What AI Agents Can Actually Automate in a Business Right Now

Almost every demo of an AI agent shows it doing something impressive and unrepeatable. Almost every agent that earns its keep in a real business does something boring, every day, that nobody wanted to do.

This post is about the boring ones: which jobs agents genuinely handle today, which they still cannot, what each costs to build in India, and how to choose a first project that pays for itself. It is written from shipping them for clients, not from watching keynotes.

The one-sentence version

Agents are good where work is high-volume, rule-shaped at the edges, and tolerant of a human check at the end. They are bad where being wrong is expensive and nobody is checking.

Everything below is an elaboration of that sentence.

What they are genuinely good at

Chasing

Documents, signatures, missing fields, overdue responses, unpaid invoices. Someone has to notice what is outstanding, contact the right person, and keep contacting them until it arrives.

This is almost always the first thing worth automating, for a reason that has nothing to do with the technology: the cost of the work being late is usually far larger than the cost of the work itself. An onboarding that stalls nine days waiting for one KYC document has not consumed nine days of labour — it has consumed nine days of revenue.

Agents are tireless, they do not feel awkward sending a fourth reminder, and they never forget. That combination is worth more than it sounds.

Routing and qualification

Reading an inbound enquiry, deciding what it is, and sending it to the right person with the right context attached.

The distinction that matters: not "is this a good lead" — that is a judgement call with money attached — but "this is an enquiry about service X, in region Y, with budget signal Z, so it belongs in this queue with these three facts summarised at the top". The agent does the reading and the summarising. A human still decides what to do.

Done properly this removes the single most common source of lost leads in small businesses, which is not bad follow-up but slow first response.

Extraction

Pulling structured fields out of unstructured input: invoices, application forms, email threads, uploaded identity documents, supplier PDFs that change format every quarter.

This used to be OCR plus a brittle template that broke whenever a vendor moved a column. It now mostly works on messy real-world input, and it is the single highest-leverage capability available to any operations team drowning in attachments. If your team retypes numbers from PDFs into a system, that is the project.

Reconciliation

Two systems disagree; something has to notice and either fix it or raise it.

Agents are good at the noticing, which is precisely the part humans are worst at, because reconciliation done well is invisible and done badly is catastrophic. Nobody gets praised for checking that the CRM and the invoicing system still agree, so nobody checks, until an audit.

Drafting

First-pass replies, summaries, file notes, status updates, meeting follow-ups.

The value is not that the draft is good. It is that a human edits instead of starting from a blank page, which is roughly a 4x difference in how long the task takes and a much larger difference in whether it gets done at all.

What they are still bad at

Worth being blunt, because this is where commissioned projects go wrong and budgets get wasted.

Judgement with consequences. Anything where being wrong costs money or trust: pricing, credit decisions, hiring, clinical, legal, anything a regulator looks at. An agent can assemble the decision — gather the facts, summarise the history, flag the risks. It should not make it.

Genuinely novel exceptions. Agents handle the exceptions you anticipated. The ones nobody described get handled confidently and wrongly, which is worse than not being handled at all, because nobody notices. Every agent needs an explicit route for "I am not sure", and most demos do not have one. Ask about it specifically.

Processes nobody can describe. If three people on the team would describe the workflow differently, no agent will fix that. You will automate one person's version and the other two will quietly route around it, and you will have paid for a system that makes the inconsistency permanent.

Anything needing real accountability. When a client, a regulator or a court asks who decided, "the system did" is not an answer anyone accepts. Keep a human in the loop wherever that question could plausibly be asked.

Work that is actually rare. Automating something that happens four times a year is a hobby, not an investment. Volume is what makes this arithmetic work.

The mental model that helps

The useful frame is not "replace a role". It is take the waiting out of a process.

Pick any workflow and mark every point where it stalls because a person has to notice something before it can move. Those stalls are what agents remove. The work itself frequently still needs a human; the waiting almost never does.

This reframing matters commercially, because it changes what you measure. You are not trying to reduce headcount — you are trying to reduce elapsed time, and elapsed time is usually where the money is.

For Aieraa's admissions platform that meant a fully automated flow across five-plus user roles. Not because admissions decisions were automated — they were not — but because nothing sat waiting for someone to check whether it could move to the next stage. For Lynk Labs, a diagnostics business, the same logic brought report turnaround under 24 hours across more than 50,000 customers: the lab work is human and always will be, but the collection scheduling, the chasing and the dispatch around it are not.

What it costs in India

Priced as automation work, because that is what this is:

Scope	Range	Typical shape
Starter automations	₹25,000 – ₹60,000	One flow, 1–2 workflows, basic notifications, 1–2 weeks
Growth automation stack	₹75,000 – ₹2,00,000	Cross-channel journeys, lead scoring, sales alerts, dashboard, 3–6 weeks
Custom CRM + automation	₹2,00,000 – ₹10,00,000+	Role-based automation, document triggers, API integrations, 6–16 weeks

Full ranges sit on the marketing automation pricing page, and the deeper systems overlap with custom CRM work.

Below about ₹50,000 you are buying isolated setup tasks rather than a dependable system. That is sometimes exactly the right purchase — automating one painful handoff is a perfectly good project — but it is worth knowing which of the two you are buying before you are disappointed.

There is also a running cost people forget: model usage. For the volumes most Indian SMBs deal with it is small, often a few thousand rupees a month, but it is not zero and it scales with success.

How to pick the first one

Start with the thing people complain about weekly. Not the most impressive candidate — the most repeated one. In our experience it is almost always chasing something.

Then apply four tests before committing:

1. Does it happen at least weekly? If not, pick something else.
2. Can you describe the steps without arguing? If the team disagrees, fix the process first — that is free and often sufficient.
3. What is the cost of it being done wrong once? If the answer is serious, keep a human approval in the flow.
4. Is there a clean "I don't know" path? If the agent cannot escalate, it will guess.

Automate that one flow end to end, including the escalation path, and run it for a month before adding anything else. Agents compound badly when stacked on an unproven foundation; one working flow will teach you more about your own process than any roadmap.

What a sensible first project looks like

A concrete shape we build often:

• An enquiry arrives by form, email or WhatsApp.
• The agent reads it, classifies the service and region, and extracts budget and timeline signals where present.
• It writes a three-line summary and attaches it to a new CRM record.
• It routes to the right queue and notifies the owner.
• Anything it cannot classify confidently goes to a human queue, flagged as unclassified rather than guessed.
• Twenty-four hours later, if nobody has touched the record, it nudges.

That is a few weeks of work at the lower end of the table above. It does not replace anyone. It typically halves first-response time, which is the number that actually correlates with won business.

Questions to ask whoever builds it

• What happens when the agent is not sure? Show me that path.
• What does it cost to run per month at our volume?
• Where is the human approval, and can we move it later?
• How do we see what it did and why? Is there a log we can audit?
• What breaks if the model provider changes or has an outage?

The fourth one matters more than people expect. An agent you cannot audit is an agent you cannot defend when something goes wrong.

A reasonable expectation

Agents will not transform your business this quarter. What they will do, if pointed at the right workflow, is remove a category of delay that you have probably stopped noticing because it has always been there.

That is a smaller claim than the demos make and a much more reliable one.

If you want a view on which of your workflows is the right first candidate, tell us how the work moves today — including where it waits. The answer is usually smaller, cheaper and duller than people expect, which is exactly why it works.