Georgia's new AI playbook tells counties to find the AI they already bought. Here's how to do that in a week
Most city halls already run AI. It came in quietly, inside the permitting system, the HR platform and the document tool, with a release note nobody read. Georgia's playbook for local governments says to find it before buying anything new, and that's a job a small office can finish in a week.

Georgia's Local Government AI Playbook, released November 12, 2026, tells cities and counties to set policy, inventory existing AI systems, form oversight and train staff before buying AI products, and to look for AI features embedded in software they already use. A practical first step for a small government is a one-week inventory: list every vendor, ask each the same four questions, and record who in the organization owns each AI feature.
What Georgia published
On November 12, the Georgia Technology Authority's Office of Artificial Intelligence released a Local Government AI Playbook. StateScoop reports it runs 53 pages, was written by Georgia's chief digital and AI officer, Nikhil Deshpande, and is aimed at cities and counties that don't have AI teams, formal AI policies or the resources to build their own governance.
The striking thing is what it tells them not to do first. Rather than buying AI products, the playbook recommends that local governments start by setting policies, creating inventories of the AI systems they already have, forming oversight committees and training employees. It specifically asks them to identify AI features embedded in software they already use, naming human resources platforms, cybersecurity products, permitting systems and document-management tools.
Its other recommendations will sound familiar to anyone who has built software for the public sector: set central guardrails first, then let departments experiment; start from the problem you're trying to solve, not the product that happens to be available; use AI to support staff judgment, not replace it. For agents that can take actions, it recommends limited permissions to start, human approval for consequential actions, detailed activity logs and a way to switch a system off quickly.
The AI you didn't buy on purpose
Here's why the inventory step matters. Picture a county of forty thousand people. It has no AI policy and no AI budget line. Ask the county manager whether the county uses AI and the honest answer is "not really."
Then look at the software. The document-management system added a summarize button in a spring update. The HR platform now suggests interview questions and screens résumés. The permitting system offers to pre-fill applications from uploaded plans. The email security product uses machine learning to quarantine messages. The 311 vendor added a chatbot to the website last year, and the clerk's office uses a free AI tool to draft meeting minutes.
None of these went through a purchasing decision labeled "AI." Most arrived as a feature in software the county already paid for. Some are on by default. You can't govern a feature you don't know you've switched on.
That's what makes the playbook's advice practical. A policy written in the abstract, for a product nobody has chosen, tends to stay in a binder. A policy written after you've found the summarize button and the résumé screener has something to grip.
Sequence of work
A one-week inventory
- Monday
Pull the list of software contracts and subscriptions from finance and IT.
- Tuesday
Send every vendor the same four questions.
- Wednesday
Ask department heads what they use that isn't on the list.
- Thursday
Record answers in one shared table, with an owner per feature.
- Friday
Mark what's on, what's off, and what nobody can answer yet.
You can't govern a feature you don't know you've switched on.
Four questions for every vendor
The inventory itself is not complicated. For each piece of software on your contract list, ask the vendor the same four questions in writing.
1. Does this product include any AI or machine-learning features, including ones added in updates? Ask them to list each one by name.
2. Which of those are switched on for our account right now, and who can turn them on or off? Some can only be changed by the vendor. That's worth knowing.
3. What data does each feature use, and does any of it leave our environment or train a model? Résumés, permit plans, emails and resident requests are the sensitive ones.
4. Can a person see and override what the feature produces, and is there a record of it? A summary that a clerk reads is different from a screening decision nobody reviews.
Keep the wording identical for every vendor so the answers can be compared side by side, and ask for replies in writing rather than on a sales call. A short written answer is easier to file, and easier to hold a vendor to later.
Then ask department heads a fifth question that vendors can't answer for you: what are your people using that isn't on the contract list? Free tools used for real work are part of the inventory.
Our note on a county AI assistant that found a wrong phone number in an old press release shows what happens when an AI feature draws on content nobody has checked in years.
Give every feature an owner
Put the answers in one table: the software, the feature, on or off, what data it touches, whether a person reviews its output, and a named owner inside the organization. That last column is the one most inventories skip. A feature with no owner is the one that surprises you.
The table does three things straight away. It tells your new oversight group what it is actually overseeing. It shows where the playbook's guardrails for agents, such as limited permissions, human approval and activity logs, already apply, because some vendor features are closer to agents than anyone realized. And it usually exposes one real problem worth solving, like a permitting backlog or a records request queue, which is a better first AI project than anything a vendor demo suggested.
Expect some answers to be "we're not sure." A vendor may not know whether a feature is on for your account, or a department may not know which free tools staff use. Don't let those stall the rest. Put them in their own column with a name and a date to follow up. The inventory is useful the day it's mostly done, and the open questions are often the most important thing on it, because they show where nobody is watching.
The same logic applies to accessibility: our note on how the accessibility deadline covers the pages your vendors run makes the case that what your vendors ship is yours to answer for.
Decision guide
Inventory first or purchase first
| Consideration | Starting with an inventory | Starting with a purchase |
|---|---|---|
| First question | What AI do we already run, and who owns it? | Which AI product should we buy? |
| Policy | Written for the features you actually have | Written for a product you haven't chosen yet |
| Risk you find | A feature switched on by a vendor update | Usually found later, by a resident |
| Next project | Chosen from a real problem the inventory exposed | Chosen because the technology was available |
Start Monday
Pull the contract list from finance and IT. Send the four questions to every vendor on Tuesday. Talk to department heads midweek. Fill in the table by Friday, with an owner for every feature and a short list of questions nobody could answer yet. It won't be perfect, and it doesn't need to be. It needs to exist before the next purchase.
If the inventory turns up a problem worth solving, like a request process residents find hard or a case queue staff can't see into, that's the kind of work our government and public sector team does: intake, case handling and the systems behind them, with permissions and traceability considered from the start. Talk to us.
Related work: Sutton|Place
We build alongside the people doing the work.
Bring us the problem you’re thinking about.