Agents that do the work, not demos that describe it

An agent is a loop: perceive the current state, plan a sequence of steps, act by calling tools, verify the result, and reflect for next time. The tools are what make it useful — querying your database, writing to your ERP, sending a message, reading a file, controlling a browser. Without tools an agent can only produce text.

The critical design decision is which of those tools the agent may use without a human. Our default is that reads are autonomous and writes are gated. An agent may query anything it is authorised for, compose a response, prepare a transaction and present it for approval — and a human confirms before anything changes in a system of record. As confidence builds on a specific task, with evaluation data to support it, we widen the boundary for that task specifically.

This is not timidity. It is that the failure mode of an over-trusted agent is not a wrong answer on screen; it is three hundred incorrect entries in your ERP that take a fortnight to unwind. Starting gated and widening deliberately is faster overall than starting open and rebuilding trust after an incident.

Most mid-sized Indian companies process a large volume of semi-structured documents by hand: purchase orders arriving by email in twelve different formats, supplier invoices, transport documents, lab reports, bank statements, customs paperwork, insurance claims. A person reads each one and types the fields into a system. It is slow, error-prone and universally disliked.

Modern document AI handles this well, including scanned documents, poor photocopies, rotated pages, tables that span pages and handwritten annotations. Our systems extract the fields, validate them against your master data — is this a known supplier, does the GSTIN check out, does the total match the line items — and post them straight through where confidence is high. Only the uncertain cases reach a human, with the document and the extracted values side by side for a two-second correction.

Typical straight-through rates are 88 to 96% after tuning, which converts a full-time data entry role into a twenty-minute daily exception review. And because every extraction is confidence-scored and logged against the source document, an audit is a query rather than a search through filing cabinets.

Every organisation carries knowledge in documents nobody can find: standard operating procedures, equipment manuals, quality specifications, contracts, past project reports, policy circulars. Somebody spends an hour a week hunting for the right version, and somebody else makes a decision without it.

We build retrieval-augmented assistants over this material with three non-negotiable properties. Every answer cites its source, with the document and the page, so the human can verify. The assistant refuses to answer rather than inventing when the retrieval returns nothing relevant — an assistant that guesses confidently is worse than no assistant. And answers respect access control, so a user only receives content from documents they are entitled to see.

Quality is measured, not assumed. We build an evaluation suite of at least 150 questions with known correct answers, drawn from real user queries, and run it on every change to the model, the prompts or the retrieval configuration. Accuracy is tracked over time and reported. Without this you have no way of knowing whether last week's prompt change made things better or worse.

Order processing: an email arrives with a purchase order attached. The agent extracts it, matches the customer, validates products and prices against the current list, checks credit availability, flags anything unusual, and prepares the sales order for a one-click human confirmation. What was ten minutes of work becomes fifteen seconds of review.

Reconciliation: bank statements, gateway settlements and ledger entries matched automatically with the exceptions — genuinely ambiguous cases — presented for judgement. Finance teams routinely report this as their single biggest time recovery.

Support triage: incoming queries classified, enriched with account context, answered directly where the answer is unambiguous and documented, and routed with a suggested response where it is not.

Procurement follow-up: an agent that watches open purchase orders, contacts suppliers on schedule for status, records responses, and escalates the ones that are slipping — a job currently done by somebody with a spreadsheet and a phone.

Report generation: narrative commentary drafted from the actual numbers, with variances identified and explained against last period and target, for a human to edit rather than write from scratch.

AI systems fail differently from conventional software. They do not throw exceptions; they produce plausible output that is wrong. That requires a different kind of engineering discipline.

Structured outputs are enforced by schema, so a response that does not match the expected shape is rejected rather than parsed hopefully. Confidence thresholds route uncertain cases to humans. Every decision is logged with its inputs, its output and its confidence, so any case can be reconstructed. Rate and cost limits prevent a runaway loop from producing a surprising bill. And personally identifiable information is redacted before it reaches a model where policy requires it.

Evaluation runs continuously against a golden set, and the results are visible to the client rather than held by us. Model providers update their models, your documents change, and your users start asking different questions — all of which can shift quality without any code changing. A system without evaluation degrades silently, which is the worst possible failure mode because it erodes trust before anyone can explain why.

We work across frontier models from Anthropic, OpenAI and Google, and open-weight models such as Llama and Mistral where they can be self-hosted. The choice is driven by the task rather than by preference: complex reasoning and long-document work go to the strongest available model; high-volume classification and extraction often run perfectly well on a smaller and dramatically cheaper one.

Cost engineering matters more than most teams expect. A naive implementation that sends full documents to a frontier model on every request can cost ten to twenty times a well-designed one. We use tiered routing — a cheap model attempts first and escalates only when confidence is low — plus caching, prompt compression and batching. Across our deployments this typically reduces inference cost by 60 to 80% with no measurable quality loss.

For clients who cannot send data to an external service — defence-adjacent work, certain financial and health contexts, or simple policy — we deploy open-weight models on your own infrastructure. Quality is somewhat below frontier models for the hardest reasoning tasks and entirely adequate for extraction, classification and retrieval, which is where most business value sits anyway.

Almost every engagement begins with a two-week discovery: we watch the process, count the volume, measure the current time cost, and identify where AI genuinely fits versus where a rule or an integration would do. You get a written assessment with a ranked list of opportunities and honest estimates, and you are free to take it elsewhere.

The first build is deliberately narrow — one process, one document type, one workflow — deployed with a human gate and measured against the baseline we established. Expansion is earned by demonstrated results rather than assumed from a roadmap.