AI automation for the work your team repeats every day

A scorecard for choosing the first process

Score each candidate process from 1 to 5 on three factors and multiply the scores. The result (1 to 125) is a rough ranking, not a business case, but it stops the loudest request from jumping the queue.

Our rule of thumb: above about 60, it is a good first project. Between 20 and 60, automate the easy part and keep a person on the rest. Below 20, write the process down and leave it manual. A low error-tolerance score does not rule a process out. It means the design needs a human approval step, which reduces the time saved, and the ROI sum has to include that.

Zapier, Make, n8n, custom code or an AI agent?

The tool matters less than people think, but the pricing model matters a lot once volumes grow. Pricing models below are as of October 2026; check each vendor’s current plans.

If you are leaning towards n8n on your own server, our n8n self-hosting guide covers the setup, backups and upgrades. If the steps genuinely vary, read our AI agent development page before choosing an agent.

An honest ROI worksheet

Most automation ROI claims ignore running costs and the time people still spend checking output. Put your own numbers into this:

• Time saved per month = runs per month × (minutes per run today − review minutes per run afterwards) ÷ 60 × loaded hourly cost
• Errors avoided per month = errors per month today × share the automation prevents × average cost of one error
• Running cost per month = platform or hosting fees + model usage + maintenance hours × hourly rate
• Net monthly benefit = time saved + errors avoided − running cost
• Payback in months = one-off build cost ÷ net monthly benefit

Three honest adjustments. Measure the baseline before you build, not from memory. Budget a few maintenance hours every month, because APIs, templates and document layouts change. And count staff time as saved only if it goes to other useful work. If the net monthly benefit is close to zero, the project is not worth doing, however good the demo looks.

Where not to use AI

• Totals, tax and arithmetic. Calculate in code. A language model can produce a plausible wrong number.
• Anything a rule already handles. A lookup table or a regular expression is cheaper, faster and testable. Use a model only for the fuzzy step.
• Decisions with legal or similarly significant effects on people, such as credit, hiring or claim denials. GDPR Article 22 restricts decisions based solely on automated processing, and keeping a person in the loop is good practice everywhere.
• Clinical judgement. AI can sort, summarise and draft. Diagnosis and treatment stay with clinicians.
• Rare, high-variance work. If every case is different and it happens monthly, the build and testing will never pay back.

Privacy: DPDP, GDPR and HIPAA

Automation moves personal data between systems, often through a third-party platform and a model provider. Map that path before you build.

• India: DPDP Act, 2023. The DPDP Rules were notified in November 2025 with a phased start: consent-manager provisions from November 2026, and most obligations, including notice, consent, security safeguards and breach reporting, from May 2027 (our DPDP checklist has the exact dates). Record what each person agreed to and let them withdraw as easily as they agreed. Our DPDP timeline and checklist tracks the dates.
• EU and UK: GDPR. You need a lawful basis for each purpose, a data-processing agreement with every vendor in the chain, and a transfer mechanism if data leaves the EEA or UK.
• US healthcare: HIPAA. Every vendor that creates, receives, stores or transmits protected health information for you needs a business associate agreement (BAA), including the automation platform, the model provider and the host. HHS guidance says a cloud provider is still a business associate even if it stores only encrypted data and holds no key. Major model providers offer BAAs for eligible API use after review. Check whether your automation platform will sign one on your plan; if not, self-host or use code.

Whatever the law, the same habits apply. Send a model only the fields a step needs, use API terms that exclude training on your data, set retention periods on logs, and keep an audit trail of what was sent where.

Worked examples (illustrative)

These are illustrative scenarios showing how we would approach common requests. They are not client case studies.

1. Purchase-order extraction for a Tiruppur garment exporter. Buyer POs arrive as PDFs in a dozen layouts. A model extracts style numbers, size-wise quantities and ship dates; code checks that quantities add up to the PO total; failures go to a merchandiser before anything reaches the production sheet.
2. WhatsApp order status for a Coimbatore pump distributor. Dealers send an order number to the company’s WhatsApp number. The workflow looks it up in the ERP and replies with dispatch status and the transporter’s LR number. Anything it cannot match goes to the sales desk.
3. Tally sync for a trading firm. Invoices raised in an online store or CRM are posted into TallyPrime as sales vouchers each night through Tally’s XML interface. Unknown ledgers and GST mismatches are listed for the accountant instead of being guessed.
4. Inbox triage for a Kochi tour operator. Email and web-form enquiries are sorted into new bookings, changes, complaints and supplier mail. New-booking replies are drafted with availability from the operator’s sheet, and staff approve each one.
5. Payment follow-ups for a Madurai wholesaler. Overdue invoices trigger WhatsApp reminders on set days with a payment link. The sequence stops on the payment webhook or any reply.
6. Monday report for a multi-branch retailer. Sales by branch and last week’s ad spend are pulled and calculated in SQL. The model writes five lines on what changed, and the figures travel with it.
7. Healthcare intake forms for a clinic. New-patient forms are extracted into the practice system, identifiers are masked before the model step where possible, and every vendor is covered by a BAA (US) or DPDP-ready consent (India).
8. Three-way invoice matching for a Hosur manufacturer. Supplier invoices are matched to the purchase order and goods receipt. Differences above a set tolerance are flagged; matches are queued for payment approval.
9. Lead routing for a B2B supplier. Leads from the website and marketplaces are deduplicated, enriched from the company’s website and assigned by product and region, with a first reply sent within minutes.

Common mistakes

• Automating a broken process. Fix the steps first. Automation makes a bad process fail faster.
• No named owner. Every workflow needs a person who receives the alert and decides what happens next.
• Building on a personal account. When that person leaves, the workflow goes with them. Use company accounts and keep workflow definitions in version control.
• No log of inputs and outputs, so nobody can explain a bad result or prove a good one.
• Skipping the shadow run. Run beside the manual process for a week or two before switching over.

How to choose an AI automation partner

• They ask about volumes, exceptions and error costs before they mention tools.
• They tell you when a rule, an off-the-shelf app or no automation is the better answer.
• They estimate monthly running costs (platform, model usage, maintenance) before you sign.
• Workflows live in your accounts, documented, with credentials you control.
• They can show how failures are logged, retried and escalated.
• Consent, data-processing terms and, where relevant, BAAs are part of the design, not an afterthought.
• They agree a before-and-after measure and report against it.