We work with public sector organisations worldwide and follow a consistent process for reviewing and entering these contracts, including compliance with our acceptable use policies and AI principles.<\/p><\/blockquote>\n
Both things can be true. The process was followed and the people closest to the work still lost confidence. Every serious AI deployment now ships with a principles document. The distance between that document and Tuesday’s deployment decision is where trust, talent, and reputation quietly drain away.<\/p>\n Google is one of several tech companies that signed contracts with the US military earlier this year to deploy AI across classified military environments. US Defence Secretary Pete Hegseth has made AI integration a priority for the armed forces, saying the US “will become an ‘AI-first’ warfighting force”. Those two facts together explain the timing of the departures.<\/p>\n The workers who spoke to the BBC were not objecting to public sector work in general. Their concern was specific: this year’s deals sharply expand the scope of what the Pentagon can use the technology for. Liguori pointed to the company’s growing work with both the US and Israeli militaries as the reason she lost confidence.<\/p>\n Scope expansion is the mechanism worth noting. Nobody re-approved the ethics when the use cases widened. The original review covered a narrower deployment, and the policy language stayed the same while the application grew underneath it.<\/p>\n Google’s response to the BBC was that it works with public sector organisations worldwide and follows a consistent process for reviewing and entering these contracts, including compliance with its acceptable use policies and AI principles. It also said it has dedicated channels for employees to safely raise concerns, covering business ethics and misconduct, with strict workplace policies to prevent retaliation.<\/p>\n Read that carefully. Every claim is about process existing, not about process catching anything. A consistent review process can be consistently applied and still approve deployments that the people building the system consider indefensible. Compliance with an acceptable use policy only means something if the policy was written tightly enough to exclude something.<\/p>\n That is the test any AI governance policy has to pass, in a plant or in a data centre. Does it name specific decisions the system is not allowed to make on its own? If your document only commits to review and oversight as general principles, it will survive contact with any deployment you happen to want.<\/p>\n The staff and experts the BBC spoke to did not say oversight was absent. They said they were concerned about the level of human oversight “at critical moments”. That qualifier carries the entire argument. A policy that promises human oversight without naming which moments are critical has promised nothing enforceable.<\/p>\n Most oversight language means a human can look at the output. That is a dashboard, not a control. The model has already acted, the batch has already moved, and someone can audit it later if they have reason to.<\/p>\n Real oversight has three parts: a named role, the authority to say no, and enough time in the process to actually look. If an AI system flags a defect, releases a batch, or approves a deviation, write down who signs, what evidence they see, and what happens when they disagree. If the answer to that last question is “the system proceeds anyway”, you do not have oversight. You have a logging feature.<\/p>\n Critics in the BBC piece questioned whether the technology is reliable enough for warfare without more testing, stronger guardrails and human oversight. The same lag shows up in plants. Deployment takes weeks, validation takes months, and the pressure to show results favours the faster of the two.<\/p>\n The fix is unglamorous. Define the failure modes before go-live, not after: what does a false pass cost, what does a false reject cost, and at what confidence threshold does the decision route to a person. Run the model in shadow mode against known outcomes long enough to see its error pattern, not just its headline accuracy.<\/p>\n Build that into your AI governance policy as a condition of deployment rather than a review that happens later. Otherwise your principles document and your production floor are two separate things, and only one of them is making decisions.<\/p>\n Write your own acceptable use boundary for AI in operations before a vendor writes one for you. Their version will be broad by design, because breadth protects them. Yours needs to be narrow and specific to your processes, your regulators, and your customers.<\/p>\n The cost case is simple. Defining governance upfront is a few days of work with quality, operations, and IT in a room. Retrofitting it after a regulator asks how a batch was released, or after a customer audit finds no decision trail, takes quarters. It also keeps the people who actually understand your systems from losing faith, which is what Google is dealing with now.<\/p>\n One page, two columns. On the left, decision classes where a human must sign off and the process cannot proceed without it. On the right, where AI acts unsupervised because the downside is recoverable.<\/p>\n That log is the difference between oversight you can prove and oversight you assert. Google says it has “dedicated channels for employees to safely raise any concerns”. Channels are not records. If an auditor cannot reconstruct who decided what and on what evidence, your policy is a statement of intent.<\/p>\n Read the AI principles and the contract terms of every vendor already embedded in your quality stack. Most MES, LIMS, and vision inspection suppliers have added AI features without changing a line of the agreement you signed three years ago.<\/p>\n Ask four things. Where does our data go and does it train your models. What changes when you push a model update, and do we get notice. Which decisions does your system make without a prompt from us. And what does your own acceptable use policy permit that we would not permit. If answers arrive as marketing language, escalate before renewal, not after.<\/p>\n![\"Engineer](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-governance-policy-lessons-f-inline-1.jpg\")
<\/figure>\nWhat Google Actually Signed, and What Its Staff Object To<\/h2>\n
The expanded contract scope staff flagged as the turning point<\/h3>\n
Google’s position: consistent review process and acceptable use policies<\/h3>\n
Human Oversight Is the Fault Line, and It Is Rarely Defined<\/h2>\n
Reviewable output versus a decision that cannot proceed without sign-off<\/h3>\n
Why testing and guardrails lag behind deployment speed<\/h3>\n
![\"Analyst](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-governance-policy-lessons-f-inline-2.jpg\")
<\/figure>\nWhat Manufacturing and Quality Leaders Should Take From This<\/h2>\n
A one-page oversight map for AI-influenced quality decisions<\/h3>\n
\n
Vendor due diligence questions worth asking before the next renewal<\/h3>\n