Fast is easy. Fast and controlled is the engineering.
Put AI to work in the engineering team with one controlled route in front of the models, and a record of what agents did.
Most developers now use AI coding assistants and many are trying agents. They are also the first to say the output is almost right and not quite. GenWA puts identity, rules, data protection and metering on the AI you route through it, and keeps a log you can read.
of developers use coding assistants or coding agents, and only 6.6% would trust AI output for important decisions.
Stack Overflow Developer Survey, 2026, over 30,000 responses from 169 countriesuse automated agent workflows, and about half trust AI output when they can easily check it.
Stack Overflow Developer Survey, 2026named AI solutions that are almost right, but not quite, as a problem they meet.
Stack Overflow Developer Survey, 2025of agentic AI projects are forecast to be cancelled by the end of 2027, on cost, unclear value or weak risk control.
Gartner, June 2025 (analyst forecast)You are asked to move faster with tools that you cannot yet fully verify, price or constrain.
of developers would trust AI output for important decisions, even as most use it daily.
Stack Overflow Developer Survey, 2026In 2023 engineers at a large manufacturer pasted source code and meeting notes into a public chatbot within about twenty days, and the firm then banned it (CNBC, Forbes). What it costs you: intellectual property outside your control.
In July 2025 a coding agent deleted a production database during a code freeze and then fabricated records (The Register). What it costs you: an outage, a restore and a hard conversation about permissions.
Assistants, plug-ins and agents arrive on personal accounts and cards. What it costs you: no register, no owner and no consistent rules.
Between January 2025 and April 2026 tokens used rose roughly 1,001% against spend up roughly 497% for Ramp's customers (vendor data). What it costs you: a budget argument without an answer.
Retirements and silent behaviour changes break prompts, tests and agent behaviour. What it costs you: regression work every time a supplier moves.
Guidance from the National Cyber Security Centre for AI agents stresses scoped credentials, logging, human oversight and a way to stop them. What it costs you: stalled projects while permissions are untangled.
A prompt is an instruction. It is not a permission system.
Telling an agent what not to do is not the same as limiting what it can do. Engineering teams that keep control give each agent its own identity, narrow credentials, a budget and a way to stop it, and they route model requests through one place where rules, data protection and metering apply. That is also what lets you say yes to more of it. GenWA deploys and implements these harnesses, including your own; we do not build your coding tools.
Controls apply to AI used through GenWA, meaning connected and routed integrations. They do not extend to tools that are not routed through it.
Six ways to give engineers the models they want, inside rules you can defend.
Guard sits in front of the models. Analyse and Evidence show what happened. Advisers, human and agentic, help you deploy it.
GenWA applies your rules and data protection to every request routed through it, deployed in your AWS or Azure account, or hosted by GenWA on AWS as agreed.
About GuardCode and secrets kept out of promptsData Protection redacts personal data and applies your rules before a request reaches an approved external model.
About Data ProtectionModels tested before they shipChange Control and the Sandbox let you test a model or release before it reaches your people.
About Change ControlOne register of what is connectedRegister lists the tools, accounts and agents connected to GenWA, the data they touch and a named owner for each.
About RegisterA Log of what agents didThe Log records what AI saw, produced and who approved it, so an incident review starts from facts.
About the LogAgents with a structure and an ownerAgentic Org Structure is being built: a bespoke or pre-built design of people and agents, with named people accountable. The Intelligence Layer, also being built, will draw answers from the Log.
About Agentic Org StructureTokens are rising faster than the money spent on them, and the invoice does not say why.
Between January 2025 and April 2026 tokens used rose far faster than spend. A bill shows the total. It does not show which team, tool or task produced it.
Animated chart comparing growth in tokens used with growth in AI spend
Tokens used (the units AI is billed in)+1,001%
AI spend+497%
Usage has grown about twice as fast as spend.
- Drafting tests, documentation and first versions of code a developer reviews.
- Explaining unfamiliar code and summarising changes.
- Triage of tickets and logs for a person to confirm.
- The output must be identical every time, such as a calculation or a migration step.
- The change touches production without review.
- The decision concerns credit, hiring or medical matters, where people decide.
Engineering already knows where determinism matters. Apply the same discipline to AI.
Three sectors where engineering faces the toughest AI constraints.
Pick the sector closest to yours.
Source code, customer data and security questionnaires
Engineering teams build on AI and are asked about it by customers. Guard keeps code and secrets out of unapproved tools, and Evidence Packs answer questionnaires from the Log.
See technology and softwareOlder core systems and conduct rules
Teams modernise carefully around systems that run the business. Rules set where AI may and may not touch them, Change Control tests changes, and the Log supports audit.
See financial servicesDesign files and operational technology
Design files and supplier detail must not leak into public tools, and customers expect proof, such as Cyber Essentials (the UK government-backed security scheme). Rules block design files from public routes.
See manufacturing and logisticsThree routes for an engineering team, from a review to a controlled build.
Advisers are human and agentic. Founder-led businesses can also apply to the Accelerator, a fourth route paid by equity, project, retainer or licence.
Our advisers, human and agentic, review how engineers use AI, set the rules for models, data and agents, and design the harness on paper first.
See advisoryProductsRules, Data Protection and the Log run in your AWS or Azure account, or hosted by GenWA on AWS as agreed. They apply your rules to AI used through GenWA and keep the record.
See GuardIncubatorWe build and run an agent function with you in an agreed engagement in a controlled environment, with scoped identities and a way to stop each agent. You own the rules, prompts and record.
See the IncubatorThree moves an engineering lead can make this month.
Three colleagues who share the decision, and three functions to bring alongside.
The people who own technology, spend and risk, and the teams who depend on engineering.
Owns the list of what is connected and the route AI takes.
For the chief technology or information officerChief financial officerOwns the number for AI spend.
For the chief financial officerCompliance and General CounselOwns the rules and the record.
For Compliance and General CounselProduct teamsChoose the models and own feature cost.
AI for product teamsCompliance teamsNeed the record that shows rules were applied.
AI for compliance teamsLegal teamsReview code, contract and data terms.
AI for legal teamsWant to see one route in front of the models?
Book a demo and we will walk through identity, rules, data protection, metering and the Log, using example data.
Keep reading
Straight answers.
How do engineering teams control AI coding tools and agents?
Give each tool and agent an identity, narrow permissions, a budget and a way to stop it, and route model requests through one controlled place. GenWA applies identity, rules, data protection and metering to AI used through it and keeps a log of what happened.
Can GenWA stop source code and secrets reaching public AI tools?
For AI used through GenWA, yes: rules and Data Protection apply before a request reaches an approved external model. It cannot control tools that are not routed through it, so we help you find those and decide what to bring across.
Does GenWA build a harness for us?
No. We are harness deployment and implementation specialists. We deploy and run harnesses, including ones you already have and those in the GenWA Toolkit, and fit them to your estate.
Which models can we use?
The approved providers are OpenAI, Anthropic, Amazon Bedrock, Azure OpenAI, Google Gemini and private models. Rules decide which model handles which task, data type and risk.
Where is it hosted?
It is hosted in your AWS or Azure account, or hosted by GenWA on AWS as agreed. You choose, and we agree it in writing.
Will this slow developers down?
It should not. The aim is an approved route that is easier than a personal account, with checks only where the risk justifies them. A person approves anything consequential.
How do we get started?
Start the AI health check or book a demo. We will show the route and the Log on example data and help you choose a first team to bring across.
Give engineers the models they want, inside rules you can defend.
Controls apply to AI used through GenWA. Hosted in your AWS or Azure account, or by GenWA on AWS as agreed.