Home / Services / AI deployment
SVC/AI — AI deployment & automation
The demo is easy. Running it is the work.
Anyone can show you a convincing chatbot in an afternoon — the tooling is that good now. What decides whether it is worth anything is the part nobody demos: whether it connects to your actual systems, what it does with your customers' data, and who fixes it at nine on a Monday.
POS/ — Where we're useful
Infrastructure people, doing AI.
Most AI consultancies are very good at prompts and models, and are meeting deployment, networking and data protection for the first time on your project. We came the other way: this practice sits on top of the same infrastructure, security and integration work on the rest of this site.
That matters because the hard problems in production AI are not model problems. They are "where does this run", "what happens when the provider is down", "who can see this data", "why did the bill triple last month", and "how do we know it is still answering correctly". Those are operations questions.
What we bring to it
- Deployment that is ours to own — self-hosted or cloud, monitored, backed up, patched
- Data governance by construction — what is sent, retained and logged, decided deliberately
- Integration into existing systems, because output in a chat window is not a workflow
- Cost control — usage visibility and limits before the invoice teaches you
- An honest "no" when a rules engine or a spreadsheet would do the job better
CAP/ — Capabilities
What we do in this practice.
AI/RDY
AI readiness review
Where AI would genuinely pay in your business, where it would not, what state your data is in, and what it would cost to run. Often the most useful thing we do, and it frequently ends in "not yet".
AI/DOC
Document intelligence
Invoices, purchase orders, lab reports, claims and KYC packets turned into structured data, with a human reviewing the exceptions rather than every page.
AI/KB
Internal knowledge assistants
Retrieval over your own SOPs, product documentation and past tickets — answering with citations, so the person asking can verify rather than trust.
AI/PRIV
Private & self-hosted deployment
Open-weight models running on your infrastructure or in an India region, for organisations that cannot send customer data to a third-party API. Network, access control, monitoring and backup included, because a model is just another production service.
AI/INT
Integration into real systems
The unglamorous half: connecting to your CRM, accounting package, calendar and messaging so output lands where work already happens instead of in a chat window.
AI/OPS
Running it afterwards
Monitoring, cost control, evaluation as prompts and models change, and a defined fallback for when the provider has an outage. AI features are production systems and decay like any other.
FIT/ — Being straight about it
Where AI pays, and where it doesn't.
Worth doing
- Extraction from documents where a human already checks the output
- Classification and routing at volumes too high for people and too irregular for rules
- Search over your own content, answered with citations the reader can verify
- Drafting — first versions a person edits, never output that ships unread
- Handling the repetitive front door — enquiries, bookings, status questions
Worth resisting
- Anything where being wrong is expensive and nobody checks the output
- Arithmetic and aggregation. A query is correct; a model is plausible
- Replacing a rules engine that works. Deterministic beats probabilistic when the rules are known
- Regulated advice — medical, legal, financial — stated by a model without review
- Anything adopted because competitors have one
The pattern separating the two columns: an AI feature works when there is a cheap way to catch it being wrong. Design that check first. If you cannot describe how a bad output gets noticed, the feature is not ready, however well it demonstrates.
FAQ/ — Common questions
Questions we get asked.
How is this different from your Data & AI practice?
Data & applied AI is about your data platform — pipelines, warehouse modelling, reporting, and LLM features built into an application we are already developing. This practice is about deploying AI into how the business operates: document processing, knowledge retrieval, and running models somewhere you control. Considerable overlap in technique, different starting point and different buyer.
Do you train custom models?
Almost never, and you should be sceptical of anyone who leads with it. For the overwhelming majority of business problems, a good general model with your data retrieved at query time beats a fine-tuned one — it is cheaper, it updates when your documents update, and it can cite its sources. Fine-tuning earns its place for consistent formatting or a specialised domain vocabulary, which is a narrower case than the marketing suggests.
Can we keep our data out of the big AI providers?
Yes, and for some organisations this is the deciding constraint. Open-weight models are now good enough for extraction, classification, summarisation and retrieval — the tasks most businesses actually need — and can run on hardware you control or in an India region. It costs more per unit of quality than a frontier API. Whether that trade is worth making depends on your data and your obligations, and we will give you a straight answer rather than a preference.
What about the DPDP Act?
India's Digital Personal Data Protection Act makes consent, purpose limitation and retention real obligations rather than good practice, and shipping customer personal data to a third-party model provider is a processing decision you need to be able to defend. We build with that in mind — what is sent, what is retained, what is logged, and where it physically sits. We are not lawyers and will not tell you that a design makes you compliant; we will tell you what it does with the data so your advisor can.
How do you stop it making things up?
Partly by design and partly by scope. Design: retrieval with citations so answers are grounded in your documents and checkable; constrained output formats; and refusing rather than guessing when confidence is low. Scope: we do not put a model in a position to state prices, make commitments or give regulated advice unverified. If you cannot describe how a wrong answer gets caught, the feature is not ready — that test decides more designs than any model choice.
Is this going to be obsolete in a year?
The conversational layer probably will be — platforms are commoditising it quickly, and we would rather say so than sell you a moat that does not exist. What lasts is the integration into your systems, the deployment, the data governance and the operational discipline. We build so that the model is a replaceable component, because it will be replaced.
RD/ — Related
AI booking & reception
The packaged version for appointment businesses — WhatsApp booking, reminders and enquiry handling.
Read →Before you buy an AI receptionist
The questions to ask, and what these systems genuinely cost to run once the demo is over.
Read →Low-code automation: where it pays, where it becomes debt
The same judgement applied to the automation layer underneath most AI deployments.
Read →CTA/ — AI work
Wondering whether any of this applies to you?
Describe the process you think could be automated and roughly how often it happens. You'll get an honest read on whether AI is the right tool — including when the answer is that it isn't.
Ask about a readiness review →