Loading…
Loading…
AI automation from a founder who builds it: where it pays back, where it fails, how to pick a first use case, keep people in control and start in 90 days.
AI automation earns its keep on judgement, not on steps. If a task can be written as rules, plain automation is cheaper and more reliable.
The best first use cases are high volume, judgement light decisions on messy inputs, with a person ready to take the cases the AI is unsure about.
Most failures are not model failures. They come from no owner, poor data, a broken process, no baseline, or running costs nobody counted.
Anything that cannot be undone should ask a person first. Confirmation, a full audit log and an easy override are what make AI automation safe to scale.
Measure outcomes against a baseline taken before launch, and count the human review time the automation still needs.
Ninety days is enough to prove one use case properly. It is not enough to automate a department, and nobody should promise that.
Search for AI automation and the first page belongs to the companies that sell it: Zapier, UiPath, Automation Anywhere, Pega, Oracle, AWS and Cisco, plus lists of AI automation tools. Most explain the idea well.
Few say where it goes wrong: that many use cases on your list need no AI at all, or that a model's running cost can quietly outgrow the saving it was bought for.
I write as someone who builds AI automation, for clients and in our own product. This is an opinionated view of where it pays back and where it fails. Our guide to business process automation with AI covers process design and the ROI arithmetic in depth.
AI automation is the use of artificial intelligence to handle the judgement inside a task that software runs on its own: reading documents, understanding language, recognising images and choosing between options. Traditional automation follows fixed rules written in advance. AI automation copes with varied, messy inputs, and passes the cases it cannot resolve to a person.
A simple test tells the two apart. If you could write the decision as an if this, then that rule and trust it, you need automation. If a capable new employee would need to read, look or weigh things up, you need AI inside the automation.
What is AI in automation, then? It is the part that replaces a human glance. An invoice arrives as a scanned PDF in one of dozens of layouts. A rule cannot read it reliably. A model can pull out the supplier, amount and due date, and ordinary automation does the rest.
You will also meet the term intelligent automation. IBM describes it as AI acting as the decision engine, combined with business process management and robotic process automation. It is the same idea from a different starting point.
Automation does what it is told. AI decides what to do when the instructions run out.
They also fail differently. Rule based automation breaks loudly: a missing field stops the run, and someone notices. AI fails quietly, returning a confident wrong answer while the process carries on. Confidence thresholds, monitoring and human review all exist to manage that difference. With automation you ask whether it ran. With AI you ask how often it was right, and what each wrong answer cost.
Four approaches now sit under the AI automation label. They differ in what they handle and how they go wrong.
Rules automation. Software follows steps and conditions written in advance: route this request, escalate that ticket. It is cheap, predictable and easy to audit. It fails when inputs vary in ways nobody wrote a rule for.
Robotic process automation (RPA). Software robots operate existing applications through their screens, the way a person would, bridging systems with no proper connection. It fails when a screen changes, and large robot estates carry heavy maintenance.
AI automation. A model makes a narrow judgement inside a fixed process: reading a document, classifying a request, spotting an event in an image. The process around it stays predictable. It fails when the model is confidently wrong and nobody checks.
AI agents. A language model plans its own steps towards a goal, chooses tools and adjusts. Anthropic draws the line clearly: workflows run models and tools through predefined code paths, while agents direct their own process and tool use. Agents fail as a confident series of wrong actions, taken quickly.
My view is firm: use the simplest approach that does the job. Rules first, a model where judgement is needed, an agent only where the path cannot be mapped in advance. Anthropic's engineering guidance agrees, advising teams to find the simplest solution possible, even if that means no agents at all. Our guides to agentic AI and to the Model Context Protocol cover agents in depth.
In our experience, AI automation pays back when four conditions hold together. Miss one and the case weakens. Miss two and it rarely survives.
High volume. The build has a fixed cost and each run costs little, so the return grows with every case. A task done twenty times a month seldom justifies a model.
Judgement light decisions. The best targets need a quick, narrow call that a trained person makes in seconds: which team should handle this request, is this document complete, is someone in a restricted zone. Work where experts themselves disagree is a poor first target.
Unstructured inputs. Free text, scanned documents, voice, images and video are where rules give up and AI earns its place. Clean, structured input rarely needs it.
A clear human fallback. Every AI system meets cases it cannot handle. The ones that pay back send those cases to a named person with full context, so an exception is resolved quickly instead of becoming a complaint.
Notice what is missing: the newest model. It matters far less than the shape of the work.
These are the AI automation examples I see justify themselves again and again.
Document intake. Invoices, claims and applications arrive in many formats. AI reads each one, extracts the fields and flags anything missing, and ordinary workflow does the rest.
Conversations in any language. Customers write the way they speak, often mixing languages. Project Sasha, the AI travel and lifestyle concierge we built for Kanoe in Spain, works this way: people talk or type in any language, on the web or on WhatsApp, and Sasha plans whole trips across 350,000 places in Vietnam.
Video that watches itself. Cameras produce far more footage than any team can watch. AI detects the events that matter, such as an intrusion, a fire or a missing safety helmet, and the system raises an alert or acts. This is what our own product, Visylix, is built to do.
Answers from your own knowledge. Staff ask in plain words and get answers from approved documents, with sources. Our guide to enterprise search with RAG explains how.
Most AI automation projects that disappoint do not fail because the model was weak. They fail for six ordinary reasons, and every one is avoidable.
1. No process owner. Nobody answers for the outcome after launch, so drift goes unnoticed and exceptions pile up. If you cannot name the owner, do not start.
2. Poor data. The model reads, or learns from, data that is incomplete or out of date. The demo used clean samples. Production never does.
3. Automating a broken process. AI makes a bad process run faster and harder to change. Fix the process first.
4. No measurement. Without a baseline taken before launch, nobody can say whether the automation helped, and the loudest opinion wins.
5. Irreversible actions without a human check. Sending money, deleting records, messaging customers or changing access should never rest on a model's judgement alone. One confident mistake can undo the whole business case.
6. Hidden running costs. Every model call is paid for, monitoring and review queues need people, and each model upgrade needs fresh checks. A case that counts only the build looks worse every quarter. Our analysis of why AI proofs of concept fail covers this cost picture in detail.
For a business, the five that matter most: AI cannot take accountability, know what it was never given, be right every time, decide what your organisation values, or repair a broken process.
Each one shapes the design. A named person answers for every automated decision. Missing context produces confident gaps, so data comes first. Every model has an error rate, so plan for the wrong answers. Trade offs between speed, cost, fairness and risk are leadership decisions that AI can only apply. And a broken process, automated, simply runs its faults faster.
The first use case decides whether the organisation trusts the second. This is the sequence I use.
1. List the candidates with real numbers: monthly volume, handling time, error rate and the systems involved. Set aside anything nobody can measure.
2. Remove anything rules can do. If a workflow would solve it, build the workflow. Save AI for the judgement.
3. Keep the high volume, judgement light tasks with messy inputs. These form your shortlist.
4. Name an owner for each. A candidate without an owner leaves the list.
5. Check the data. Can you gather a few hundred real examples where the right answer is known? If not, your first project is a data project.
6. Weigh the cost of a wrong answer. Prefer cases where a mistake is caught by the next step, not by a regulator or a customer.
7. Pick the one with a clear human fallback and a baseline you can measure today. That is your pilot.
Control is not a brake on AI automation. It is what lets you give it more work. Three mechanisms do most of it.
Confirmation before high impact actions. The system proposes, a person approves, and only then does it act. Reading and drafting can run freely. Paying, deleting, messaging customers and changing access should ask first.
A complete audit log. Every action records what was done, on whose authority, from what input, and when, so any incident can be reconstructed in minutes.
An easy override. A person can reject, reverse or pause any automated decision, and switch the automation off, without calling an engineer. The EU AI Act sets this bar for high risk AI systems: the people overseeing them must be able to understand the system's limits, stay alert to automation bias, override or reverse its output, and stop it safely. It is a sound standard even where the law does not apply.
Then bound the autonomy: a narrow set of tools, a cap on how often each automation can act, and limits it cannot exceed on its own. Widen them only as its record earns it.
My clearest lessons come from Visylix, our automated AI video surveillance platform. It is AI automation in its fullest form, running on the customer's own premises.
It watches: 22 AI analytics built in house watch the cameras for intrusion, fire and smoke, weapons, missing safety equipment, crowds, number plates and more. It decides: Sudarshan rules combine conditions, sequences in time and access control events, and an operator can run a rule on real history before arming it to see exactly what it would have done. It acts: alerts, incidents in the operator console, signed webhooks into other systems, talkdown through the camera speaker and siren relays.
Above it sits Radha, an AI copilot that runs fully on premise, with 111 tools that operators direct in plain words. Thirteen high impact actions always ask for confirmation, Radha has no delete tool, and every action is recorded in the audit log.
Three lessons carry over to any AI automation. Detection alone is not value: an alert nobody acts on is noise, so design the decide and act steps as carefully as the AI. Let people correct the system: in Visylix, marking a false alarm tightens thresholds automatically, never below the operator's limit. And make the safe path the default. Project Sasha adds a fourth: meet people in whatever language and channel they choose, not through a menu built for the system.
Honest measurement starts before launch. Record today's volume, time per case, error rate and cost per case.
Then measure outcomes, not activity. A count of documents processed says nothing about value. The time from invoice to payment, and the share paid without correction, says a great deal.
Count what remains: review queues, exceptions, corrections, model calls, monitoring and fresh checks after each upgrade. Compare against the simplest alternative, often a well built rule, not against doing nothing. And watch the share of AI decisions people overturn. A rising override rate is the earliest warning you will get.
The NIST AI Risk Management Framework, a voluntary framework from the US standards body, organises this work into four functions: govern, map, measure and manage. It makes a useful checklist for anyone putting AI into a live process.
Most organisations use all three over time, matched to each use case.
AI automation tools. Workflow platforms such as Zapier, Make, n8n and Microsoft Power Automate, and RPA platforms such as UiPath, Automation Anywhere and Blue Prism, now include AI steps. They suit common tasks between systems with ready connectors, owned by your own team. They strain when the logic grows, when volume makes pricing per task expensive, or when data cannot leave your network.
An AI automation agency. A specialist builds and often runs automations for you, usually on those same platforms. It suits standard use cases when you need speed and lack people in house. Ask who owns what is built, and what happens when the engagement ends.
Building your own. A custom system makes sense when the process is specific to your business, your data is the advantage, integration is the hard part, or the work must run on your own infrastructure. It costs more at the start and gives you full control.
Buy for common work, use an agency for speed on standard work, and build where the automation sets you apart. Our guide to custom AI versus off the shelf AI covers that decision in depth.
Ninety days is enough to put one AI automation into real use and know whether it pays.
Days 1 to 30: choose and baseline. Run the selection sequence above, name the owner, record the baseline and gather real examples. Write down what success means, as a number.
Days 31 to 60: build the smallest complete version. Connect the AI to the real systems, set a confidence threshold, design the handover to a person and add confirmation for anything that cannot be undone.
Days 61 to 90: run it alongside people, then let it lead under oversight. Start with the AI suggesting and people deciding. When agreement and the cost of errors meet your bar, let it act on routine cases and keep people on the rest.
At day 90, decide on evidence: widen it, fix it or stop it. Stopping a use case that does not pay back is a good result.
At Aptibit, we start with the problem, not the technology. We define the outcome in measurable terms before any AI work begins, review what data exists and whether AI is the right tool, and say plainly when simple rules will serve you better.
When the work needs action across your systems, we build AI agents connected to your CRM, ERP and databases, with confirmation before high impact actions, a full record of every action and handover to a person whenever the agent is unsure. A focused agent with a few tools typically takes 4 to 6 weeks, and it can run entirely on your own infrastructure.
When a pilot has stalled, we close the gap to everyday use: data preparation, integration, monitoring and adoption. A focused proof of concept takes 2 to 4 weeks, and a full product typically takes 3 to 6 months. You see working software every two weeks, with senior engineering at India cost levels, well below US and Western European rates. Code, models, data pipelines and documentation belong to you under an IP assignment signed before work begins.
AI automation earns its place on high volume work that needs a quick, narrow judgement on messy inputs: documents, free text, conversations in many languages, images and video. It shines where a person stands ready for the hard cases, where outcomes are measured against a baseline, and where confirmation, audit logs and easy override let you widen its authority as it earns trust.
It disappoints where rules would do the job, where nobody owns the outcome, where data is poor or the process is broken, and where success was never defined. It is risky wherever it can take irreversible actions without a person checking, and it loses money quietly when running costs, review queues and fresh checks after upgrades were left out of the business case.
AI automation is software that runs a task on its own and uses artificial intelligence for the parts that need judgement, such as reading a document, understanding a message, recognising what is in an image or choosing between options. Ordinary automation follows fixed rules. AI automation handles varied, messy inputs and passes the cases it cannot resolve to a person.
Automation carries out instructions written in advance, and AI makes judgements where no fixed instruction fits. Automation asks whether a step ran; AI asks whether a decision was right. They also fail differently. Rule based automation usually stops visibly when something is wrong, while AI can return a confident wrong answer, which is why AI automation needs thresholds, monitoring and human review.
Common examples include reading invoices and claims in any format and extracting the key fields, sorting incoming emails and tickets by intent and urgency, handling customer conversations in many languages, answering staff questions from approved documents, flagging suspect records before they reach a report, and video surveillance that detects intrusions, fires or missing safety equipment and raises an alert on its own.
For a business using AI automation, five limits matter most. AI cannot take accountability for a decision, know what it was never given, be right every time, decide what your organisation values, or repair a process that is broken. Good AI automation is designed around these limits, with a named owner, good data, a plan for errors, leadership decisions on trade offs and a sound process.
An AI automation agency designs, builds and often runs automations for a business, usually on workflow platforms such as Zapier, Make or n8n with AI steps added. An agency suits standard use cases when you need results quickly and lack the people in house. Before signing, ask who owns what is built, how it is monitored and what happens when the engagement ends.
The best tool depends on the work. Workflow platforms such as Zapier, Make, n8n and Microsoft Power Automate suit common tasks between popular applications. RPA platforms such as UiPath, Automation Anywhere and Blue Prism suit work across older systems without proper connections. A custom build suits processes specific to your business, or data that must stay on your own infrastructure. The use case matters more than the brand.
Very nearly. Intelligent automation is the term favoured by enterprise automation vendors. IBM describes it as AI acting as the decision engine, combined with business process management and robotic process automation. AI automation describes the same idea from the AI side: automation in which a model handles the judgement. The labels differ, but the principles of ownership, data, oversight and measurement are the same.
For a business, start with one high volume task that needs a quick judgement on messy inputs, name an owner, measure how it is done today and run a 90 day pilot with a person ready for the hard cases. For an individual, learn one workflow platform, then learn to call an AI model inside a workflow, and build small automations for real problems you know well.
Yes. Aptibit builds AI agents and AI automation connected to your existing systems, with confirmation before high impact actions, limits on what each agent can do and a full record of every action. A focused agent with a few tools typically takes 4 to 6 weeks, and systems can run entirely on your own infrastructure. We also run AI automation in our own product, Visylix, every day.