Keystone AI · Custom AI development · Chicago

The Keystone Field Guide to Working AI

Keystone AI designs, builds, and runs custom AI systems for operating businesses.

This guide covers the seven kinds Keystone builds: what each one does, where it lives, and how to handle it.

Assessed in two weeksBuilt in weeks, not quartersRun and measured monthly

Let’s
talk

Every build works for the people already in the seats, and a person approves anything that leaves.

Swapped into production. Nobody's Tuesday got worse.Expedition footage · reel 1
Field notes · Keystone AIPreface

AI earns its place in a business two ways: it lifts the burden off people who have better things to do, and it opens channels that could not exist before.

That is how this guide is organized. Part one is the burden: the collecting, re-keying, chasing, and answering that bury skilled people. Part two is the new channels: work found before it is public, concepts made in an afternoon, a voice available at midnight, a market measured instead of guessed. In both, the person keeps the decision. The plates that follow are the seven kinds of work we have found this true for. Habitats are named by industry. Specimens are shown in conversation.

Landed, measured, and written down. Cause of death: nobody was counting.Expedition footage · reel 2
Sightings log
Entries through 22 Sept. 2026

What is even possible now.

What AI has been seen doing for ordinary businesses, kept to get your own thinking going.

SightedDone, not yet routine
  • A leak triaged from the tenant's video.
  • Month-end arriving as a draft.
  • Forty product concepts, all inside the license.
LegendNot caught yet
  • A jobsite that inspects itself by drone.
  • A quote written from the customer's drawing.
  • A product line proposed by its returns.

Bring a legend of your own. ↗

Read before handling

Handling notes: the unsupervised agent

Dangerous when loose
Temperament
Confident, tireless, and wrong about one thing in twenty. Beautiful animal. Will send the email.
Never
Give it send, submit, or sign. Not as a setting. It should not have the limb.
Always
Feed it a golden set of real examples before it meets live work, and again whenever the model changes.
Approach
Asked first. Then scheduled. Then event-driven, only after the approval log shows it can be trusted.
If bitten
Pause the queue, take the job back, correct the draft. A non-technical person must be able to do all three.

she's a beauty. keep your hands where the gate is.

Part one

Lifting the burden.

The collecting, re-keying, chasing, and answering that bury skilled people. These builds take the pile and hand back the decision.

Plate IOfficium chartarum
Fig. 1. The pile in; one draft out; a person's yes
Fig. 1. Drawn from the field; client screens stay privateHabitat: the office
Plate I · Office work

Office work

Officium chartarum, the paper office

Software that drafts the paperwork under each seat, so the person in it can do the job they were hired for.

What it gives back: the coordinator's afternoons. The estimator's numbers get the estimator's attention instead of the formatting.

HabitatConstruction, facilities, manufacturing, professional firms

checked with the coordinator: the packet used to take her Friday. Now it takes her a yes.

Plate IIVisio operaria
Microscope frame of a soil sample with organisms outlined in green
Fig. 1. A microscope frame, organisms outlined
Inventory table listing items, quantities, confidence and the number of video frames behind each line
Fig. 2. An inventory drafted from a walk-through video
Report card showing 27 organisms counted and 19 marked uncertain for a person to check
Fig. 3. The count, and what it is unsure of
Fig. 1 to 3. Outlined, itemized, counted; each awaiting a personHabitat: photos, video, microscopy
Plate II · Seeing

Seeing

Visio operaria, the working eye

AI that reads what your people photograph, so nobody has to type what the camera already saw.

What it gives back: the hour after the walk-through. The crew films the room once and the record writes itself for them to confirm.

HabitatRestoration, soil labs, facilities, landscaping

ask: what would the adjuster need to see? build for that, not for the demo.

Plate IIIVox conversans
Fig. 1. A conversation, and the hand-off
Fig. 1. Drawn from the field; client screens stay privateHabitat: voice, chat, coaching
Plate III · Talking

Talking

Vox conversans, the conversing voice

Conversational agents that hold a real conversation, remember it, and know when to hand off to a person.

What it gives back: the calls nobody could take. A coach available at midnight. A phone answered during the spring rush. A person reached when it matters.

HabitatCoaching, consumer apps, service businesses

the good ones say “let me get someone” without making it awkward.

Part two

Channels that could not exist before.

Work found before it is public. Concepts made in an afternoon. A market measured instead of guessed. These builds are not faster versions of an old job; they are jobs nobody could staff.

Plate IVInventio operis
Fig. 1. Next year's work, marked before it is public
Fig. 1. Drawn from the field; client screens stay privateHabitat: permits, filings, trends
Plate IV · Finding work

Finding work

Inventio operis, the finder of work

Standing intelligence that reads the sources where next year's work shows up first, and hands your people a ranked list.

What it gives back: a business development function for a company that never had one. Nobody was reading forty municipal agendas a week. Now something is, and a person decides what to chase.

HabitatConstruction, facilities, consumer products

everyone hears about the job when it hits the bid board. this hears about it before.

Plate VFabrica imaginum
Fig. 1. The brief, as a spec
Fig. 2. Variant
Fig. 3. Variant
Fig. 4. Variant
Fig. 5. Parts, from a description to a supplier
Fig. 1 to 5. Brief, variants, and the partsHabitat: images, video, 3D
Plate V · Making

Making

Fabrica imaginum, the image works

Generation pipelines with judgment built in, so a designer reacts to forty concepts instead of drawing one.

What it gives back: the afternoon the team never had. A brief becomes a wall of judged options by lunch; the designer spends her skill on choosing and refining, not on the first draft.

HabitatConsumer products, toys, apparel, marketing, manufacturing

the prompt box is where most generation projects die. the pipeline is where they live.

Plate VIMachina mercatus
Fig. 1. An ad campaign run as an experiment
Fig. 1. The experiment, countedHabitat: sites, search, ads, content
Plate VI · Marketing machinery

Marketing machinery

Machina mercatus, the market engine

The machinery behind marketing, engineered and measured, so a small company can run what only big ones could staff.

What it gives back: a marketing department's output without a marketing department. A site that gets found, content that gets published, ads run as experiments, all counted.

HabitatAny business that has bought marketing and cannot say what it returned

“awareness” is not a unit. count the conversations.

Plate VIIDisciplina operantis
Fig. 1. Seedling in a pot; grows the next build
Fig. 1. No screenshot; this one happens in your conference roomHabitat: your conference room
Plate VII · Training

Training

Disciplina operantis, the operator's craft

Your people become AI operators at the level their role requires.

What it gives back: capability that stays in the building. Each session leaves your people able to do more, and usually pointing at the next thing worth building.

HabitatOffice, field, leadership

the goal is not that your people understand AI. it is that they get their Tuesday back.

Anatomy of a build

Cutaway of a specimen, any specimen.

Every plate above is built on the same skeleton. Most of it is not AI, and all of it ends at a person. That is the point.

  • a
    Connections

    Read-only, into the systems you already use. Write access is a decision, not a default.

  • b
    Plain software

    Collecting, matching, formatting, routing. Written once, verified by a person, trusted like a jig.

  • c
    Judgment organs

    The agents. Small, typed, cited. Used only where a clerk with a checklist would have to stop and think.

  • d
    Memory

    An answer library that compounds with every job. Half the value of the first build; the reason the second costs less.

  • e
    The gate

    Output is a draft. The build cannot send, submit, or sign. Enforced by the skeleton, not promised by a policy.

  • f
    Evaluation

    A golden set of real examples the build must pass before it goes live, and keep passing when models change.

Worst-case scenarios
Survival handbook, business edition

How to survive the usual dangers.

Numbered instructions for situations that come up more often than anyone admits. Illustrated for use under stress.

How to survive an AI vendor demo

  1. Ask to see it run on your own documents, not theirs. A demo that cannot survive your PDFs is a slideshow.
  2. Ask for the approval log and the human override. If the presenter says "it's fully autonomous," that is the danger, not the feature.
  3. Ask which parts are AI and which are plain software. If the answer is "it's all AI," the reliability argument has just left the room.
  4. Ask what it cannot do. A system that can do anything can do the wrong thing.
  5. Do not sign in the room.

a demo is a legend. ask to see it landed.

How to escape a runaway automation

  1. Find the pause. If there is no pause a non-technical person can press, you are not escaping, you are waiting.
  2. Take the job back into a person's hands. A good system lets you do this per item.
  3. Read the log to find what it did and when. If there is no log, assume everything.
  4. Correct the draft, not the policy. Then feed the correction back as a test case so it cannot happen the same way twice.
  5. Ask why it had the ability to act at all. Remove the limb, not just the behavior.

the gate is not a setting. it is a missing limb, on purpose.

How to tell an agent from a workflow in the dark

  1. Feed it the same input twice. A workflow does the same thing both times. An agent might not.
  2. Give it a document it has never seen. A workflow breaks loudly. An agent produces something plausible; check whether it is right.
  3. Ask who chose the next step. If a person wrote the path in advance, workflow. If the model chose at run time, agent.
  4. Ask what it costs to run. Near zero is a workflow. Per-run model usage is an agent.
  5. The safe animal is the hybrid: workflow skeleton, agents only at the judgment points, a person at the gate.

How to survive the pile on Monday

  1. Name the seat the pile lands on. Not the department; the person.
  2. Count the hours. Collecting, re-keying, chasing, formatting. Write the number down.
  3. Multiply by what that hour costs. That is the build's budget and its deadline.
  4. Build for that seat only. One job, real data, a person approving. Measure against the number.
  5. Only then look at the next seat. Piles are patient. Budgets are not.

start with one seat. it is arithmetic, not courage.

The expedition

One job first. Then whatever earns its keep.

The route is arithmetic: the hours or the revenue a job is costing you, against what a build returns.

Assess. Where the hours go, seat by seat. What your software already does. How work reaches you. A ranked build list you can act on with or without Keystone.

Build. The top item on the list, shipped as production software on real data with real logins, measured against a number agreed before the build starts.

Expand. Hours returned, drafts approved, leads qualified. When a build pays for itself, the next goes in beside it. When it doesn't, you can walk.

The keeper

Founder-led, in the field.

David Lai, Keystone's founder, does the assessment in your office and writes the first build himself. Before Keystone he carried a quota selling to multi-site businesses, so the pipeline is understood from the inside, not from a deck.

— kept in the field, Chicago

  • Built for the people in the seats

    Every build starts from someone's actual Tuesday and ends with that person deciding more and re-keying less.

  • Reported in counts

    Every claim names a unit. If it can't be counted, it doesn't go in the report.

  • Built to hand over

    Each client gets a separate instance and a written offboarding process. Your code, your data, your accounts.

  • Deterministic where possible

    Agents only where judgment is required. Enforced by the architecture, not promised by a policy.

Why "Keystone" Keystone oak leaf mark

Small presence, outsized effect.

In nature, keystone species hold entire ecosystems together. An oak feeds hundreds of species. A beaver builds habitat for thousands.

Native ecosystems are the founder's lifelong interest, and they shaped how Keystone builds software. A business rarely needs to be rebuilt. It needs a small, well-placed thing at the spot where everything else leans. Hence the field guide.

One focused build, placed where the work piles up.

Correspondence

Which job in your company is buried?

Tell us about it. Thirty minutes on where the hours go, how work reaches you, and what a first build would look like.

Let’s
talk

or write to david@keystoneai.io

Set in Fraunces, Archivo, and IBM Plex Mono. Plates drawn in ink and code. Chicago, MMXXVI.