GenAI & Agent Build
AI agents built into real work, measured before they scale.
An AI assistant helps one person at a time. An AI agent does a defined piece of work, every day, with the results measured. Adamant Tech designs and builds those agents, from the first measured proof to production, and runs the same way of working on its own platform, 12website.
Where agents earn their keep
- Review and triage. Sorting, routing and first-pass checks on high-volume work, with a person approving the outcome.
- Document and e-mail intake. Inbound documents and e-mail in mixed formats are read, the key fields are extracted, and a person sees where each value came from.
- Internal knowledge. Assistants that answer from the organisation’s own documents and policies, with sources shown.
Other workflows are scoped the same way: a clear task, a measured baseline and a named person who owns the outcome.
How the work runs: measure, then automate
The work moves in phases. Each phase has a fixed fee, a written deliverable and a go or no-go at the end. The client can stop after any phase and keep everything built so far.
- Discovery, about two weeks. Time with the people who do the work today, a review of real samples, the fields or outcomes that matter, and the success measure and stop point, agreed before any build. Data protection is covered here: where the data may go, and the POPIA cross-border basis for the model provider.
- Proof of concept, about four to six weeks. A batch run over historical work, compared item by item with what people did. The output is an accuracy report and a business case built on measured numbers.
- Pilot, about eight to twelve weeks. The agent runs on live work for real users, with a review screen and monitoring. It ends with a recommendation to run or to stop, based on the correction rates seen.
- Run. Automation is widened only where the numbers have earned it, then handed over to the client’s IT or supported on a retainer.
What stays true in every phase
- A named person on the client side owns each outcome. Agents prepare the work and show their sources; what needs a person’s approval is set per phase in the proposal.
- Every extracted value carries a confidence score and a pointer to where it came from.
- Nothing writes to the client’s systems of record until the pilot result is approved. In the pilot, a person confirms every write.
- It runs in the client’s own cloud tenant where possible, with the model provider the client has approved, on serverless infrastructure that costs little when idle.
- Code written for the client is delivered to the client’s repository. Ownership is set out in the written quote; Adamant Tech’s pre-existing methods and tooling remain its own.
What the client commits to
AI projects fail when nobody owns them. Before the build starts, Adamant Tech asks for a named business owner, a subject-matter expert for a few hours a week, real (or masked) samples paired with the known right answers, a sandbox or a signed data agreement, and a security contact.
Price
A fixed fee per phase, quoted after a scoping call. How client data, access and AI models are handled is set out in Security and data. Delivered by the principal, remotely or on-site by arrangement, including across time zones.
Not included
Licences for the client’s cloud or model provider, which are billed to the client directly, and legal advice. Accuracy is measured and reported, not promised.