AI employee architecture / Technical reference

TECHNICAL REFERENCE

The technical system behind an AI employee.

Inspect how work moves through context, procedures, approved capabilities, human decisions, review records, and controlled improvement.

RETURN TO THE BUYER VIEW BOOK A WORKING SESSION >

ONE SYSTEM · THREE GUIDED VIEWS

Inspect how the whole system carries the work.

Choose a view, apply an illustrative workflow, and open any node for the business outcome, technical mechanism, and review evidence.

CHOOSE A VIEW

01Run the work 02Control the work 03Improve the system

APPLY A WALKTHROUGH

RFQ + estimating Shared inbox intake Compliance answer

MANUFACTURING

RFQ received to approved estimate draft

A request arrives with drawings and a due date. The AI employee assembles the job history, follows the estimating procedure, checks approved systems, and pauses when an assumption needs an estimator.

ILLUSTRATIVE WALKTHROUGH · NOT A CLIENT RESULT

EXPLORE HARNESS ENGINEERING >RELATED PRODUCTION WORK >

Run the work

Request arrivesThe shared inbox receives drawings, quantities, and the requested delivery date. The queue assigns an owner and due time.
Relevant job history is assembledCurrent specifications, similar prior jobs, rate cards, and open questions are retrieved with source dates attached.
The estimating procedure runsThe AI employee follows the approved checklist, identifies missing inputs, and prepares the work for specialist review.
Approved systems are queriedRead-only connectors retrieve inventory, supplier, and prior-job data required for the draft.
The estimator reviews the exceptionsA missing finish specification and any pricing assumption are routed to the estimator before the draft can move forward.
The draft returns to the queueThe estimate draft, open assumptions, citations, and approval history are stored with the next action and owner.

Control the work

The estimating role receives read access to job history and inventory for this request.
Retrieval stays within approved suppliers, rate cards, and prior-job records.
Inventory and prior-job lookups run through approved read-only connectors.
Pricing assumptions pause with the designated estimator before the draft advances.
The system retains the sources, open assumptions, approval, and next owner.
An interrupted estimate can resume from the last recorded checkpoint without repeating approved work.

Improve the system

Estimator edits and approval outcomes could become signals for a later system review.
Repeated missing finish specifications could surface as a recurring intake gap.
The system could propose a required-field check for the estimating procedure.
An estimating owner would evaluate the proposed check and the examples behind it.
An approved check could be released as a named procedure version with a rollback point.

PROFESSIONAL SERVICES

Client inquiry to partner-reviewed response

A new inquiry reaches the shared inbox. The AI employee collects the required facts, applies the intake procedure, drafts from approved precedent, and escalates the relationship decision.

ILLUSTRATIVE WALKTHROUGH · NOT A CLIENT RESULT

EXPLORE SKILLS AND GATEWAYS >

Run the work

Inquiry enters the shared queueThe email, attachments, sender, and response deadline become one tracked intake item.
The client briefing is assembledThe AI employee retrieves service criteria, relevant precedent, and any permitted relationship history.
The intake checklist runsRequired details are extracted, missing questions are drafted, and the inquiry is routed to the right practice owner.
Approved records are checkedScoped connectors read the CRM, precedent library, and scheduling system without exposing unrelated records.
The practice owner decidesThe AI employee presents the facts, open risks, and a draft response. The owner approves, edits, or declines.
The response and handoff are recordedThe approved draft, decision basis, and next owner return to the shared queue before anything is sent.

Control the work

The intake role receives access to the shared queue and permitted relationship records.
The briefing excludes unrelated client matters and records outside the requester's scope.
CRM, precedent, and scheduling lookups run through scoped connectors.
The relationship decision remains with the designated practice owner.
The system retains the approved draft, decision basis, and handoff state.
A paused inquiry can resume from its last reviewed state with the same supporting record.

Improve the system

Partner edits and routing changes could become signals for a later intake review.
Repeated requests for the same missing detail could reveal a checklist gap.
The system could propose a new intake question and routing instruction.
A practice owner would assess the proposal against representative inquiries.
An approved revision could enter service as a monitored checklist version.

REGULATED KNOWLEDGE WORK

Policy question to cited answer

A team member asks how a current obligation applies. The AI employee retrieves the governing sources, resolves the relevant relationships, drafts a cited answer, and routes it for review.

ILLUSTRATIVE WALKTHROUGH · NOT A CLIENT RESULT

EXPLORE CONTEXT ENGINEERING >RELATED PRODUCTION WORK >

Run the work

The question is capturedThe request, business context, jurisdiction, and response deadline enter a controlled work queue.
Governing sources are retrievedCurrent policies, agreements, amendments, and related entities are assembled with citations and effective dates.
The review procedure runsThe AI employee applies the approved issue checklist and separates supported findings from unresolved questions.
Authorized repositories are queriedSearch and knowledge connectors read the approved corpus and return only records within the requester's scope.
A reviewer checks the interpretationThe cited draft and unresolved points go to the designated reviewer before the answer is distributed.
The answer remains traceableThe final answer keeps its citations, review history, and the source versions used at the time.

Control the work

The compliance role receives access to the approved corpus for the requester's scope.
Retrieval filters limit sources by jurisdiction, effective date, and permitted repository.
Search and relationship resolution run inside the approved knowledge environment.
Interpretive findings pause with the designated reviewer before distribution.
The answer retains citations, source versions, reviewer changes, and approval state.
A later review can return to the cited source set and the prior approval checkpoint.

Improve the system

Reviewer corrections and citation changes could become signals for a later system review.
Repeated source-version conflicts could reveal a retrieval or procedure gap.
The system could propose an effective-date check for the review procedure.
A policy owner would evaluate the change against approved examples and exceptions.
An approved check could be released as a monitored procedure version with rollback history.

THE OPERATING LOOP

Follow one responsibility from request to result.

The architecture stays fixed while the selected workflow changes the example inside each step.

REQUEST, EVENT, OR SCHEDULEDELIVERABLE + WORK RECORD

01Work arrivesChannels + triggers + schedulesRequest arrivesOPEN
02Build the briefingContext + session and durable memoryRelevant job history is assembledEXPLORE
03Apply the procedureSkills + orchestrationThe estimating procedure runsEXPLORE
04Use approved capabilitiesTools + connectors + delegationApproved systems are queriedEXPLORE
05Reach a decisionRules + approvals + escalationThe estimator reviews the exceptionsEXPLORE
06Return the resultState + trace + checkpointsThe draft returns to the queueEXPLORE
MModel intelligenceReasoning engineThe model reasons over the assembled job briefing and drafts the next estimating stepEXPLORE

Work arrives

Channels + triggers + schedules

The AI employee receives a message, file, system event, or scheduled responsibility through a channel your team already uses.

HOW IT WORKS

Gateways and event adapters turn each request into a tracked responsibility with an owner and due state.

IN THIS WALKTHROUGH

Request arrives. The shared inbox receives drawings, quantities, and the requested delivery date. The queue assigns an owner and due time.

WHAT YOU CAN INSPECT

Authorized sender, trigger, channel, and task owner

OPEN THE DEEP DIVE >

Build the briefing

Context + session and durable memory

Current source material, prior decisions, and live task state are assembled before the model responds or acts.

HOW IT WORKS

Retrieval, session recall, and durable memory assemble only the material required for the responsibility.

IN THIS WALKTHROUGH

Relevant job history is assembled. Current specifications, similar prior jobs, rate cards, and open questions are retrieved with source dates attached.

WHAT YOU CAN INSPECT

Sources included, dates, exclusions, and prior state

OPEN THE DEEP DIVE >

Apply the procedure

Skills + orchestration

Versioned instructions define the steps, checks, handoffs, and specialist roles used to carry the work forward.

HOW IT WORKS

A reusable skill coordinates the approved sequence, checks, and specialist handoffs for the work.

IN THIS WALKTHROUGH

The estimating procedure runs. The AI employee follows the approved checklist, identifies missing inputs, and prepares the work for specialist review.

WHAT YOU CAN INSPECT

Procedure version, assigned role, checks, and next step

OPEN THE DEEP DIVE >

Use approved capabilities

Tools + connectors + delegation

Scoped connectors let the AI employee read or update only the systems and records required for the task.

HOW IT WORKS

A governed capability library exposes approved tools, connectors, and specialist tasks to the procedure.

IN THIS WALKTHROUGH

Approved systems are queried. Read-only connectors retrieve inventory, supplier, and prior-job data required for the draft.

WHAT YOU CAN INSPECT

System, permission, requested operation, and result

OPEN THE DEEP DIVE >

Reach a decision

Rules + approvals + escalation

Rules define what the AI employee may complete, where it must pause, and who has authority to approve an exception.

HOW IT WORKS

Policy checks compare the work with defined authority and route exceptions to the designated reviewer.

IN THIS WALKTHROUGH

The estimator reviews the exceptions. A missing finish specification and any pricing assumption are routed to the estimator before the draft can move forward.

WHAT YOU CAN INSPECT

Boundary reached, exception raised, approver, and decision

OPEN THE DEEP DIVE >

Return the result

State + trace + checkpoints

Task state, sources, tool calls, approvals, and the final result remain available for review and the next handoff.

HOW IT WORKS

Persistent state and checkpoints retain the result, supporting evidence, current status, and next owner.

IN THIS WALKTHROUGH

The draft returns to the queue. The estimate draft, open assumptions, citations, and approval history are stored with the next action and owner.

WHAT YOU CAN INSPECT

Status, evidence, actions, approvals, and next owner

OPEN THE DEEP DIVE >

Model intelligence

Reasoning engine

The model interprets the assembled briefing and helps produce the next useful step.

HOW IT WORKS

The surrounding system supplies context, instructions, tools, and decision boundaries before requesting reasoning.

IN THIS WALKTHROUGH

The model reasons over the assembled job briefing and drafts the next estimating step.

WHAT YOU CAN INSPECT

Briefing supplied, requested task, model response, and next action

EXPLORE HARNESS ENGINEERING >

THE CONTROL PLANE

See the boundaries around every action.

Identity, data scope, execution policy, human authority, and recovery remain inspectable across the responsibility.

DEFINED AUTHORITYREVIEWABLE STATE

C1Scoped identityCredentials + permissionsThe estimating role receives read access to job history and inventory for this requestOPEN
C2Data boundariesRetrieval scope + secret handlingRetrieval stays within approved suppliers, rate cards, and prior-job recordsEXPLORE
C3Controlled actionIsolated execution + tool policiesInventory and prior-job lookups run through approved read-only connectorsEXPLORE
05Reach a decisionRules + approvals + escalationPricing assumptions pause with the designated estimator before the draft advancesEXPLORE
06Return the resultState + trace + checkpointsThe system retains the sources, open assumptions, approval, and next ownerEXPLORE
C6Review and recoveryTrace + checkpoints + rollbackAn interrupted estimate can resume from the last recorded checkpoint without repeating approved workEXPLORE

Scoped identity

Credentials + permissions

Each responsibility operates with a defined identity and the minimum access required for its work.

HOW IT WORKS

Credential stores and role policies scope which systems and operations are available to the AI employee.

IN THIS WALKTHROUGH

The estimating role receives read access to job history and inventory for this request.

WHAT YOU CAN INSPECT

Identity used, permission granted, expiration, and requested operation

EXPLORE SECURITY >

Data boundaries

Retrieval scope + secret handling

The system limits which records enter the briefing and keeps credentials outside the working context.

HOW IT WORKS

Retrieval filters, tenant boundaries, and secret references constrain data before the model receives it.

IN THIS WALKTHROUGH

Retrieval stays within approved suppliers, rate cards, and prior-job records.

WHAT YOU CAN INSPECT

Repositories searched, filters applied, records excluded, and secret reference

EXPLORE CONTEXT ENGINEERING >

Controlled action

Isolated execution + tool policies

Approved actions run inside a bounded environment with explicit tool and network policies.

HOW IT WORKS

Execution boundaries separate the task from unrelated systems and restrict the operations a tool can perform.

IN THIS WALKTHROUGH

Inventory and prior-job lookups run through approved read-only connectors.

WHAT YOU CAN INSPECT

Runtime boundary, tool policy, network scope, operation, and result

EXPLORE LOCAL AI >

Reach a decision

Rules + approvals + escalation

Rules define what the AI employee may complete, where it must pause, and who has authority to approve an exception.

HOW IT WORKS

Policy checks compare the work with defined authority and route exceptions to the designated reviewer.

IN THIS WALKTHROUGH

Pricing assumptions pause with the designated estimator before the draft advances.

WHAT YOU CAN INSPECT

Boundary reached, exception raised, approver, and decision

OPEN THE DEEP DIVE >

Return the result

State + trace + checkpoints

Task state, sources, tool calls, approvals, and the final result remain available for review and the next handoff.

HOW IT WORKS

Persistent state and checkpoints retain the result, supporting evidence, current status, and next owner.

IN THIS WALKTHROUGH

The system retains the sources, open assumptions, approval, and next owner.

WHAT YOU CAN INSPECT

Status, evidence, actions, approvals, and next owner

OPEN THE DEEP DIVE >

Review and recovery

Trace + checkpoints + rollback

A reviewer can see what happened, resume interrupted work, or return to an earlier approved state.

HOW IT WORKS

Traces and checkpoints preserve inputs, decisions, state transitions, and restorable versions.

IN THIS WALKTHROUGH

An interrupted estimate can resume from the last recorded checkpoint without repeating approved work.

WHAT YOU CAN INSPECT

Checkpoint, interruption, prior state, recovery action, and reviewer

READ ABOUT REVIEW RECORDS >

A GOVERNED LEARNING LOOP

See how review signals could become a safer procedure.

This illustrative view shows a possible path from outcomes to a reviewed, versioned system update.

ILLUSTRATIVE CAPABILITY · NOT A PRODUCTION CLAIM

WORK RECORD + REVIEWER FEEDBACKAPPROVED MEMORY + PROCEDURE VERSION

I1Observe outcomesResults + reviewer feedbackEstimator edits and approval outcomes could become signals for a later system reviewOPEN
I2Find a repeatable patternPattern detection + evaluationRepeated missing finish specifications could surface as a recurring intake gapEXPLORE
I3Propose an updateMemory + skill refinementThe system could propose a required-field check for the estimating procedureEXPLORE
I4Review the changeHuman validation + approvalAn estimating owner would evaluate the proposed check and the examples behind itEXPLORE
I5Version and monitorRelease + observation + rollbackAn approved check could be released as a named procedure version with a rollback pointEXPLORE

Observe outcomes

Results + reviewer feedback

Completed work and reviewer changes can reveal where the operating system needs attention.

HOW IT WORKS

Outcome records collect corrections, exceptions, and approval decisions for later analysis.

IN THIS WALKTHROUGH

Estimator edits and approval outcomes could become signals for a later system review.

WHAT YOU CAN INSPECT

Result, reviewer change, exception, and source record

Find a repeatable pattern

Pattern detection + evaluation

Repeated corrections can be separated from one-off preferences before any system change is proposed.

HOW IT WORKS

Evaluation compares similar outcomes and identifies a bounded, evidence-backed opportunity for improvement.

IN THIS WALKTHROUGH

Repeated missing finish specifications could surface as a recurring intake gap.

WHAT YOU CAN INSPECT

Examples reviewed, recurrence, exclusions, and confidence basis

Propose an update

Memory + skill refinement

The system can draft a specific memory or procedure change without activating it.

HOW IT WORKS

A proposed update names the affected memory or skill, its evidence, and the expected behavior change.

IN THIS WALKTHROUGH

The system could propose a required-field check for the estimating procedure.

WHAT YOU CAN INSPECT

Proposed diff, evidence, affected responsibilities, and expected outcome

READ ABOUT MEMORY AND SKILLS >

Review the change

Human validation + approval

An accountable owner evaluates the proposal, its evidence, and its operating impact.

HOW IT WORKS

The review gate supports approval, revision, rejection, or a limited trial before release.

IN THIS WALKTHROUGH

An estimating owner would evaluate the proposed check and the examples behind it.

WHAT YOU CAN INSPECT

Reviewer, decision, edits, test scope, and approval record

READ ABOUT APPROVALS >

Version and monitor

Release + observation + rollback

An approved update can enter service as a named version with monitoring and a recovery path.

HOW IT WORKS

Versioned activation records what changed, where it applies, how it performs, and how to reverse it.

IN THIS WALKTHROUGH

An approved check could be released as a named procedure version with a rollback point.

WHAT YOU CAN INSPECT

Version, release scope, observed results, owner, and rollback point

PRODUCTION EVIDENCE

Compare the reference with current systems.

Production engagements stay distinct from the illustrative workflows used inside the map.

PRODUCTION ENGAGEMENT · MANUFACTURING

Infinite AI OS: an AI operating system in 90 days

Four named AI employees working in Microsoft Teams on a six-layer system: company memory, knowledge map, file intake, employee profiles, tool connectors, and observability.

4 AI employees live · 90 days to foundationVIEW PRODUCTION EVIDENCE >

PRODUCTION ENGAGEMENT · HEALTHCARE ENTERPRISE

AgentHub: contracts you can question

A shared prompt library paired with a contract-review agent that uses hybrid retrieval with field-level citations.

Field-level citations · 50-document grounded corpusVIEW PRODUCTION EVIDENCE >

TECHNICAL LIBRARY

Open the part you need to inspect.

Six focused explainers connect the map to implementation choices.

01Context engineering

Deciding exactly what your AI sees: the right data, structured and retrievable, at the right moment.

EXPLORE >

02Harness engineering

The scaffolding around a model (skills, tools, memory, and guardrails) that turns raw intelligence into dependable work.

EXPLORE >

03Hybrid RAG

Your unstructured files, converted into vector databases and knowledge graphs your AI can actually reason over.

EXPLORE >

04AI employees

Agents that live where your team works (Teams, Slack, email) and carry whole workflows, not just chats.

EXPLORE >

05Skills & gateways

Your procedures, written once as reusable skills, shared by every agent, delivered through the channels you already run on.

EXPLORE >

06Local AI

Models on your own hardware: private by default, with a flat monthly cost instead of a metered bill.

EXPLORE >

SUPPORTING ARTICLES

Read the operating details.

VIEW ALL ARTICLES >

ARCHITECTURE · 8 MIN READ

What makes an AI employee work

Follow the six responsibilities that turn a model into a system that can carry a company workflow.

READ THE ARTICLE >

ARCHITECTURE · 7 MIN READ

Context, memory, and skills do different jobs

See what belongs in a task briefing, durable memory, and a reusable procedure, plus the failure mode of mixing them together.

READ THE ARTICLE >

CONTROLS · 8 MIN READ

Permissions, approvals, and audit trails

A plain-language guide to the boundaries that determine what an AI employee can access, decide, and leave behind for review.

READ THE ARTICLE >

Bring one workflow. We will sketch the system around it.

A 30-minute working session maps the work, required access, human decisions, evidence, and likely architecture.

BOOK A WORKING SESSION