AISTORSServicesAI agents and assistants
Service 01Inside the system, not beside it.
For operations, service and support teams whose queues are growing faster than their headcount. You have work that arrives in a queue, follows a rule somebody already wrote down, and still needs a person to read it, retype it into a second system and close it. We build agents that do that work inside the systems you already run, with a human approval step wherever being wrong carries a cost.
Book a 30-minute call
If the left column is not you, tell us on the call and we will say what would actually help instead.
Four things go wrong with agent projects. Each one has a specific answer, agreed before any code is written.
What we do. People are the copy and paste between two systems nobody ever connected. We put the agent inside those systems, reading and writing through their own interfaces under your credentials, so the retyping stops instead of moving to a new screen.
What we do. Demonstrations pass because a person chose the examples. We score the agent against a fixed set of your own real cases, labelled by your team, before it ships and again on every change after that.
What we do. So it does not act alone. The boundary of what the agent may decide by itself is agreed in writing before the build and becomes a contract term. Consequential actions pass an approval gate, and the rollback path is tested rather than assumed.
What we do. We record cycle time, error rate, volume and hours before anything is built, in your systems, with the counting method written down. The same four measures are reported afterwards, so the renewal conversation has numbers in it.
Six agent types, all of which act inside your systems rather than describing what you should do.
Gartner names three causes of cancellation. Each one is addressed before the build, not after.
Cycle time, error rate, volume and hours on the target process, recorded in your systems before anything is built. This is what makes "unclear business value", the second cause Gartner names, impossible by construction.
Which systems hold the record, which have usable interfaces, where the records disagree. Gartner predicts 60 percent of AI projects will be abandoned through 2026 for want of AI-ready data, so this is settled before a model is selected.
A fixed, versioned set of your own cases, labelled by your team, with a target accuracy agreed before the build. Without this there is no way to tell a model change from a regression, which is Gartner's third cause: inadequate risk controls.
One process, in your accounts, with approval gates on consequential actions and cost instrumented per run. Gartner expects 80 percent of tangible agentic return by 2028 to come from specialised agents, so the first build is deliberately narrow.
The same four measures the Diagnostic recorded. If the numbers do not move, that is reported plainly and the second build is not automatically approved. Escalating cost is Gartner's first named cause of cancellation, and cost per completed task is tracked from the first week.
Built on your platform and your model contracts. Self-hosted where the data cannot leave your estate.
Azure
Azure AI Foundry, Azure OpenAI, AI Search for retrieval, Functions and Logic Apps for orchestration, Entra ID for agent identity, content safety filters.
AWS
Bedrock with Agents and Knowledge Bases, Bedrock Guardrails, Step Functions and Lambda, OpenSearch for retrieval, IAM roles scoped per action.
Google Cloud
Vertex AI with Agent Builder, Gemini models, Vertex AI Search, Workflows and Cloud Run, BigQuery for grounding on your own data.
DigitalOcean
GenAI Platform and managed inference, DOKS for orchestration, managed Postgres with pgvector for retrieval, or open-weight models self-hosted on your own droplets.
Model providers we work with include OpenAI, Anthropic, Google Gemini, AWS Bedrock and Azure AI Foundry, on enterprise or business terms with training on your data disabled. Where residency or sensitivity requires it, open-weight models run fully self-hosted in your own tenancy. The full disclosure is on the AI capability page.
You do not need these numbers to recognise the problem. They are here because somebody in your approval chain will ask.
17%
of organisations have deployed AI agents, while more than 60 percent expect to within two years. Gartner describes this as the steepest adoption curve among the emerging technologies it measured.
Source: Gartner 2026 CIO and Technology Executive Survey, via Gartner's Hype Cycle for Agentic AI, 2026.
40%
or more of agentic AI projects are predicted to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
Source: Gartner, 25 June 2025.
2 in 3
enterprises expect service providers to build and operationalise their priority use cases. One third are already scaling agentic deployments.
Source: BCG, The $200 Billion AI Opportunity in Tech Services, 2026.
Read those three together and the brief is obvious. Almost everyone intends to do this, most have not started, a large share of what does start gets cancelled for reasons that are all controllable, and enterprises have already decided not to do it alone. Gartner also expects 80 percent of tangible agentic AI return by 2028 to come from specialised, domain-specific agents rather than general-purpose ones, which is an argument for narrow scope and sector depth over platform breadth.
Intent is near-universal. Deployment is rare. Cancellation is common. All three are published numbers.
You do not, at the start, and neither do we. That is why the boundary of what the agent may do alone is agreed before the build and written into the contract, why every consequential action passes an approval gate, why the rollback path is tested rather than assumed, and why accuracy is scored against a fixed set of your own cases. The question is not whether an AI system will be wrong. It is what happens in the minute after it is.
Because Gartner's stated causes of agentic project cancellation are escalating costs, unclear business value and inadequate risk controls, and a baseline addresses two of the three directly. Without one, there is no way to defend the spend at renewal and no way to decide what to scale. The second project then gets approved on the same evidence as the first, which is none.
Yes, and it is usually preferable. We build on your contracts, in your tenancy, on your accounts. If you have no provider relationship yet we will help you choose one on the merits of your workload and residency requirements, and we hold no reseller agreement or commission on any of them.
Yes. Open-weight models run fully self-hosted in your own tenancy, which is slower to build and usually costlier to run than an API, and we will tell you the trade-off in numbers rather than in principle. Region is treated as a requirement rather than a preference, and every third party in the path is listed before work starts with a standing right for you to refuse any of them.
It is planned for, because it is certain. OpenAI announced retirements from ChatGPT effective 13 February 2026 and Anthropic's published policy is at least 60 days notice on publicly released models. The version is pinned, deprecation notices are tracked, and a successor is scored against your evaluation set before cutover. That handling sits in managed operation.
Yes, and it happens. Where the process is unstable, the data unreachable, or the volume too low to repay the build, a no-go with the reasoning attached is the outcome of the Diagnostic. You keep the report, the baseline method and the plan. We would rather lose the build than sell one that fails.
Thirty minutes, no obligation. Describe one process that is done by hand today. We will tell you whether it is a measurement problem, an integration problem or neither, and whether an agent is the right answer at all.
Or write to [email protected].
Managed AI and cloud operations
The same four measures every month, plus drift detection and model upgrade handling. What keeps the agent working after go-live.
Service 06Cloud cost and FinOps
Token and GPU spend attributed to the team that caused it. Escalating cost is the first reason agent projects get cancelled.
DisclosureAI capability and data handling
Which providers, on what terms, with training disabled, residency respected and every sub-processor named.