You are drowning in manual work.
Quiet software that handles the busywork.
- Workflow & task agents
- Data pipelines & sync
- Human-in-the-loop controls
What you walk away with
The dull work handled, a person still in charge of anything that matters, and a record of everything that ran.
The dull work, done quietly
The repetitive jobs that eat your team's day, handled quietly in the background, so your people spend their time on work that needs them.
A human checkpoint that matters
A clear human checkpoint on anything that matters: you approve, edit, or reject before it goes out.
A log of everything that ran
A log of everything the automation does, so nothing happens in the dark and you can see what ran and why.
Your apps talking to each other
Your different apps talking to each other, so the same information stops being typed in three times by three people.
Fewer tired mistakes
Fewer of the small, tired mistakes that creep in when a human repeats the same boring task for the hundredth time.
This will not pay for itself in saved salaries
It is the first thing most people selling this will tell you, and the numbers do not support it. Here is what we think is actually true.
A regular salaried employee in India earns about ₹20,702 a month (Periodic Labour Force Survey 2023-24, the national household survey, quoted in the Economic Survey 2024-25). The person doing your invoice entry also answers the phone, chases a supplier, handles the one order that arrived wrong, and can be told something new on a Tuesday. Software does none of that.
So automating one task saves a fraction of one person and does not remove the person. If somebody tells you this pays for itself in saved wages, ask them to name the wage. We would rather lose the argument here than have you find it out in month four.
What the case actually rests on
- Deadlines you cannot miss. A business over ₹10 crore turnover has thirty days from the invoice date to report it to the government portal, and after that the portal will not accept it at all (GSTN advisory, 27 March 2025, effective 1 April). That is not lost time. That is lost money, and it does not come back.
- Mistakes in the places you never look. You watch the bank balance and your customers ring you directly, so money errors and customer errors surface in days. Supplier reconciliations, compliance filings, and the error folder nobody opens do not. That is where a wrong number runs for three weeks.
- The work nobody has time to do at all. The weekly figures you keep asking for and keep not getting, because the person who would build them is re-typing invoices.
What we have written about automation
Free, ungated, and written to be useful whether or not you ever get in touch. No email required.
- “Human in the loop” isn’t a safeguard. It’s a person who stops looking.Give an agent the power to issue a refund and put an approval button in front of it. After four hundred correct refunds, the button is not an approval gate any more - it is a reflex. The safety has to come from what the agent cannot do, not from who says yes.Read 9 min
- For four of these six jobs, you should hire a person.Automation maths is local. Every payback figure you have been shown was worked out somewhere people are paid four or five times what they are paid here, and at Indian salaries the same sum comes out the other way. Here is the formula, and here is where it stops working.Read 10 min
- WhatsApp automation isn’t cheap. It’s cheap per message.Nobody multiplies. A price small enough stops being read as a price at all, which is exactly why it is quoted that way. Meta also changed the unit under everyone in July 2025 and raised India’s marketing rate twice since. Here is the sum the rate card does not do for you.Read 9 min
- Automation in Plain English: Software That Does the Boring PartsAutomation is not robots taking over. It is handing the dull, repeatable jobs to software so your people can do the work that needs a person.Read 4 min
- What Is an AI Agent, Really? Your Tireless Junior AssistantAn agent is not magic and it is not a person. It is software that can take a goal, work out the steps, and check its own work, with you holding the final yes.Read 4 min
- Keeping a Human in the Loop: Automation You Can TrustThe question with automation is not "can it run on its own?" It is "where does a person need to stay in control?" That line is the whole design.Read 4 min
My team spends half the day on the same dull jobs: copying details from emails into a system, chasing the same updates, retyping the same numbers. It is slow, it is boring, and that is exactly when mistakes creep in.
Every business has a layer of repetitive work that nobody enjoys and that does not need a human brain, just a human's time. Re-typing an order from an email. Moving a row from one app to another. Sending the same three follow-ups. Pulling the same weekly figures into the same report.
It is the kind of work too small to hire for and too constant to ignore, so it eats your best people's hours and burns them out. The fair worry is whether a machine will quietly make a mess you cannot see. So we build the answer in: a person stays in control of anything that matters.
What "human in the loop" means
It means a person stays in control of anything that matters. The software does the repetitive first pass, then hands the real decisions to you to approve, edit, or reject. You are never handing the keys to a machine and hoping.
How it works
Think of it as a quiet assistant that does the first pass, then hands the important decisions to a person. The work gets done, and you stay in charge.
For low-risk, repetitive steps, the software just does them and logs what it did. For anything that touches money, a customer, or a real decision, it prepares the work and a person presses the button. You take the dull eighty percent off your team's plate and keep the twenty percent that needs judgement.
How these things actually fail
Not one of these is exotic, and none of them is about the software being clever enough. We design against them by name.
Nobody reads the error messages
A broker’s own system sent 97 automated error emails naming the exact fault before the market opened. Nobody had designed them as alerts, so nobody read them; forty-five minutes of trading later the firm had lost $460 million (US Securities and Exchange Commission order 34-70694, 16 October 2013). Your version of this is a folder filling up with failure notices nobody opens. It is the most common thing we find and the cheapest to fix.
It ran on the tidy sample
The two hundred records used to prove it worked were the clean ones. Real data has blank fields, two spellings of the same supplier, dates in three formats, and the one branch that never filled in the address. Ninety-two per cent of practitioners in a peer-reviewed study of 53 people across India, Africa and the USA had hit a problem of this shape (Sambasivan and others, ACM CHI 2021).
Something upstream changed and nothing complained
Somebody adds a field or renames a column. The automation keeps running on the old map and keeps producing output that looks fine. An automation that STOPS is cheap. One that carries on quietly being wrong is not, and only a number computed a second way will catch it.
The exceptions cost more than the work did
The software takes the ordinary cases and hands back the awkward ones, out of order and stripped of context. Your people now do only the hard cases, without the rhythm of the easy ones. If nobody has priced the exception path before you sign, it has not been designed.
Nobody could overrule it
Research on automated matching in Indian welfare records found offices refusing to reverse a decision even when shown evidence it was wrong (Amnesty International, April 2024). The small version is a staff member saying “the system won’t let me.” When a person cannot correct the machine, an error stops being an error and becomes policy.
The person who understood it left
The rules and the reasons lived in one head. The thing still runs, but nobody can change it and nobody can tell whether it is right. We could not find a single credible measurement of how often this happens - every source was a vendor selling the remedy - so we name it and do not put a number on it.
Every number on this page, and where it came from
This is the most exaggerated corner of the internet. A number published by a company that sells this software is an advertisement, so we went to governments, regulators and peer-reviewed work instead, and we show you the source next to the figure rather than in a footnote.
30 days
A business over ₹10 crore turnover has thirty days from the invoice date to report it to the government portal. After that the portal will not accept it at all, and the credit behind it is gone.
SourcedGSTN advisory on the official e-invoice portal, dated 27 March 2025 and effective 1 April 2025, read on einvoice6.gst.gov.in itself rather than in a summary of it. Read the source ↗
₹20,702 a month
Average earnings of a regular salaried employee in India. This is the number that decides whether automating a clerical task pays, and it is the reason we will not tell you this one pays for itself in saved salaries.
SourcedPeriodic Labour Force Survey 2023-24, the national household survey, quoted in the Economic Survey 2024-25 (Ministry of Finance). Read the source ↗
47%
Share of newly created business records carrying at least one critical error, when managers checked their own last hundred. Only 3% of the resulting scores met the loosest acceptable standard.
SourcedNagle, Redman and Sammon in Harvard Business Review, 11 September 2017; 75 executives each checked their unit’s last 100 records against 10-15 critical fields. Self-collected and not a random sample of firms, which is why it is a rate and not a cost. Read the source ↗
88% against 10%
Spreadsheets found to contain at least one error when audited, against the roughly one-in-ten chance the people who built them thought they had of making one. The gap, not the percentage, is the finding.
SourcedRaymond Panko, University of Hawaii, pooling field audits of 113 spreadsheets since 1995. Academic, nothing sold; it pools small audits across decades and "at least one error" includes trivial ones. Read the source ↗
97 emails
Automated error messages a broker’s own system sent before the market opened, naming the exact fault. Nobody had designed them as alerts, so nobody read them. Forty-five minutes of trading later the firm had lost $460 million and was sold inside a year.
SourcedUS Securities and Exchange Commission order 34-70694, 16 October 2013, paragraphs 17 and 19. A regulator’s findings, not a retelling. Read the source ↗
92%
Practitioners who had hit at least one "data cascade" - a small problem introduced early that stays invisible until it surfaces much later and much more expensively. Nearly half had hit two in a single project.
SourcedSambasivan and others, peer-reviewed at ACM CHI 2021; 53 practitioners interviewed in India, East and West Africa and the USA. A count of people who reported it, not a rate across projects. Read the source ↗
3.3%
Share of jobs worldwide in the highest band of exposure to this kind of software. The exposed work is clerical, admin and payroll - exactly what gets automated first - and women hold nearly twice the exposure of men, 4.7% against 2.4%.
SourcedInternational Labour Organization with Poland’s NASK, Working Paper 140, May 2025. Inter-governmental, nothing sold; global with regional cuts, so it is not a measurement of your town. Read the source ↗
67%
Indians who expect a company to tell them when it is using this kind of software. Concern in India rose 14 points in a year while excitement rose 2 - the fastest-rising worry of the thirty countries surveyed.
SourcedIpsos AI Monitor 2025, fieldwork 21 March to 4 April 2025; India sample about 2,200 urban adults. Ipsos is a research firm and does not sell AI systems, but this is the urban general public, not business owners. Read the source ↗
Zero
Coloured words on this page. We checked every element in the page at three screen widths: no coloured text, no coloured borders, and the brand purple only on button fills. We counted the reference site our founder sent us the same way, with the same script, before we picked anything.
We counted itCounted by us on 28 August 2026 with scripts/aia-smoke.mjs, which re-runs against this page at 360, 768 and 1440 and fails if the number moves.
Four numbers we found, checked, and will not use
You will meet all four of these if you shop around. We went looking for where they came from and could not stand behind any of them.
“70% of big change projects fail” - McKinsey, Kotter, Bain - it is attributed to all of them
A researcher at the University of Brighton went to the five most-cited sources and found each one either states the number with no evidence or cites another source that states it with no evidence. The trail ends at a 1993 business book whose authors call it their own unscientific estimate. It has been repeated for thirty years and was never measured.
“85% of AI projects fail, according to Gartner” - Gartner
The nearest real statement is a 2018 Gartner prediction that through 2022, 85% of AI projects would deliver ERRONEOUS OUTCOMES because of bias in the data, the method or the team. That is a forecast about wrong answers, not a count of dead projects, and its window closed in 2022. Everyone quoting it is quoting a rewrite.
“87% of data projects never reach production” - A 2019 article, cited as research by hundreds of companies
The article is a sponsored piece reporting a conference panel, in which an executive said a trade magazine had said only 13% reach production. A magazine, quoted out loud by a vendor, written up in an advert, now cited as a study. There is no sample and no method at the end of it.
“It costs ₹1 to prevent a bad record, ₹10 to fix it, ₹100 to ignore it” - A 1992 book, in every retelling
Every source repeats the attribution and none cites a page or a method, and the heaviest repeaters all sell data-quality software. The SHAPE is uncontroversial and matches every case we read - an error gets dearer the later you find it - but the three numbers have no measurement behind them, so we will not print them as fact.
Four things we cannot tell you
- How many hours a month a business your size actually loses to re-typing. Nobody has measured it. The figure in circulation traces to a vendor’s blog with no stated sample, so the only honest version of this number is one we measure in your office, on your work, and show you.
- What a small automation error costs when it runs wrong for three weeks. We looked hard. The three famous costs-of-bad-data figures all collapse under their own provenance, and nothing credible has replaced them at the size of business this page is written for.
- How often this kind of project is abandoned in Indian owner-run firms of twenty to three hundred people. Every abandonment rate in circulation is drawn from large North American and European companies with formal procurement departments. None of it describes you.
- Whether smaller firms fail at this more or less often than large ones. There is a fair case either way - less process to break, but less slack to absorb a failure - and no data. Anyone who tells you which way it goes is guessing.
Proof
The unsubscribe was finished before the first campaign.A promotional messaging system for a multi-branch operator, built so the way out worked before the first message was ever sent.Before the first sendopt-out shipped first4 placesplaces kept in step7 phrasesways to opt outRead the study ↗Before you decide
Whether or not you ever talk to us, these are worth knowing.
- 01Time one round of the job by hand, honestly, including the interruptions. The number is usually two or three times what anyone guesses, and it is the number the whole decision turns on.
- 02Automate the work you repeat constantly, never the job you do twice a year. A rare task automated is a maintenance burden you bought.
- 03Ask where a person stays in control. Anything touching money, a customer, or a real decision should stop and wait for a human. If nobody can tell you where that checkpoint is, there is not one.
Is this you?
This is for you if
- +Owners watching skilled, expensive people spend hours on copy-paste work a machine could do.
- +Teams drowning in repetitive admin who want their time back without adding headcount.
- +Businesses where the same data is entered into several systems by hand, with the errors that brings.
Honestly, not if
- ×Anyone wanting to automate a real judgement call with no oversight. We design the human back in, not out.
- ×A one-off task you do twice a year. Automation pays off on the work you repeat constantly, not the rare job.
What this actually is
- What you get
- The repetitive jobs that eat your team’s day, handled quietly in the background, with a human checkpoint on anything that matters and a log of everything that ran.
- Week one
- We sit with the person who does the job and watch one full round of it, timing every step. That is where the real cost shows up.
- At the end
- Your team stops re-typing and starts approving. You can see what ran, when, and why, so nothing happens in the dark.
- Price
- It starts with a free diagnostic call. The first fix is then priced on the outcome it delivers, not the hours it takes. After that, a monthly retainer: the next fix, then the next.
Ready to talk
Thirty minutes, free. Show us one job your people do by hand.