Pilots are easy. Operations are not.
Agents doing real work, run to a service standard with a named owner on every workflow.
Everyone is piloting agents. Few are running them the way you run any other part of the operation, with an owner, a measure of success and a way to stop them. The first failures are operational, not technical.
of agentic AI projects are forecast to be cancelled by the end of 2027, because of cost, unclear value or weak risk controls.
Gartner forecast, June 2025of US desk workers said they had received AI-produced work that needed redoing (survey).
BetterUp and Stanford, in Harvard Business Review, 2025of organisations with AI agents say an agent has exceeded its permissions (vendor-sponsored survey of 445).
Cloud Security Alliance and Zenity, April 2026Six operational problems that appear once agents leave the pilot.
of agentic AI projects are forecast to be cancelled by the end of 2027 (Gartner, June 2025).
Gartner forecast, 2025A pilot starts with enthusiasm and no owner. What it costs you: when the sponsor moves on, the workflow has no one accountable and drifts.
Work produced by AI that needs redoing can use up the time it was meant to save. What it costs you: hidden labour that appears in no business case.
Activity is counted, outcomes are not. What it costs you: a pilot kept alive or killed on opinion.
The same automation is built separately by several teams. What it costs you: duplicate effort, uneven quality and no shared standard.
Agents are given access and then act. A prompt is an instruction, not a permission system. What it costs you: an error with a customer or a supplier and no record of how it happened.
Agent workloads use many times the tokens of a chatbot. Gartner put the multiple at 5 to 30 times per task (March 2026). What it costs you: costs that surprise you at scale.
An agent is a new team member. It needs a job description, a manager, a budget and an off switch.
The question is rarely whether the model is clever enough. It is whether anyone is responsible for the work. Treat each agent the way you treat a new hire: a defined job, a named manager who approves consequential actions, a limit on what it can spend and touch, a weekly review and the ability to stop it quickly. Agents multiply whatever foundation they sit on. If the foundation is a clear process with an owner, they help. If it is a gap, they scale the gap.
GenWA view. Drawn in part from Gartner (2025) and UK National Cyber Security Centre guidance on AI agents (2026).
Six ways to run agents like the rest of the operation.
Contact is available. Comms is in early access. Skill Teams, Library and Agent Catalogue are being built. We confirm in writing what is live for you and do not describe a building item as live.
Contact runs on Amazon Connect: live adviser assist, voice, monitoring of every call against your rules and card payments. Monitoring is a capability, not a guarantee of compliance.
About ContactCustomer letters and emailsComms drafts letters and emails, checks each against your rules and routes to a named person for sign-off where it matters.
About CommsAgent teams for defined jobsSkill Teams are being built as agent teams for one defined job, with a named owner and a way to stop them.
About Skill TeamsKnowing which jobs to start withAgent Catalogue is being built as a list of agent jobs, owners and needs, to help you choose on value and risk.
About Agent CatalogueA record of what each agent didLog keeps a tamper-evident record of what AI saw and did and who approved it.
About LogOutcome by workflowValue Register records task, outcome, cost and time saved for each workflow.
About Value RegisterThree moves for the operations lead.
Three routes, matched to how operations buy change.
A review of which of your workflows are good first candidates, with owners, limits and measures defined before anything is built.
See advisoryProductContact and Comms for customer-facing work, hosted in your AWS or Azure account, or by GenWA on AWS as agreed.
See ContactIncubatorWe define, build and run an agent function for you, with a named owner on every workflow. You own the rules, prompts and record, and you can take it back.
See the IncubatorThe three colleagues an agent function depends on.
The named owner of the workflow, who approves consequential actions and reviews results each week.
Provides the systems access, identities and limits each agent needs.
For the CTO and CIOCompliance leadAgrees the rules and the approval steps so the record stands up later.
For Compliance and General CounselA coding agent deleted a company's production database during a code freeze.
In July 2025 an AI coding agent used by a software founder deleted a live production database during a code freeze and then generated misleading records. The provider added separation between development and production environments and a way to restore. The lesson for an operations lead is practical: scope what an agent can touch, keep a record of its actions and be able to stop it.
The Register and AI Incident Database, July 2025. Public incident. GenWA had no involvement.
Want to see agents doing real jobs under control?
Book a demo and we will show a workflow with its owner, limits, approvals and record.
Keep reading
Straight answers.
How do I run AI agents like any other part of the operation?
Give each agent a defined job, an identity and limits, and a named person who owns the workflow and approves consequential actions. Review results weekly and stop the agent if it is not working. GenWA keeps a record of what each agent did.
Who is accountable if an agent gets it wrong?
A named person in your business owns each workflow and signs off consequential actions. AI is not used for decisions on credit, hiring or medical matters; people decide. The record shows what was done, on what information and who approved it.
Our pilots work well enough. Why add controls now?
Good. The question is what happens at scale: who owns it, what it may spend, how it is reviewed and what the record shows. Cloud Security Alliance found 53% of organisations with agents had one exceed its permissions (vendor-sponsored, 2026).
What is live today and what is being built?
Contact is available. Comms is in early access. Skill Teams, Library and Agent Catalogue are being built next. We confirm in writing which products are live for you and do not describe a building item as live.
How much AI output ends up being redone?
One 2025 study reported in Harvard Business Review found 40% of US desk workers had received AI work that needed redoing. Your figure will differ. Value Register records outcome and time by workflow so you can see where AI helps and where it creates rework.
What do you need from us to start a workflow?
A workflow owner, a clear description of the job and its definition of done, and access to the relevant system or channel. Integration effort is agreed with you before we start. We start with the health check to see what already runs.
Do agents replace my team?
Not by design. People keep decisions that need judgement and approve consequential actions. We do not apply fixed percentage splits between human and agent work. The aim is consistent output on routine steps, with your people on the exceptions.
Run agents like any other part of the operation.
We confirm in writing which products are live for you before you commit.