We build AI agent systems that take real work off your team’s plate and run it across the tools you already use, with human oversight where it matters.
Your process isn't broken. The execution of it is.
The work gets done. Just slowly, inconsistently, and with too many dropped balls in between. What should be a system is still dependent on busy humans pushing tasks from one tool to the next.
Manual, today
manual_lead_followup
Inbound leads sit in an inbox waiting for someone to get to them. The ones that get contacted are copy-pasted between CRM, email, and Slack. The rest go cold.
14:02lead #8842 submitted form
14:47lead #8843 submitted form
T+48h11 of 14 still uncontacted
T+5hlost to competitor
5 hr avg response time3 / 14 leads contacted
With an Agent System
lead_followup · bishop-v1
A form submission triggers a workflow: score, enrich, qualify against ICP, draft a personalized first-touch, route to the right owner, log to CRM, notify Slack. You approve the drafts you want to approve.
T+0sform submission → parsed, source tagged
T+12senrichment complete · 92% ICP fit
T+1mfirst-touch drafted, owner assigned
T+4mCRM synced · Slack notified
4 min avg response time14 / 14 leads contacted
/ The unlock you're actually getting
You can stop doing work. Start orchestrating it.
This system combines agents, automation logic, integrations, and human review into one working system. Each one is built around a specific operational outcome like lead follow-up, onboarding, reporting, or internal handoffs.
a chatbot
a Zapier flow
a prompt library
an AI demo
a workflow-level system
IN
parse_intake_form[Auto]
Extract fields, tag source
02
score_against_icp[Auto]
Apply fit rules, assign priority
!
review_flagged_cases[Human]
Ambiguous leads go to owner
03
route_to_crm · notify[Auto]
Create record, alert owner
!
approve_outreach_draft[Human]
Human signs off before send
OK
send · log · handoff[Auto]
Close the loop, record everything
/ HUMANS WERE NOT MEANT FOR COPY-PASTE
MANAGE_CONNECTIONSidle
S
Slack
OAuth 2.0
Connected
A
Apollo
API Key
Connected
H
HubSpot
OAuth 2.0
Connected
G
Gmail
OAuth 2.0
Connected
ENRICH_CONTACTSidle
source--
matched--
emails found--
titles verified--
status--
Agentworkflow:prospecting
agent agent-v1
model claude-sonnet-4-6
HUBSPOT_SYNCidle
action--
contacts--
list--
duplicates--
sync--
DRAFT_OUTBOUNDidle
template--
personalized--
tone check--
queued--
review--
SANDBOXidle
ENRICH_BATCHwaiting
for prospect in targets:
apollo.enrich(
email=prospect.email,
linkedin=prospect.li
)
filtered = [p for p in prospects
if p.title in ICP_TITLES]
CRM_PUSHwaiting
for contact in enriched:
hubspot.create_contact(
email=contact.email,
company=contact.company,
source='agent-workflow'
)
slack.notify('#pipeline')
Agent · agent-v1workflow: prospecting
> apollo.search querying ICP targets...
> filter: SaaS, 10-50 headcount, US
> found 142 accounts
> apollo.enrich resolving contacts...
> matching titles against ICP rules
> 87 contacts enriched(61% match)
> suppressing 12 duplicates from CRM
> hubspot.sync pushing 75 new contacts...
> list: Q2-SaaS-Outbound
> 75 contacts synced
> draft.generate building sequences...
> personalizing from account briefs
> tone: direct, no jargon, founder-voice
> 75 drafts queued for human review
> slack.notify #pipeline "75 prospects ready"
> awaiting approval before send
> workflow complete(4m 12s)
connectionsdone
enrichmentdone
crm syncdone
outbounddone
workflowprospecting
agentbishop-v1
modelclaude-sonnet-4-6
runtime4m 12s
replaces~4hr / week manual
/ WHERE AGENTS ACTUALLY HELP
Where agent systems usually pay off first
Most businesses do not need more technology. They need one important workflow taken off human hands. And they need it yesterday. These are the three places we usually find it.
Automations fire notifications. Agent systems run your workflow.
Task-level automation saves a step. AI Agent Systems move the job forward. They handle the sequence, handoffs, decisions, and exceptions that usually grind momentum to a halt.
Task-Level Automation
One rule. One action. One beep.
Send a notification. Update a field. Move a row. Useful, but the workflow around it is still stuck with a human.
triggerstatus = "Closed"
actionslack.send_message
channel#ops
message{{deal_name}} marked Closed
AI Agent System
End-to-end workflow. Judgment built in.
New lead submitted via form. Qualify, route, onboard. Decisions, handoffs, approvals, exceptions, all in one run.
01parse intakedone
02score against ICP92% fit
03route to ownerKurt F.
04create CRM contactsynced
05notify #inboundsent
06review flagged caseawaiting human
07schedule follow-upqueued
/ THE OPERATING RULES
Four rules. They keep AI builds useful, scoped, and under control.
AI builds fail when the scope is vague, the workflow is unclear, and nobody owns the result. We avoid that by tying every system to one operating workflow, one outcome, and a defined boundary.
01 / PRIORITIZATIONworkflow_audit.trace
The workflow always comes first. Tools come second.
The starting point isn't the tool. It's the workflow: where work enters, where it stalls, what decisions get made, and what the next step should be. Tool selection happens after mapping. Never before.
Map entry point: where does the work enter the system?
Locate the stall: where does it wait on a human?
Identify the decisions: what judgment is actually required?
Define the next step: what should happen after each decision?
02 / TECH STACKstack_ownership.config
Build on top of your stack. Don't replace it.
Core systems (CRM, inboxes, calendars, sending) stay client-owned. We build workflow logic and orchestration in a separate, portable layer on top. No brittle rebuilds where mature software already exists.
Workflow LayerOrchestration · routing · agentsowned by Growth Marshal
Integration LayerAPIs · data mapping · syncowned by Growth Marshal
InfrastructureCRM · inbox · calendar · sendingowned by Client
03 / CONTROLapproval_gates.policy
Judgment stays with the human. Repetition moves to the system.
Most workflows still need approvals, judgment, and exception handling. Approval gates are built into the workflow, not bolted on. Volume handled by the system. Decisions handled by the owner.
parse_intake_formauto
score_against_icpauto
review_flagged_caseshuman
route_and_notifyauto
approve_outreach_drafthuman
04 / SCOPEscope_control.lock
One workflow first. Additional systems earn their way in.
We don't begin with "let's automate the business." We begin with one important workflow, scoped before kickoff, with one owner and one outcome. Expansion follows delivery, not the other way around.
Phase 1 · ship this
inbound lead routing
CRM record creation
owner notification
Later · earn its way in
outbound prospecting
reporting dashboards
/ Fit Check
A strong fit or not quite there yet? Here’s the honest cutoff.
Agent Systems work when there's a clear workflow, real operational pain, and an internal owner. If that’s not you, we’ll tell you before you spend a dollar.
Dimension
/ Strong fit
/ Not a fit
DimensionWorkflow clarity
Strong fitRecurring workflow with identifiable friction points
Not a fitVague AI strategy with no specific workflow target
DimensionOperational pain
Strong fitLosing time to manual coordination, follow-up, or admin
Not a fitExpecting full transformation without process clarity
DimensionGrowth model
Strong fitMore output without immediately adding headcount
Not a fitLooking for a generic chatbot or AI demo project
DimensionTool readiness
Strong fitBusiness tools already in use; they just don't talk to each other
Not a fitAvoiding all software investment even when the workflow needs it
DimensionAI expectations
Strong fitPractical automation tied to measurable outcomes
Not a fitWants to automate everything before proving one workflow
DimensionOwnership
Strong fitClear internal owner willing to run the workflow long-term
Not a fitExpects Growth Marshal to substitute for internal ownership
/ What Success Looks Like
Business gets better when work moves to an agent system
Common outcomes include faster lead response, more consistent prospecting, fewer dropped handoffs, cleaner records, and less founder dependency. On the call, we’ll give you a realistic range for the workflow you’re considering.
5h4m
lead response time, inbound
B2B SaaS · 10 to 50 headcount
87%
reduction in manual time per workflow
median across shipped systems
0
dropped handoffs across operational runs
logged + auditable end-to-end
1
workflow to start, then expansion
scope controlled before kickoff
/ FAQ
You probably have questions. We have some answers.
If you don't see yours here, just ask!
What does an Agent System actually cost?
In general, a deployed agent will land in the $200 to $2,000 per month range, depending on a blend of factors, including workflow complexity, run volume, AI models in use, the number of tool integrations, and the status of the underlying data and tech stack at kickoff. We will quote you a fixed estimate after scoping your requirements.
How long does it take to ship the first system?
Four to fourteen days for most workflows, from kickoff to a system running in production. The variation is mostly about your data and integrations, not our build speed. A clean stack with a well-defined workflow can ship in under five days. More complex integrations and approval chains will take longer. We tell you which end of the range you're on before we build.
What happens to my data?
Your data stays where it lives. Your CRM, inbox, calendar, and project tools never move. Agents reach into those systems through credentials you control and can revoke at any time. Agents take the specific action the workflow calls for, with every step logged.
The workflow logic, the agent runtime, and the orchestration layer all run inside a secure, hosted environment with encryption in transit and at rest, scoped least-privilege access, and full audit logging on every run. Your data is not used to train AI models, ours or anyone else's. You retain ownership of everything that runs through the system.
What happens when the system breaks or makes a wrong call?
Every system ships with logging, alerting, and a kill switch. If an agent step fails or returns low-confidence output, the workflow either escalates to a human owner or pauses and waits for review. Most "wrong calls" never reach production because approval gates sit in front of any irreversible action. When something breaks in the wild, we see it the same day, not the same quarter.
What do I have to do during the build?
Less than you think. We need a kickoff session, access to the relevant tools, an internal owner who can answer workflow questions, and your approval at the design and launch gates. The actual building is on us. Most clients spend ~15 minutes on a call with us once every 3-4 days during the build, mostly on review and approval, not on production work.
Will this disrupt my team's existing work?
No. The system runs alongside your current process until your team is ready to let it carry the load. We deploy in a non-disruptive mode first, with humans in the loop on every step, and only widen the autonomy as the workflow proves itself. Your team learns what to hand off to the system, not to wrangle it.
What if my workflow changes after launch?
Workflows always change. Each system is built in modular pieces so individual steps can be swapped, expanded, or retired without rebuilding the whole thing. The post-launch optimization window extends throughout the life of the engagement, and changes are scoped as small follow-on work and rolled into the managed service arrangement.
How is this different from Zapier, Make, or a low-code automation tool?
Those tools are good at moving data from point A to point B. They are not built to reason, draft, qualify, summarize, or handle exceptions. An Agent System uses Zapier-class automation where it fits, and uses AI agents where judgment is required. The result is one system that handles the entire workflow end to end, not a wall of disconnected zaps that still need a human to babysit them.
Why not just hire a freelancer to build this?
A freelance build typically optimizes for a specific tool, ships a working demo, and moves on. What's left behind is brittle, undocumented, and tied to whoever built it. Marshal builds production systems with logging, audit trails, approval gates, and a portable architecture. You also never have to think about staff turnover or tool changes, because we manage the service end to end.
How do I know the system is working?
There is a technical answer and a common-sense answer.
First, the technical. Every system ships with built-in observability: run logs, success and failure counts, time-to-completion, and the specific outcomes the workflow was scoped to improve. We agree on the success metrics at kickoff, so "working" is not a matter of opinion at the end.
Now, common sense. Did the work get done? Was it done correctly? Was it done within your policy and procedure guidelines? If those answers aren't obvious, then we built something that isn't actually solving a problem.
Is this only for tech companies or B2B SaaS?
No. Agent Systems work in any business with a recurring workflow, a real operational bottleneck, and an internal owner. We've shipped systems for professional services, agencies, e-commerce, and founder-led B2B companies. The pattern that matters is the workflow shape, not the industry label.
What if we don't have a workflow scoped yet?
That's the point of the consult. We map your operational surface, find the workflow with the highest leverage, and tell you whether it's a fit. If it isn't, we say so. If it is, you leave the call with a scoped problem and a real plan to solve it, not a vague promise.