AI Operations · Assessment → Pilot → Managed service
Put AI to work across your business—without handing it the keys.
We map how your work actually happens, connect the systems you already use, and build AI workflows your teams own and our engineers operate — on the systems and tools you already run. Every consequential action stops at a named person for approval.
Enquiry → invoice · illustrative
Built around how your company already works
We start from your processes, not from a tool. We interview the people doing the work, map where the decisions and the handoffs actually are, connect the systems that already hold your records, and build the workflows with whatever you have already committed to — or propose a stack if you have none. The result is yours: workflows, prompts, policies and runbooks in your repositories, under your accounts.
We work inside what you already run
Your CRM, ERP, HRIS, service desk, document stores and cloud stay authoritative for your records. We connect to them under scoped identities and documented ownership. Your business does not get copied into something of ours.
Models and connectors are replaceable
Hosting and model choice follow your data rules, not our preference — your cloud, ours, or self-hosted on hardware we already run. Workflows are built so a model or a connector can be swapped without redesigning the process around it.
Leaving is designed in
Exit arrangements are written into the agreement at the start rather than negotiated at the end. The service identities are created in your systems, and you can revoke any of them without calling us first.
DevOps and SRE do not stop — this runs on them
Infrastructure as code, CI/CD, observability, incident process and on-call are what make an AI workflow safe to point at real records, and they are what we have always done. AI Operations is that same discipline applied to a new kind of workload: versioned changes, tested releases, monitored runs, a rollback that works. The platform and fractional DevOps work carries on unchanged, and a client can take either one on its own — but a team that already has the operational practice starts the AI work weeks ahead of one that does not.
It starts in one department and does not stay there
The first workflow is chosen on business value, repetition, data readiness and the cost of getting it wrong — not on which team asked first. Here is the kind of work that qualifies.
Sales
CRM · Proposals- Source-linked account briefs before a call
- Incoming enquiries qualified and routed
- Follow-ups and proposals drafted for approval
- CRM records updated from approved outcomes
Discounts and contract terms route to the person who owns them.
Customer support
Service desk- Permitted knowledge retrieved with source links
- Requests classified and queued
- Replies drafted for an agent to approve
- Exceptions routed to a named owner
Refunds, account access and anything unsupported escalate.
Finance
ERP · Accounting- Invoice fields extracted, orders and receipts matched
- Discrepancies explained against the source records
- Collection reminders and management reports prepared
- Duplicate invoice checks before anything posts
Payment execution stays inside your authorised finance process.
People and HR
HRIS- Policy questions answered under role-appropriate access
- Onboarding checklists prepared and coordinated
- Internal communications drafted
- Administrative tasks tracked to completion
Employment decisions stay with authorised people. Personal data stays restricted.
Operations and procurement
Suppliers- Requests triaged against approved limits
- Supplier information compared side by side
- Fulfilment coordinated, delays flagged early
- Purchasing recommendations prepared
Commitments above the agreed limit need a human signature.
IT and engineering
Our home ground- Service requests coordinated and triaged
- Incident investigation assisted with real context
- Changes drafted and put through review
- Infrastructure cost reviewed, routine tasks automated
High-impact production changes follow your existing change process.
Four phases, each one ending in something you can hold
You can stop after any phase. The work produced in it is yours either way.
Discover and prioritise
We interview leaders and the people doing the work, then map applications, data, decisions, repetitive tasks, exceptions and whatever AI is already in use unofficially.
You receive: process inventory, opportunity backlog, readiness and risk assessment, current performance baseline, recommended pilot scope.
Establish the foundation
Model access, identity and roles, secrets management, knowledge permissions, connector inventory, audit logging and separate environments — set up before any workflow is built.
You receive: architecture, data handling rules, access matrix, initial policy set, evaluation plan, incident and recovery procedures.
Build and validate a pilot
One bounded workflow, built with your team. Tested against normal work, missing data, hostile inputs, permission failures, duplicate events and a dependency going down. Historical cases first, then supervised live work.
You receive: working integrations, test evidence, user guidance, and a release decision signed off by the business owner.
Roll out and operate
Users and reviewers trained, runbooks published, approved execution enabled, volumes expanded gradually. Then we keep it running: failures, quality, cost, access and feedback watched, models and connectors maintained through controlled changes.
You receive: monitoring and on-call coverage, regular outcome reviews, and the next workflow chosen from the backlog.
Security, checks and balances, monitoring, and someone operating it
Four things, each a deliverable rather than a reassurance. They are set up in phase two, tested against the pilot, and kept running under the managed service.
Security
AI gets its own identities, its own limits, and never more reach than the person asking for the work.
- Scoped service identities, least privilege, SSO where your systems support it
- Credential rotation, and separation between environments and between customers
- Knowledge retrieved under the requesting user's own permissions — nobody sees more through AI than they already could
- Data classification, redaction, storage location, retention and deletion agreed per client, including what a model provider is allowed to keep
- Retrieved documents, emails and tool responses treated as untrusted content; prompt-injection attempts are part of the test suite
- The list of allowed actions enforced outside the model, where it cannot talk its way past it
- MCP and similar connectors are an interface, not a control — identity and policy checks sit around them
Checks and balances
Every workflow is registered at a declared level of autonomy. Raising a level is something a person signs, not a setting that drifts.
- Assist, Prepare, or Execute within limits — detailed below
- The reviewer sees the exact action and its source evidence, not a summary of it
- Deterministic rules handle calculations, routing and policy checks where a model should not
- Schema validation, business-rule checks and reconciliation; a model's own confidence score is never approval evidence
- Payments, contractual commitments, employment decisions and high-impact production changes stay inside your authorised process
- Workflows, prompts, models and policies are versioned and tested before release, with an approved version to fall back to
- Audit record per run: requester, workflow version, sources used, tool calls, policy outcome, reviewer, executed change, result
Monitoring
The signals we already watch on infrastructure, pointed at AI work — plus the ones only AI needs.
- Errors, latency, queue age and volume per workflow
- Permission denials and policy failures, treated as security signals rather than noise
- Review outcomes: how often a reviewer rejects or edits, which is where quality drift shows up first
- Spend per run and per workflow, so "is this worth running" has an answer instead of an opinion
- Evaluations re-run on every change to a model, a prompt or a connector
- Connector and knowledge-source health, with the last successful access per integration
- Incidents and exceptions routed to a named owner, with pause controls and a manual fallback
- Retries with duplicate prevention; an uncertain outcome is reconciled before anything is retried
Managed AI
Somebody is on the other end of it. This is the part most AI projects skip, and the reason so many stall after the demo.
- Monitoring and incident response under agreed coverage and response times
- Model, prompt, connector and evaluation maintenance through controlled changes
- An agreed improvement capacity each month, rather than a change-request queue
- Outcome reviews against the baseline measured before the pilot started
- Workflows that stop justifying their cost or risk get paused or retired — on our recommendation, not only on yours
- The next workflow chosen from the backlog, with you
- Run by the same team that carries your infrastructure on-call, not a separate desk
Three levels of autonomy, and a line that does not move
Every workflow is registered at one of three levels. Raising a level is a decision someone signs, not a setting that drifts.
Assist
Search, summarise and recommend. No business record changes. This is where anything uncertain starts.
Prepare
A draft or a proposed change. An authorised person sees the evidence and the exact action before it runs — not a summary of it, the action itself.
Execute within limits
Specifically permitted routine actions, with validation, budgets, rate limits and escalation rules. Granted per action, after the pilot has shown the action is safe.
Never automated, at any level
- Payments
- Contractual commitments
- Employment decisions
- High-impact production changes
Recorded for every run
- requester
- workflow version
- sources used
- tool calls
- policy outcome
- reviewer
- executed change
- result
Retention and log access are set per client, and evidence exports are mapped to your requirements. We do not claim that using this makes you compliant with anything — that is your auditor's call, and we give them something to read.
Where the approvals actually land
An approval is only useful if it reaches the right person in the place they already work — the service desk ticket, the chat channel or the mailbox they have open anyway. We agree that surface during the assessment and wire the pilot to it.
In the tools you already use
The request arrives in your service desk, chat tool or inbox, carrying the proposed action, the evidence behind it and the policy result — approve, reject or send it back from there. Nothing new to log into, and no second place to check.
The record behind it
Whichever surface it arrives on, every run keeps the same timeline: inputs, decisions, tool calls, approvals and results, with incomplete, denied, failed and partially completed runs shown as plainly as the successful ones.
Who gets asked, and what if they are out
Each workflow has a named owner and an authorised approver for the action in question, so nothing waits on whoever happens to be reading. Escalation rules and queue age are watched, because an approval sitting unanswered for three days is a failure the same way an error is.
Three ways in
Prices are set after the assessment, because the number depends on how many systems have to be touched and how much of the work is already written down.
AI Operations Assessment
- Process mapping across the functions you pick
- Readiness and risk assessment
- Opportunities ranked by value and feasibility
- Architecture direction and a pilot proposal
- Your current baseline, measured
AI Workflow Implementation
- Foundations: identity, permissions, logging, environments
- The integrations that workflow actually needs
- The agreed workflows, built with your team
- Testing against failure, not just the happy path
- Training, runbooks and handover
Managed AI Operations
- Monitoring and incident response
- Model, prompt and connector maintenance
- Evaluations re-run on every change
- Agreed improvement capacity each month
- Outcome reviews, and the next workflow chosen
The awkward questions first
Related Services
Start with an assessment, not a platform.
Two weeks of mapping tells you whether any of this is worth building for you. If the answer is no, we will say so.
Book an AI Operations Assessment