Services — Advise · Build · Tailor · Modernise
Onega advises on where AI pays, builds the agents and models that do the work, and modernises the applications they depend on. Our own engineers deliver every engagement inside your environment, on a governed, self-hosted stack, and hand over a runbook at the end.
Deployed, not demoed.
01 — Practices
Advise
Decide where AI pays before you pay for it. We map candidate use cases, measure the work they would change, classify each under the EU AI Act, choose the architecture and write the business case your board and works council can check.
Starts withAI Opportunity Assessment · 3–4 weeks
Build
Agents that do real work, within limits you set. We design, build and evaluate AI agents and multi-step workflows that use your tools and data, with a person approving what matters and a kill switch your security team holds.
Starts withDiscovery Sprint · 2 weeks
Tailor
Models that know your business and run where your data lives. We ground models in your documents, tune instructions and policy, fine-tune open-weight models on your approved data, and prove every change on an evaluation set built from your own examples.
Starts withModel Evaluation Sprint · 2–3 weeks
Modernise
Bring AI to the systems you already run. We assess your existing applications, add AI through one governed API where a rewrite is not needed, turn legacy functions into safe tools for agents, and migrate step by step where it pays.
Starts withModernisation Assessment · 3–4 weeks
02 — How engagements run
Every engagement starts with a fixed-scope, fixed-price assessment. Pilots are measured against the baseline agreed in that assessment, and scale-up happens only on evidence.
Baseline, data flows, risk register with an EU AI Act classification, architecture, acceptance criteria and a fixed-price proposal. You may stop here.
One bounded scope built inside your environment, evaluated on a holdout sample, run under supervision and handed over with a runbook.
A Forward Residency adds the next workflows, agents or applications, with an engineer embedded in your team.
Monitoring, evaluation refresh, model and policy changes under change control, and a named support channel.
03 — What you can hold us to
The people who assess your case are the people who build it. Every deliverable is something you can run, test or audit.
Agents and models run behind policy, budgets, redaction and an audit trail, in OneVeer or in the controls you already operate. AI proposes; people decide what matters.
Frontier or open-weight, in the cloud or on your hardware, chosen per use case under your policy. Configuration, evaluation sets and runbooks are handed over.
Baselines before builds, evaluation before release, and a stop option at every stage. We publish methods, not promises.
04 — The stack underneath
Services are how we learn which problems repeat. Products are how we solve them faster the next time. Where it fits, your agents and models run behind OneVeer, and OneMail or OneDesk join when the work lives in a shared mailbox or at the desk. Where your own platform is the right answer, we build on that instead.
One governed API for every agent and model, local or in the cloud: a kill switch scoped to one agent, one model or everything; single-use approvals for tool calls; PII redaction; budgets; an AI bill of materials.
For work that arrives in a shared mailbox: designed around evidence for every recommendation, automation levels from Observe to Autonomous, and Dynamics 365 as the system of record.
For work at the desk: a native workspace where people review and send what AI prepares, with an encrypted local store and a device-wide AI kill switch.
Every engagement is expected to return something reusable to the stack: a connector, a deployment recipe, an evaluation set or a documented pattern. That is how the tenth deployment should cost less than the first.
05 — Examples
Advise
A review of twenty proposed AI use cases, narrowed to the three worth piloting, each with a measured baseline and an EU AI Act classification.
Build
An agent that reads an incoming request, looks up the customer record and prepares the case for an operator to approve.
Build
A retrieval assistant over maintenance manuals and service history that cites the page it used and says so when the sources are silent.
Tailor
An open-weight model fine-tuned to classify requests in five languages, scored against the larger model it replaces and served on hardware you own.
Modernise
Summaries, validation and drafting added to a fifteen-year-old order application through one governed API, while it keeps running as it does today.
Modernise
A step-by-step plan for a legacy application, with AI-assisted code analysis and characterisation tests written before anything moves.
Illustrative scopes, not customer references. Onega names customers and publishes results only with their written permission.
06 — Boundaries
In each case we offer a fixed-scope assessment with a stop option instead.
07 — FAQ
No. We recommend OneVeer, OneMail or OneDesk where they fit the problem and build on the platforms you already run where they do not. Each recommendation is written into the assessment with its reasons and the alternative we considered.
Frontier models from the major providers, open-weight models that run on your own hardware, and the models your organisation has already approved. The choice is made per use case on your evaluation set, your cost limits and your data constraints. Behind a governed alias it can change later without rebuilding the application.
Ownership is agreed in the statement of work before any build starts. It covers customer-specific code, your data, evaluation sets, model weights trained on your data, and Onega’s own background components and products. You always receive the configuration, documentation and runbook needed to operate the result or to exit.
In your environment: your hardware, your tenant or a sandbox you approve. There is no production access before written authorisation, synthetic or approved samples come first, and a data-processing agreement with a subprocessor list is in place before any personal data is processed.
Assessments and pilots are fixed-price and agreed before work starts. Residencies and Operate and Improve are monthly, with a defined scope. Software licences, compute and third-party model charges are priced separately. We do not publish prices.
Yes. We can work beside the partner who runs your systems, and implementation partners can be certified to deliver the method through the Onega Partner Network.