UntrenchTalk to us →
No. 001 — 2026
For local and state government

Most AI projects
don’t work.

Ours do, because we do
the unglamorous parts.

The pitch

Untrench builds agentic systems for cities, counties, and the agencies that serve them. We focus on the work where being wrong is expensive. Permits. Grants. RFPs. Code. Constituent mail. If a chatbot can handle it, use a chatbot. If it has to be right, that’s us.

Try a live demo →Tell us your problem
Built by

Polco. Three decades and ~1,000 cities of civic data and trust. We know what good looks like in local government, and what gets sold to it that isn’t.

Built for
  • City & county staff
  • Planning & permitting
  • Grants & budget offices
  • Procurement & legal
  • Council & constituent services
More capacity for the work that matters◊Built for cities, counties, and special districts◊Your codes, not the internet’s average◊99.5% beats 75%, especially under public records law◊~1,000 cities served, 50 states, 30+ years◊If a chatbot solves it, use a chatbot◊
§ 01 / Approach

There are two kinds of AI work.
We do one of them.

Option A · free, fast, fine

You paste it
into a chatbot.

  • +Works most of the time. Free or close to it.
  • +Great for drafts, summaries, and low-stakes lookups.
  • −Hallucinates code citations under public records request.
  • −Doesn’t know your municipal code, your funders, your forms, your council.
  • −No audit trail. No integrations. No one to call when it’s wrong.
Use this for the council packet.
Option B · what we do

You hire the team that already
knows local government.

  • →Right answers, with citations to your code, your minutes, your ordinances.
  • →Tailored to your jurisdiction, your funders, your prior decisions, your voice.
  • →Real interfaces, not a chat box. Defensible records. Reviewable diffs.
  • →Evals, guardrails, and a human in the loop wherever the law requires one.
  • →Built to be 99.5% right. And to tell you, plainly, when it isn’t.
For the work that ends up in the public record.
§ 02 / Why us

We knew we
had to build this.

We’re Polco. For 30 years we’ve helped nearly 1,000 cities, in every state, understand what their residents actually think and make better decisions with real data. That’s our day job, and we’re good at it.

We didn’t set out to build agentic AI products. We started Untrench because we kept watching the same thing happen: vendors selling local governments expensive, brittle “AI solutions” that hallucinate code citations, ignore the actual jurisdiction, and produce work nobody on staff would sign their name to. Cities deserve better than that. We knew we were the right people to build it.

The name is the promise. Cities don’t have a time problem. They have a capacity problem. Too few staff, too much work, no room to do anything new. We give that capacity back, built on your codes, your prior decisions, your funders, your voice, and the institutional knowledge we’ve earned across a thousand civic relationships.

Polco, in numbers
  • 30+
    years of civic data
  • ~1,000
    cities served
  • 50
    states represented
  • 1
    reason we built this
§ 03 / Examples

Five things we’ve built
for cities like yours.

Each one is a real working agent in a real local-government workflow. They aren’t videos and they aren’t mockups. You’ll be poking at the actual system, with the actual reasoning, on real example data. Permit review is live now. The rest are in active builds.

01
Live
Critique · permitting

Permit review

Reads a permit submission against the actual code in force in your jurisdiction. Flags conflicts with citations to your municipal code. Drafts the reviewer’s letter.

Try the demo↗
PERMIT-2026-0114 · ADU · Berkeley, CA
Issue · setback
Rear setback shown at 3’-6”. BMC §23.304.060.A requires 4’-0” minimum for ADUs over 800 sq ft.
Issue · egress
Bedroom 2 window sill at 46”. CRC R310.2.2 requires ≤44” from finished floor.
Note
Solar-ready provisions appear satisfied per CALGreen §4.201.1.
02
In build
Search + draft · grants

Grant finder & writer

Reads your city’s capital plan and prior awards. Surfaces federal, state, and foundation grants you actually qualify for, with fit reasoning, then drafts the narrative in your voice, not in AI voice.

Coming soon
CITY · pop. 38,400 · capital plan FY26
USDOT Safe Streets & Roads for Allfit · 0.94
EPA Community Change Grantsfit · 0.88
FEMA BRIC pre-disaster mitigationfit · 0.81
HUD CDBG rural set-asidefit · 0.42 (pop. cap)
Draft · project narrative §2
“The Maple Street corridor, identified in our 2024 Local Road Safety Plan as the city’s highest-injury arterial, aligns directly with USDOT’s 2026 SS4A priority on systemic redesign in communities under 50,000…”
03
In build
Read + draft · procurement

RFP response

For when your city is the one responding, or evaluating. Reads the RFP, builds the compliance matrix, flags the requirements you don’t meet, and drafts response sections weaving in your past wins.

Coming soon
RFP-2026-DOT-441 · compliance matrix
L.4.2.1
SOC 2 Type II (current)
MET
L.4.2.2
FedRAMP Moderate ATO
GAP — 4mo
L.4.3.1
Past performance, >$2M, public sector
MET ×3
M.5.1
Section 508 conformance report
MET
04
In build
Triage + respond · constituent mail

Constituent triage

For council offices, mayors, and 311 lines that get more mail than they can answer. Classifies casework, policy comment, and FOIA. Routes correctly. Drafts replies in your office’s voice. Every decision logged, every draft reviewable.

Coming soon
INBOX · district 7 · today
CASEWORKResident — SNAP recertification not received→ benefits team
POLICYComment on draft zoning amendment→ policy + log
FOIARecords request — 2024 contracts >$50k→ counsel
THANKS“thank you for the town hall last week…”→ auto-reply
847 processed · 23 escalated · 0 dropped
05
In build
Resolve · municipal code

Code conflict audit

Reads your full municipal code and surfaces what contradicts what, what’s redundant, and what just got quietly amended last Tuesday. Every flag has a citation chain to the source ordinance.

Coming soon
MUNI CODE · audit · 14,221 sections
Direct contradiction · 1
§17.44.030 requires 1 parking space per ADU. §17.44.090 (added 2024) prohibits parking minimums in transit zones. Same parcel, two rules.
Redundant clauses · 6
Definition of “habitable space” is restated, with minor variation, across §15, §17, and §22. Recommend consolidation.
Stale references · 3
§19.12 cites 2018 IBC; jurisdiction adopted 2022 IBC in March.

Note · Permit review is live now. The other four are in active builds. Static previews above are sized to real outputs.

§ 04 / Other shapes

The five above are examples,
not a menu.

Most of what we build is shaped to a specific city, county, or agency. The patterns repeat (review, draft, triage, audit), but your code, your forms, your funders, and your voice are always the source of truth.

Custom build

Tell us what part of your week
you’d hand off if you could.

Two questions, no demo deck. If it’s the kind of work we do well, we’ll say so. If it isn’t, we’ll point you somewhere better.

Start there →
City-wide rollout

Rolling this out
across departments?

Multi-department deployments, integrations with your existing systems (GIS, permitting, ERP, CRM), procurement support, and the evals to prove it works at scale. We do those.

Let’s talk →
§ 05 / What we believe

Five things we’ll
say out loud.

i.

Most “AI strategy” is theater.

Pilots that never ship. Demos rehearsed on cherry-picked inputs. Decks that say “agentic” forty times. We’re not interested. We ship things people use on Tuesdays.

ii.

Without your data, the AI doesn’t work.

A general model that knows the average of the internet will give you the average of the internet. That’s exactly what you don’t need. Your municipal code, your prior reviews, your funder relationships, your reviewer’s voice: that’s where the accuracy actually lives, and connecting it correctly is the part most vendors get wrong. The AI is the easy part. The data plumbing is the job, and it takes real expertise to get right.

iii.

A chat box is not a product.

Sometimes the right interface is a form. Or a queue. Or a dashboard. Or a reviewable diff. We build whatever shape the work actually has, and we don’t default to a chat window because it was easy.

iv.

Evals or it didn’t happen.

If we can’t measure how often it’s right, we don’t ship it. If the number drops after a model update, we know within an hour. The unsexy infrastructure is the whole job.

v.

The boring parts are the work.

Schemas. Permissions. Audit logs. Fallbacks. Versioning. The stuff nobody puts on a slide. That’s where 75% becomes 99.5%, and that’s what you’re hiring us for.

§ 06 / Pricing

Clear pricing,
stated up front.

We don’t do “contact us for pricing.” You’re running a city. You don’t have time for a five-week sales cycle to find out if we’re in your budget. Here’s how it works:

Step 1 · one-time
Required

Implementation

We tailor the agent to your jurisdiction. Your code, your forms, your prior decisions, your reviewers’ voice. Data integrations included.

One fee
Quoted in writing after a 30-min scoping call. No surprises.
  • →Agent configured against your municipal code in force
  • →Integrations with your existing systems (permitting, GIS, CRM, ERP)
  • →Voice & format calibrated against your prior reviewer letters
  • →Eval suite built on your real submittal history
  • →Staff training and rollout support
Step 2 · ongoing
Your choice

Hosting

Once it’s built, the agent is yours. Run it on your own infrastructure, or let us run it for you.

Self-hosted
$0
per year, ongoing
You take the code and the model config. Run it where you want.
We host
$5K–15K
per year, all-in
Range depends on agent complexity, data volume, and uptime needs.
  • →Hosted plan includes infrastructure, model costs, monitoring, and updates
  • →Continuous evals against your data. We see model drift before you do.
  • →SLA, support, and a real human to call
  • →Switch to self-hosted any time. The code is yours either way.
Do the math

One reviewer’s hour per week, at fully-loaded cost, pays for the hosted plan. The cities we work with typically get back five to fifteen. We’re happy to walk through your specific numbers on a call.

House rules
No per-seat fees.
Add every employee in your city. We won’t charge more.
No per-token fees.
Use it as much as you want. Model costs are our problem on the hosted plan.
No surprise increases.
Multi-year pricing locked at signing. We renegotiate openly or not at all.
§ 07 / FAQ

Reasonable
questions.

Why shouldn't we just use ChatGPT for this?+

If the work doesn't punish a 25% error rate, please do. ChatGPT is great for that. We're for the work where being wrong is expensive: in council embarrassment, in litigation risk, in records requests, in dollars. And being right requires knowing things a general model doesn't: your code, your prior decisions, your forms, your funders.

How does this fit our procurement?+

We're available through several cooperative purchasing agreements, can respond to RFPs directly, and have done piloted-then-purchased deployments before. Polco has been through this thousands of times. We'll send the procurement person to the procurement person.

How long does a project take?+

A tailored deployment of one of our existing products (e.g. permit review configured for your code) is usually 4–8 weeks to a working system. Custom agents run 8–16 weeks. City-wide rollouts are quarter-scale projects. We’d rather give you a real number after a 30-minute call than a fake one now.

Who owns the model? The data? The outputs?+

You do. We don't train on your data. We don't sell it. We don't pool it across jurisdictions. Outputs are subject to your records-retention rules, not ours, and we build with public records law in mind from day one.

What happens when the underlying models change?+

Every system we build comes with an eval suite that runs continuously against your data. When a model update shifts behavior, we see it before you do. That's not a feature; it's the only responsible way to run this kind of software for a public agency.

Are you going to tell us we need a chatbot?+

No. We might, sometimes, build something with a conversational element because it's genuinely the right interface for one part of the workflow. But the product is the workflow. The review, the draft, the queue. Not the chat.

§ 08 / Contact

Tell us what your week
is full of.

Four short questions. A real human reads every one. If we’re a fit, we’ll say so within two business days. If we’re not, we’ll tell you that too, and probably point you toward someone who is.

No sales sequence. No drip campaign. Just a reply.