Enquiry triage
Reading incoming enquiries and classifying by service, urgency and sector, so routing is instant and a person confirms.
Automation · Delhi NCR & remote
AI is good at reading. It is bad at being sure.
Conventional automation follows rules. AI can handle tasks where the rule cannot be written down — reading, summarising, classifying. It is also confidently wrong sometimes, and that changes where it belongs.
Short answer
AI business automation applies language models to tasks that resist explicit rules: reading and classifying enquiries, summarising long documents, extracting data from unstructured text, drafting responses. It suits work where an occasional error is recoverable, and it does not suit work where being confidently wrong is expensive.
Handles unstructured text
Human check where it matters
Recoverable errors only
Cost modelled first
Free AI feasibility review
Get a free quote
Four details. A real reply the same day, from me.
Wondering where AI actually fits?
Get a Free Quote
WhatsApp
01 The premise
The one property that decides everything.
Conventional automation fails loudly. A rule that cannot match throws an error and somebody investigates. AI fails quietly and confidently — it produces a plausible answer that happens to be wrong, in the same tone as all its correct answers. That single property determines where it should and should not be used.
It follows that the right AI tasks are ones where an occasional wrong answer is recoverable and visible. Classifying an enquiry into a category: if it gets one wrong, a person notices and reclassifies it, and the cost is a minute. Extracting a figure from a purchase order that then triggers a payment: if it gets one wrong, the cost is considerably higher and nobody may notice for weeks.
That is the whole design principle. Use AI where it reads, drafts, summarises, classifies and suggests. Put a person between it and anything irreversible. Where a conventional rule can do the job reliably, use the rule — it is cheaper, faster and it fails in a way you can see.
02 Straight talk
Where AI earns its place.
Sorted by whether a confident error is recoverable.
Good AI tasks
- Classifying enquiries. By service, urgency or sector. A wrong one is noticed and corrected in seconds.
- Summarising long text. Call notes, documents, threads. The summary is read by a person who has the original.
- Extracting data for review. Pulling fields out of unstructured documents, presented for confirmation rather than acted on.
- Drafting replies. A first draft a person edits and sends. Saves time without removing judgement.
- Routing and triage. Deciding who should see something. Cheap to correct, useful when right.
Bad AI tasks
- Anything irreversible. Payments, orders, cancellations. A confident error here is expensive.
- Facts nobody will check. It will produce plausible wrong numbers, and confidence is not a signal of accuracy.
- Customer answers, unsupervised. A wrong answer sent in your name is your answer.
- Regulated advice. Medical, legal or financial. The risk is not worth the saving.
- Jobs a rule already does. If a rule works, use the rule. It is cheaper and it fails visibly.
03 Applications
What people actually build.
Practical applications with recoverable failure modes.
Call and chat summaries
Turning a long WhatsApp thread or call note into a short summary attached to the CRM record.
Document extraction
Pulling structured fields out of invoices, purchase orders or specifications, presented for confirmation before use.
Draft responses
First drafts of replies to common enquiries, edited by a person before sending. Time saved, judgement kept.
Internal search
Answering staff questions from your own documents, with the source shown so the answer can be verified.
Review and feedback analysis
Reading hundreds of reviews or survey responses and reporting themes rather than counting stars.
Free AI feasibility review
What would it cost if it got one wrong?
That question sorts good AI automation from bad faster than any technical assessment. Tell me the task and I will tell you honestly whether it belongs here.
- An honest answer, including "use a rule instead"
- Human checkpoints before anything irreversible
- Running costs modelled before committing
- Same-day reply, all 7 days
Free review
Get a free quote
Four details, and a real reply today.
04 Controls
What has to be designed in.
The controls that make AI automation safe enough to run unattended.
Controls
- A human checkpoint before anything irreversible. Non-negotiable. AI suggests, a person confirms, the system acts.
- Confidence handling. Uncertain cases routed to a person rather than guessed. The model should be allowed to decline.
- Source citation. Where an answer comes from your documents, show which one, so it can be verified rather than trusted.
- Logging of inputs and outputs. What was asked and what was produced. Essential for investigating a wrong result later.
- Cost modelling. Per-request charges look small and scale surprisingly. Model realistic volumes before committing.
- A fallback path. What happens when the AI service is unavailable. It will be, occasionally.
- Data handling decided. What is sent to a third-party model, and whether that is acceptable for your data. A real question, not a formality.
— Automation
Related services.
These sit next to AI automation and are usually bought with it. Same person doing the work in each case.
— Questions
AI automation — your questions.
Straight answers, including the ones that cost me work. If yours is not here, ask it — I reply the same day.
Tasks where the rule cannot be written down — reading unstructured text, summarising, classifying by meaning rather than by keyword, drafting. Conventional automation needs an explicit rule; AI can handle work where the pattern exists but resists being specified.
Reliable enough for tasks where an occasional error is visible and recoverable, and not reliable enough for anything irreversible. The important property is that it fails confidently rather than loudly, so it should never be the last step before something expensive happens.
Usually per request, and the per-request cost looks trivial until you multiply it by real volume. Ongoing running cost should be modelled honestly before building, because it is a recurring charge rather than a one-off build cost.
It depends on the provider and the arrangement, and it is a genuine question rather than a formality. Some models can run privately at higher cost. What data is sent to a third party should be a deliberate decision, made before building rather than after.
Normal automation wherever a rule can do the job — it is cheaper, faster, and it fails in a way you can see. AI is for the tasks that genuinely resist rules. Using AI where a rule would work is a common and expensive fashion.
For genuinely repetitive factual questions, with a clear route to a human and a source shown, it can help. For anything specific it should draft rather than send. A confidently wrong answer sent in your name is your answer, and customers do not accept the model as an excuse.
Free consultation · No obligation
Use AI where being wrong is cheap.
Describe the task and what a wrong answer would cost. That single question usually settles whether AI belongs there or whether a plain rule would serve you better.
- I will recommend a plain rule when that fits
- Human checkpoints before anything irreversible
- Running costs modelled before building
- Data handling decided deliberately
Start here
Send your enquiry
Name, number, email, what you need.