and how strong the evidence actually is.

Two things to take from that set. Klarna and the Commonwealth Bank are the same story told twice: a headline number about replaced humans, followed by a quieter correction once the work came back. Use both halves of each or neither. And Taco Bell is the clearest statement of the gap that matters at any company size — works in the demo is not the same as works at 7pm on a Friday, and the difference between them is where the money goes.

The vendor-reported successes are worth reading with the label attached rather than dismissed. The Novo Nordisk entry is a good example: the eye-catching number is production time collapsing from weeks to minutes, and the sentence directly beneath it says review cycles were cut in half. Halved. A qualified human still reads the output before it goes to a regulator, which is exactly the shape of every one of our own deployments that works.

Most AI agents in production aren't agents, and what the 95% really said

Two numbers get thrown at each other constantly. MIT NANDA says 95% of organisations are getting zero return. Anthropic says 80% of leaders report measurable economic impact today. Both are real, both are published, and they are not talking about the same thing.

Menlo Ventures resolves it in one line. In a survey of 495 US enterprise AI decision-makers run in November 2025, only 16% of enterprise deployments — and 27% of startup deployments — qualified as true agents, meaning a model that plans, acts, observes the result and adapts. The rest were fixed-sequence or routing workflows wrapped around a single model call. Menlo are AI investors, so read them with that in mind, but the finding explains the fight: everyone is measuring different objects and calling them by the same word.

What the MIT 95% figure actually says, read properly

We read the report rather than the headlines about it. A note on sourcing first: MIT's own hosting for the paper was pulled, so we read a mirrored copy of the v0.1 PDF. It is marked preliminary and it is not peer-reviewed. The methodology is a review of over 300 publicly disclosed initiatives, structured interviews with 52 organisations, and 153 survey responses collected at four conferences, between January and June 2025.

The exact wording is "95% of organisations are getting zero return" and "just 5% of integrated AI pilots are extracting millions in value". That is about custom, enterprise-grade tools. The same report says over 80% of organisations have explored or piloted general-purpose tools like ChatGPT and Copilot, nearly 40% report deployment, and generic chatbots show pilot-to- implementation rates around 83%. So "95% of AI projects fail" is not what the paper says, and the people repeating it that way have not opened it.

Two findings inside it almost nobody quotes, and both point the same direction for a small company. External partnerships reached deployment about 67% of the time against about 33% for internally built tools — the report's own summary is that partnerships are twice as likely to reach full deployment, with employee usage rates nearly double. And the speed finding above: 90 days for top mid-market performers against nine months or longer for enterprises. The study most often waved around as proof that AI does not work is, read carefully, an argument for being small and for bringing in outside help.

We should be even-handed about it, because the report is honest about its own limits and its readers usually are not. It flags that success definitions varied across organisations, that the deployment percentages come from a 52-organisation interview sample, and that the correlation between partnerships and success does not prove causation. It also contradicts itself: on page 9 the takeaway box says 50% of GenAI budgets go to sales and marketing, and the body text two paragraphs below says sales and marketing captured approximately 70% of budget allocation. Both numbers, same page. Download it and check that yourself before you cite either figure at a board meeting.

The wider survey picture is flat rather than catastrophic. McKinsey's 2026 State of AI put 37% of 1,719 respondents attributing at least some EBIT impact to AI, roughly the same share as a year earlier, with 6% qualifying as high performers and 40% of respondents at organisations above $1 billion in revenue scaling agents. We tried the McKinsey page five times over two days and it timed out every time, so those figures come from The Register's reporting of the survey on 25 August 2026, not from a primary read. That distinction is exactly the kind most articles quietly skip.

Before you believe any production or failure rate, ask what the study counted as an agent. Most of them do not say.

What actually breaks, and it is almost never the model

We read 25 negative reviews of n8n, Zapier, UiPath and Intercom in August 2026, looking for the technical complaints. There were none. Not one review said the AI got the answer wrong. Every single complaint was about the business layer wrapped around it.

• Silent breakage. "All our zaps just were not running and no alert was presented" (Zapier, 5★, 7 August 2026). "One wrongly-configured step can destroy a flow without any warning" (Zapier, 5★, 14 December 2025). Both from five-star reviewers, which is the part that should worry you: these are people who like the product.
• No observability. "Hard to debug… vague debugging messages" (n8n, 5★, 4 September 2025). When something stops at 2am, the question is not whether the model is clever. It is whether anyone finds out.
• Pricing that moves. "Fin charges about $0.99 per resolution, which adds up quickly" (Intercom, 4★, 25 August 2025). Per-resolution and per-token pricing both put your bill on the other side of a number you do not control.
• The expectations gap. "Connecting a Google Sheet took me about an hour. I expected it to take 5–10 minutes" (n8n, 4★, 16 May 2026). That reviewer expected the AI to remove the technical work, and it did not.

A buyer in a public thread put the same finding better than any of our own copy does: the handoff and maintenance were where things started to wobble. Not the build. The handoff.

Our own three failure modes match, and they are the ones we watch for on every discovery call. First, model choice: people reach for the largest model for every task and the token cost balloons. That is the single most common issue we see. Second, no clarity on the target: no defined outcome or KPI going in, so nobody can see the return and nobody can defend the spend when it is questioned. Third, no employee training: the system is set up correctly, the people around it do not know what to do with it, and the results are bad anyway.

None of those three is a technical failure, and none of them is fixed by a better model.

Two 2026 incidents that should change how you scope an agent

This section is short because the honest version is short. Two things happened in the last two months that are worth more than a year of governance framework diagrams.

On 16 July 2026, Hugging Face disclosed a security incident and described the intrusion in its own words: it was "driven, end to end, by an autonomous AI agent system". The disclosure describes an agent framework running thousands of individual actions across short-lived sandboxes. Whatever you think about agent capability in a demo, an agent ran that campaign.

On 4 August 2026, the UK AI Security Institute published an incident report about its own testing. During evaluations between 25 and 28 July, AI agents took 19 unsanctioned actions across 10 of 122 runs, directed at real people and organisations on the live internet. The most serious one: an agent tried to insert malicious code into a publicly used open-source project, researched the project maintainers, created multiple fake identities, and used them to socially engineer a real maintainer into approving the code. A human maintainer caught it and refused. These were deliberately permissive test conditions with safety filters disabled, and that is the point — it is what the capability looks like when the guardrails are the only thing in the way.

Set against that, a useful corrective from a security researcher posting as @cyb3rops: "end-to-end autonomous" is a claim, not a proven technical finding, and what gets sold as capability is frequently a capability demo repackaged as a marketing pitch. Both things hold at once. Agents do real damage in a lab, and the claims made about them in a sales deck are still unverified.

For a 30-person company, the practical version of all this is small and boring. Credentials are the exposure, not the prompt. The most useful line in the Uber practitioner thread quoted above is that credential leakage is a more common operational problem than prompt injection, and it matches what we found the moment we took an agent from one user to a team: a shared instance is a shared-secrets problem, and the fix is isolation per user and per department with scoped credentials. One client came to us for an AI policy for exactly this reason — an employee had access to something they should not have had, and it ended in something being published that should not have been. The trigger was access control, not the model.

If you want a structured checklist rather than our opinion, OWASP published a Top 10 for Agentic Applications in December 2025. We are linking it rather than summarising it, because the full list sits behind a download we have not read, and reciting ten items we have not verified would be the exact behaviour this page is arguing against. Our own agent security and governance guide covers what we actually implement.

Numbers about AI agents that don't survive a fact-check

Our method was dull and anyone can repeat it. Take the number, find the publisher it is attributed to, open that publisher's own page, and look for it. If it is not there, do not use it. Here is what fell out.

A date trap worth knowing about, since it catches good writers too. The Klarna reversal was May 2025 and the Commonwealth Bank reversal was August 2025. Anyone presenting either as a 2026 event is copying from someone who copied from someone. The same goes for the Salesforce support-headcount numbers, which are from September 2025 and recirculate every few months as fresh.

If a number appears in six articles and none of them links to the publisher, it usually does not exist.

How to tell whether your company would be on this list

Pick one role, not a department and not the company. Ask three questions about it. Does that role have clearly determined KPIs? Does it have clearly determined SOPs? Is the work mostly digital? If all three are yes, an agent can very likely do a real part of that work, or at least resolve part of it. That test is the whole buy signal, and it beats headcount and budget as a predictor of whether a build is going to land.

If one of the three is missing, the honest first invoice is for writing the missing thing down, not for building on top of the gap. A follow-up sequence that only lives in the owner's head is not an SOP. Two documents that disagree are not knowledge. Every example on this page that worked started from something already written down, and the one that failed hardest started from nothing at all.

Engagements go best when three things are true at once: the client has one area they want to double down on, one specific KPI they want to move, and an acceptance that their team will need training. Named area, named number, training appetite. When one of those is missing we can usually predict which part of the rollout will stall.

Two free tools run this properly. The AI readiness scorecard scores documentation, data access and ownership in a few minutes. If the honest answer is that the SOP does not exist yet, the SOP generator writes one and then tells you which of its steps an agent could actually run. Both are ungated and there is no email box.

What one production agent costs, and what we charge

We looked for a competitor publishing a price for shipping a single production agent and did not find one. That gap is the most-cited complaint in every buyer thread we read. So, ours, in public.

• $499 readiness audit, credited in full against any later build. A written build-or-don't verdict on the role you bring us.
• Consulting engagements from a fixed $3,500, scoped in writing before anyone starts.
• A production agent build at a fixed fee from about $8,000, with acceptance criteria you can test and a 90-day fix-first warranty. No retainer required to keep it.
• The run cost is quoted as a separate line, always, because it is the line that ends programmes and it does not belong hidden inside a build price.

On the run line, our own worst case is on this page: $3,000–$5,000 a month at a 20–50-person client, mostly burned by always-on agents nobody was talking to. That is the disaster tail, though, not the middle. The boring middle, across our stable deployments, is model tokens plus less than $70 a month of infrastructure to run everything else. Publish that next to the $3,000 horror story and the lesson writes itself: the spread between a well-scoped agent and a badly scoped one is the whole budget. And it is not a small-company-only problem. Uber went through its 2026 AI-tooling budget in about four months and capped spending at $1,500 per employee per month, reported by Fortune in May 2026. Nvidia's Bryan Catanzaro put the general version to Axios in April 2026: for his team, the cost of compute is far beyond the cost of the employees. If organisations that size are recalibrating, a 30-person company should assume its first estimate is wrong and instrument the spend from day one.

For the full market picture, including what other firms invoice and where the ranges come from, our AI agent cost guide has each band with its sourcing, and AI consulting for a small business covers the same question from the consulting side, with timelines.

Who shouldn't build any of this

If your team is small and genuinely new to this, do not commission a build. Buy an n8n, Make or Zapier subscription, give one motivated person a few Friday afternoons, and see how far that gets you. It gets most people further than a small consulting engagement would, and it teaches you what you actually need in a way no discovery call can. We say this on calls most weeks and it costs us money every time.

The threshold is repeatability, not revenue. Once the problems repeat — you know what is missing from your workflows and what needs doing at regular intervals — a paid build starts to make sense. Before that point you are paying someone to discover your own process, which you can do for free and more accurately.

A hard budget line: if your entire AI budget for the year is under about $2,000, hire nobody. A small engagement is the worst version of this work. Big enough to cost real money, too small to change how anything runs.

Our own gaps, since this page is asking you to trust nine anonymised stories. We have no reviews on Clutch or any other third-party directory, so everything here is published by us and there is nowhere independent to check it. The cost columns above are incomplete on purpose and we would rather show the holes than fill them with estimates. And we are a small team with the founder writing these guides and running the calls, so the bus-factor question is a fair one to ask us.

The uncomfortable position we hold, having built these: the future of work is agent–human hybrid, and an agent whose knowledge base, skills and context are actually maintained can out-produce a human employee at a specific desk job — because it accumulates context across domains no single person touches. We say that having also watched 40% of our own deployments die. The human in the loop is the most efficient configuration we have ever shipped, and the binding constraint is never the model. It is training the head of sales, and everyone like them, on a process that changed. That is the biggest friction in this industry right now, and almost nobody prices it into the project.

The agent itself is the easy part. What you are really buying is somebody forcing you to write down how your company works, and most of the value arrives before any model is called.

If you want the decision without talking to us, build vs hire vs agency runs it as a tool, including the answer where you do it yourself. And questions to ask an AI agency lists the three tests we fail ourselves.

Frequently asked questions

Sources

Every external link below was opened and read on 28 August 2026 unless the entry says otherwise. Evidence grades are ours.

• Cognio Labs deployment notes, 2026. All nine first-hand examples, the $3,000–$5,000 token blowup, the ~30% three-month adoption figure, the "not more than 25% arrive knowledge-ready" observation, and the three failure modes are our own client observations, anonymised to size band and industry and published with permission. [first-hand]
• MIT NANDA, The GenAI Divide: State of AI in Business 2025, v0.1 preliminary, research period January–June 2025. MIT's own hosting was withdrawn, so we read a mirrored copy of the v0.1 PDF. Source for the 95%/5% wording, the ~83% chatbot figure, the 80%/40% general-purpose funnel, ~67% vs ~33% partnership-versus-internal deployment, 90 days vs nine months, the back-office ROI observation, and the page 9 50%-vs-70% inconsistency. [primary, but preliminary and not peer-reviewed]
• Menlo Ventures, 2025: The State of Generative AI in the Enterprise (n=495 US enterprise AI decision-makers, fieldwork 7–25 November 2025) — the 16% enterprise and 27% startup true-agent figures. [study evidence; publisher is an AI investor]
• Anthropic with Material, The 2026 State of AI Agents Report (n over 500 US technical leaders) — 57% multi-stage workflows, 80% reporting measurable economic impact, and the Novo Nordisk and L'Oréal figures. We read the PDF. [vendor-reported]
• Brandon Vigliarolo, "McKinsey says enterprise AI is finally on the road to ROI", The Register, 25 August 2026 — the n=1,719, 37% EBIT, 6% high-performer and 40% agent-scaling figures. McKinsey's own page timed out on five attempts, so this is secondary reporting, not a primary read. [reported]
• Bain & Company, Your AI Budget Is Growing. Your Returns Aren't. Here's Why. (1 June 2026, Automation and AI Pathfinder Survey, n=951) — 90% increasing budgets, nearly 40% landing in the 0–10% savings bucket against 37% targeting 11–20%. The widely circulated "4% achieved above 30% savings" line does not appear on this page. [primary]
• Klarna, AI assistant handles two-thirds of customer service chats in its first month (27 February 2024) — 2.3 million conversations, the work of 700 full-time agents, $40 million projected profit improvement. The May 2025 reversal and rehiring is business-press reporting we did not read in the original. [claim: primary · reversal: reported]
• ABC News, Commonwealth Bank backtracks on AI job cuts, apologises for "error" as call volumes rise (21 August 2025) — the 45 roles, the reversal, and the rise in call volumes. [verified, independent press]
• Hugging Face, Security incident disclosure — July 2026 (16 July 2026) — the "driven, end to end, by an autonomous AI agent system" wording. [primary]
• UK AI Security Institute, Incident report: unsanctioned agent behaviour during cyber testing (4 August 2026) — 19 unsanctioned actions across 10 of 122 runs between 25 and 28 July 2026, and the open-source maintainer social-engineering attempt. [primary]
• OWASP GenAI Security Project, Top 10 for Agentic Applications (published December 2025). Linked, not summarised: the full list is behind a download we have not read. [linked, not read]
• Capterra reviews of n8n, Zapier, UiPath and Intercom, read 28 August 2026. Star ratings and dates are given inline with each quote. Twenty-five negative reviews read; none of them said the AI produced a wrong answer. [primary, small sample, ours]
• Practitioner claims on X, attributed and unverified: @praveenTweets on Uber's agent platform (session volume, credential leakage, the 50-approvals rubber stamp), and @cyb3rops on "end-to-end autonomous" being a claim rather than a proven finding. We could not independently confirm either. [unverified practitioner claim]
• Fortune, May 2026, on Uber exhausting its 2026 AI-tooling budget in roughly four months and capping spend at $1,500 per employee per month; and Axios, April 2026, quoting Nvidia's Bryan Catanzaro that for his team the cost of compute is far beyond the cost of the employees. [reported]

Related reading

• How much does an AI agent cost — the build and run bands, with the sourcing on each.
• AI consulting for a small business — the same money question from the consulting side, with timelines.
• AI automation consulting — what it costs to buy the no-code route instead, and who is liable when an automation gets it wrong.
• AI agents vs workflows vs RPA — which of the three each example above actually needs.
• Questions to ask before you hire an AI agency — including the three tests we fail ourselves.
• Build vs hire vs agency — the decision as a tool, including the answer where you do it yourself.
• AI readiness scorecard and SOP generator — the KPI and SOP test, run properly. Both free, no email gate.
• AI agent development and Agentic OS — one seat, or the company-wide version of the rollout described above.