What to Automate With AI First: 8 Business Processes Mapped
By Akshit Agrawal, Co-founder, Bitvyas · Published 29 September 2026 · Updated 3 October 2026 · Facts as at 29 September 2026
Automate first the process that happens often, has rules you can write down, is cheap to get wrong, and already has its data in your systems. At Bitvyas, collections, reconciliation and client-document chasing usually pass all four tests; recruitment screening and contract review pass only with a person signing off every decision.
That is the short answer. The rest of this article is the test we use with clients, in India, the UK and the US, and eight common processes mapped against it: what the work looks like today, where AI takes it over, and what stays with a person.
Why does the first process matter so much?
Because the first project decides whether there is a second one. If it is expensive, vague or risky, the programme stalls.
Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing "escalating costs, unclear business value or inadequate risk controls". McKinsey's State of AI survey, published 25 August 2026, found 89% of respondents' organisations use AI regularly in at least one function, but only 37% attribute any EBIT impact to it.
Use is common. Measurable results are not. In my experience the gap starts with picking a process that looks impressive rather than one that repeats every week and has a clear finish line.
Key facts (as of 29 September 2026)
| Fact | Figure | Source |
|---|---|---|
| Organisations using AI regularly in at least one function | 89% | McKinsey State of AI, 25 Aug 2026 |
| Respondents attributing any EBIT impact to AI | 37% | McKinsey State of AI, 25 Aug 2026 |
| Agentic AI projects Gartner expects to be cancelled by end-2027 | Over 40% | Gartner, 25 Jun 2025 |
| EU AI Act rules for high-risk uses such as recruitment apply from | 2 Dec 2027 | European Commission, AI Omnibus in force 27 Jul 2026 |
| NYC law on automated hiring tools enforced since | 5 Jul 2023 | NYC DCWP |
| UK statutory late-payment interest (business to business) | 8% + Bank of England base rate | GOV.UK |
| India: MSME payment deadline where an agreement exists | 45 days maximum | MSMED Act 2006, s.15 |
What is the 4-question test?
Score each process 0, 1 or 2 on four questions. Eight is the maximum. A process scoring 6 or more is a good first candidate; 4 or 5 can work with a person approving every output; below 4, wait.
- Volume. Does it happen often? Daily or weekly work across many items scores 2. A monthly task with a handful of items scores 0, because the saving never pays for the build.
- Rule clarity. Could you write the rules on one page? "Match on GSTIN, invoice number and amount, within ₹1 tolerance" scores 2. "Use your judgement" scores 0. AI handles messy inputs well; it still needs a clear definition of done.
- Cost of error. What happens if it is wrong, and who notices? If a wrong output is cheap, reversible or checked before it leaves the building, score 2. If a mistake reaches a customer, a regulator or a court unchecked, score 0, unless a person reviews every output.
- Data available. Is the input already digital and reachable? Invoices in Tally, Zoho Books, Xero or QuickBooks, a CRM, a shared inbox: score 2. Data on paper, in someone's head or behind a system with no export: score 0.
Two notes on using it. First, score the process as it runs today, not as it should run. Second, the test ranks candidates; it does not replace a look at the actual week. We map a team's week before we recommend anything.
How do the 8 processes score?
The scores below are illustrative: typical for a business where the process runs weekly and the data sits in standard tools. Yours will differ, which is the point of scoring them yourself.
| Process | Volume | Rules | Cost of error | Data | Typical total | Person keeps |
|---|---|---|---|---|---|---|
| Collections follow-up | 2 | 2 | 1 | 2 | 7 | Disputes, payment plans, stopping supply |
| Reconciliation | 2 | 2 | 1 | 2 | 7 | Unexplained differences, write-offs |
| Client documents | 2 | 2 | 2 | 1 | 7 | Partner calls on late or incomplete files |
| Sales follow-up | 2 | 1 | 2 | 2 | 7 | The first call, pricing, the close |
| Insurance renewals | 2 | 2 | 1 | 2 | 7 | Re-marketing advice, coverage changes |
| Operations triage | 2 | 1 | 2 | 1 | 6 | Priority calls, anything unusual |
| Recruitment screening | 2 | 1 | 0 | 2 | 5 | Every shortlist and rejection decision |
| Contract review | 1 | 1 | 0 | 2 | 4 | Every legal judgement and sign-off |
Process 1: Collections follow-up
By hand: someone runs an ageing report, decides who to chase, writes the emails and remembers to follow up. The polite customers get chased; the difficult ones get put off.
Where AI takes over: it reads the ledger daily, sends reminders in your tone on a set schedule, reads replies, and sorts them: "paid", "will pay on the 15th", "dispute", "wrong contact". It logs promises to pay and chases again when one is missed. In the UK it can state the statutory late-payment position politely; in India, if you are a registered MSME supplier, it can flag buyers past the 45-day limit in section 15 of the MSMED Act.
What stays with a person: disputes, payment plans, credit holds and any customer relationship worth a phone call. See AI chase unpaid invoices. The full step-by-step map, with UK and India late-payment rules, is in Collections with AI: what it chases, what your team still handles.
Process 2: Reconciliation
By hand: comparing supplier statements, bank lines and the purchase ledger line by line at month-end, often in a spreadsheet with lookups that break on a stray slash in an invoice number.
Where AI takes over: it normalises invoice numbers, matches on several fields with tolerances, explains each difference in plain words ("likely typo", "timing", "short-paid") and drafts the query to the supplier. For Indian GST, the same engine runs purchase register against GSTR-2B; we wrote up that build in GSTR-2B reconciliation automation.
What stays with a person: approving adjustments, write-offs and anything the rules cannot explain. See AI reconcile supplier statements.
Process 3: Client documents
By hand: at every close or filing season, an accounting or audit practice requests the same documents from every client, chases the silent ones and files what arrives against the right engagement. The chasing often lands on senior people.
Where AI takes over: it sends the checklist per client, reads what comes back by email or WhatsApp, recognises a bank statement from a sales register, checks the period and completeness, files it and sends a precise follow-up for what is missing. For GST-period work, see what the GSTR-3B lock means for timing.
What stays with a person: the call to a client who is badly late, and judgement on whether a file is good enough to start. See AI collect client documents.
Process 4: Sales follow-up
By hand: enquiries arrive through a website form, a shared inbox and WhatsApp. Someone has to notice them, work out who is serious and reply before the prospect talks to a competitor.
Where AI takes over: it replies within minutes, asks the qualifying questions you would ask, updates the CRM and books a call for the ones that fit. It follows up on quotes that went quiet.
What stays with a person: the first real conversation, pricing and the close. Rules score 1 because "serious buyer" needs a clear definition before you start. See AI qualify and answer leads.
Process 5: Insurance renewals
By hand: an agency's revenue is mostly renewals, tracked in an agency management system and a calendar. Renewals are lost quietly when nobody re-markets a policy or reminds a client in time, and certificate requests interrupt the day.
Where AI takes over: it works the renewal calendar 90, 60 and 30 days out, gathers updated details from the client, prepares the comparison pack and issues routine certificates of insurance from templates. This applies equally to a US independent agency and a UK broker.
What stays with a person: advice on cover, any change in risk and the renewal conversation itself. See AI handle policy renewals.
Process 6: Operations triage
By hand: requests arrive by email, WhatsApp, a form and in person. Urgency is set by whoever asks most often, and nobody has one view of what is open.
Where AI takes over: it reads every incoming request, tags it, routes it to the owner, answers the routine ones from your own documents and warns before something breaches its deadline.
What stays with a person: priority calls between teams and anything unusual. Data scores 1 because requests arriving by phone or in person still need someone to log them. See AI sort and route requests.
Process 7: Recruitment screening
By hand: a recruiter reads every CV against a loose brief, reviewers use different yardsticks, and booking one interview takes a chain of emails.
Where AI takes over: it scores each CV against a written rubric, shows its reasons line by line, and books interviews in the calendar. The scheduling part is low-risk and can run on its own.
What stays with a person: every shortlist and every rejection. Cost of error scores 0 because hiring decisions carry legal weight: New York City has enforced bias-audit and notice rules for automated hiring tools since 5 July 2023, and under the EU AI Act, recruitment AI is high-risk, with those rules applying from 2 December 2027. In India, candidate data falls under the DPDP Act and Rules. Design for human decision from day one. See AI screen CVs and book interviews.
Process 8: Contract review
By hand: contracts arrive from every team at once. The same non-standard clauses get missed under time pressure, and the quality of review depends on who reads it.
Where AI takes over: it compares each contract with your playbook, flags missing or unusual clauses, marks the risk level and routes it: clean to the business owner, flagged to counsel.
What stays with a person: every legal judgement and sign-off. It scores 4, so run it only with review on every output. Law firms in India need a tool built for Indian law; that is what our sister product Lawgix AI does. See AI triage incoming contracts.
What should you not automate first?
Anything that scores below 4, and three kinds of work in particular:
- Work you have not written down. If two people do it two ways, agree the process first. Automating a disagreement makes it faster, not better.
- Rare, high-stakes decisions. A year-end judgement call or a single large negotiation happens too rarely to repay the build.
- Processes whose data lives on paper. Digitise the input first, or start with document intake as its own project.
How do you start once you have picked one?
Start with one process, one team and a clear measure: hours spent per week, days to close, invoices past due. Run the AI alongside the current way for two to four weeks, with a person approving every output. Then remove approvals only where the error rate says you can. Keep the approval step permanently for anything in the "person keeps" column above.
Bitvyas builds this kind of automation inside the tools a business already runs, for growing businesses and enterprises. All eight processes above have a live demo on our use cases page, and services sets out how we scope a first workflow.
Sources
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- McKinsey, The state of AI, published 25 August 2026: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- European Commission, "AI Omnibus enters into force": https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force
- NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools: https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page
- GOV.UK, Late commercial payments: charging interest: https://www.gov.uk/late-commercial-payments-interest-debt-recovery/charging-interest-commercial-debt
- Late Payment of Commercial Debts (Interest) Act 1998, s.5A (fixed sums £40, £70, £100): https://www.legislation.gov.uk/ukpga/1998/20/section/5A
- Micro, Small and Medium Enterprises Development Act 2006, ss.15–16: https://www.indiacode.nic.in/handle/123456789/2013
- PIB, DPDP Rules 2025 notified, 14 November 2025: https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf
Figures as of 29 September 2026. Legal and regulatory points are summaries, not legal advice.
See it working
Keep reading
Still doing this by hand?
Bitvyas builds AI that takes this kind of work off your team. It reads the documents, does the matching and drafts the next step, inside the tools you already use. A person still approves anything that matters.
Tell us what your team repeats