How we work
The right amount of AI. Not the most.
We start with how work actually moves through your business. Then we decide what should stay human, what should be a plain automation, and where AI genuinely earns its place — and build the simplest thing that works, inside tools you own, with your approval on every action until it has earned more.
Does this sound familiar?
Six questions we ask every owner.
- 01Where does a lead, booking, or job stop moving — and who notices?
- 02What only gets done because one particular person remembers to do it?
- 03What breaks the week that person is out?
- 04Where does someone re-type the same information into a second system?
- 05Which software do you pay for every month that the team barely uses?
- 06What do you only find out about after a customer complains?
The engagement
Six phases, run one workflow at a time.
Each workflow moves through these on its own clock — one can be in pilot while the next is still in design. Small, low-risk workflows move through quickly. The pilot never gets skipped.
We map how the business actually operates — what decisions live in the owner's head, what falls through when the usual person is out, what the team does every day that nobody has ever written down. Then we score what we found by value, risk, and readiness.
You get: A ranked list of opportunities and a baseline for each: what it costs today in time, money, or errors. You pick the first one to three workflows.
For each workflow we decide what should stay human, what should be a plain automation, and where AI genuinely earns its place — then design the target process, what the system may and may not do, and every point where a person approves.
You get: A design you sign off, including which actions need your approval. Nothing gets built until you do.
We build the simplest pattern that meets the design, inside the tools your team already uses — Teams, Slack, Google Workspace, Microsoft 365, your CRM and scheduling system. It is tested against real examples from your business before it touches anything live.
You get: A working workflow that has passed tests on your own data, with security checks clean.
It runs alongside your team first, so we can compare its output to what people actually did. Then it goes live with your approval on every action that sends, books, charges, or changes something. We agree the pass criteria before the pilot starts, not after.
You get: A weekly pilot report: runs, errors, time saved, and cost.
You get the parts you own, and the training to run and change them without us. Handoff is not done until your owner can do that unaided.
You get: A runbook, owner training, an ROI report against the baseline, a data inventory, a credential list (names and owners, never values), alerting, and an offboarding plan.
We monitor, fix, and keep the stack current. When a category of action has been approved unchanged for long enough, we propose giving it more room — and you decide, in writing. Edits your team makes to drafts become improvements to the system.
You get: A recurring outcome report against the baseline, and a one-page proposal for any change that matters.
Right-sizing
Most of it shouldn't be an agent.
Agents are the expensive, hard-to-maintain end of the range. For every workflow we pick the lowest level that does the job, and only move up when there's a concrete reason — never for “future flexibility.”
Stays human
Judgment, relationships, or decisions where a mistake is expensive and hard to undo.
e.g. Talking a long-time client through a complaint.
Plain automation
Every step is known in advance. No judgment between steps.
e.g. A review-request text two hours after a completed appointment.
Automation + one AI step
A fixed process where one step needs reading, writing, or scoring.
e.g. Drafting a reply to a new Google review in your voice, for you to approve.
AI agent
The path isn't known in advance — what it finds decides what it does next.
e.g. Finding out why yesterday's revenue didn't show up in the report.
Multi-step workflow with human pauses
It branches, waits days for a person, and picks up where it left off.
e.g. A coaching loop: score, manager review, feedback, re-score two weeks later.
Autonomy
Autonomy is earned on evidence, not granted on day one.
How capable the AI is has no bearing on how much it's allowed to do. That's set by how risky the action is and by its track record in your business.
It starts by proposing
A new workflow that sends, books, charges, or changes anything drafts the action and waits. A person approves before it happens.
Approval happens where you already work
One tap in Teams, Slack, text, or email, showing the exact message, recipient, or amount — with approve, edit, or reject. High-volume work comes as one daily list, not thirty pings.
Silence never means yes
An unanswered approval holds, reminds, then escalates to a named backup. It never goes ahead on its own.
More room is earned, and can be taken back
When an action type is approved unchanged at a high rate over a sustained period, we propose letting it run without per-item approval. You confirm in writing, and anything it does still lands in a daily digest. If errors rise, it steps back down.
What we watch for
The first things that break in production aren't the model.
After running agents 24/7 against real operations, the failures are predictable. The reasoning is rarely the issue. These five patterns are what we design around from day one — and what the audit looks for before any code is written.
The handoffs break before the model does
An agent qualifies the lead, but the CRM update silently fails. A voice assistant books the appointment, but the calendar field is the wrong format. The agent looks like it worked. The workflow didn't finish. Reasoning is rarely the failure point — transitions between systems are.
Source data gets messy fast
Agents are only as reliable as the operating environment around them. Old SOPs, duplicate customer records, half-updated docs, conflicting notes. None of that is the agent's fault — and all of it makes the agent look unreliable. In multi-agent systems, small data errors compound across steps.
Exception handling is the whole game
The demo path always works. Production is all edge cases. People reply out of order. A customer asks two things at once. An API times out. A staff member changes a record mid-flow. Without explicit rules for retries, fallbacks, and human review, trust leaks fast.
Ownership goes fuzzy at 2 a.m.
When a 24/7 workflow fails, whose job is it to notice? Ops? Sales? The founder? The most underrated source of production failures isn't a bug — it's nobody owning the outcome end to end. Agents need a named human who watches them.
Too much autonomy, too early
Owners want full automation on day one. Almost every workflow needs a staged rollout: first assistive, then partially automated, then higher autonomy once the error patterns are understood. Skip the staging and you get cleanup work, not leverage.
Who owns what
Built to be yours by default.
What your team runs day to day lives in tools you can see and change. We only host something ourselves when there's a stated reason — usually that it's too complex to maintain in-house, or a security control only holds if we run it.
You own and run
- →Your data, in your own accounts — CRM, Microsoft or Google tenant, storage
- →Standard automations, built in n8n you can open and edit
- →Your Claude plugins and skills: SOPs, brand voice, connectors
- →Connections to your systems, hosted on your side where practical
We run for you
- →Agent and multi-step services too complex for in-house upkeep — unless your IT team wants them
- →Monitoring, cost tracking, and testing, with a dashboard you can see
- →Failure alerting, with the bot handed over or rotated at handoff
- →Security patches and model and tool reviews, under the monthly plan
Security baseline
Built for production. Not for demos.
Every engagement ships with the same baseline. It's part of delivery, not an add-on.
Credentials
- →Every secret lives in a secrets manager — never in code, never emailed
- →Separate credentials, service accounts, and billing for every client
- →The credential list you get at handoff names each one and its owner, never its value
Access
- →Connections are read-only unless writing is part of the design you approved
- →Your approval on every write until you sign off on more autonomy
- →An audit log and a kill switch on every workflow that writes or sends
Data
- →A one-page inventory: what each workflow touches, where it's stored, and which model sees it
- →Emails, documents, web pages, and customer messages are treated as data, never as instructions
- →Sensitive fields are redacted before anything reaches monitoring
The build itself
- →Pinned dependency versions and automated security scanning on every repo
- →Every change version-controlled, with a rollback path
- →An offboarding plan from day one: revoke every access, hand over or shut down every component, return or delete your data
After launch
Proven with numbers. Changed on evidence.
Every workflow is measured against its baseline from its first live run — time, cost per run, errors, and the business outcome where it can be measured. Renewals and expansion get argued from those numbers. We don't switch tools or models because something newer came out.
Maintenance
Security patches, bug fixes, dependency updates, and model upgrades that pass the same tests. Covered by the monthly plan and reported to you. Security fixes don't wait for sign-off.
Modernization
Replacing a tool or redesigning a workflow. Only when there's a reason — a security issue, a vendor change, reliability below target, or a saving that pays back within a year. You get a one-page proposal and approve it in writing.
Expansion
A new workflow or capability. It goes through the same design and approval path as the first one, scoped separately.
Tech stack
Open standards. Portable across models.
Workflows
- n8n — client-owned, editable
- Claude plugins and skills
- Durable execution only where a workflow truly needs it
AI
- Hosted frontier models by default (Claude)
- Local models when your data can't leave your control
- Models re-tested quarterly; switched only when they pass at the same or lower cost
Integrations
- MCP — an open standard, not a proprietary connector
- Microsoft 365 / Teams, Google Workspace
- Most CRM and scheduling systems via API
Monitoring
- Per-run traces, cost, and error rate
- Time saved against your baseline
- Self-hosted when data sensitivity requires it
Common questions
Things owners ask before they sign.
If yours isn't covered here, reply to hello@operably.ai and we'll get you a real answer the same day.
Is everything going to be an AI agent?+
No — most of it won’t be. Most of the work we find is best served by a plain automation, or an automation with one AI step. Some of it should stay with a person. We only use an agent where the path genuinely can’t be known in advance, because simpler is cheaper to run and easier for you to own.
How much of my team’s time will this take?+
Discovery starts with one 60–90 minute session plus follow-up questions. During build, about 30 minutes a week for review. During the pilot, someone approves actions — usually a few taps a day. After that it depends on how much autonomy you choose to grant.
What if we don’t have clean data or documented processes?+
That’s normal — most owner-operated businesses we work with don’t. Discovery maps your operations as they actually run today, including what’s in someone’s head and what’s undocumented. The system we build accommodates the messy reality first, then we clean up where it makes sense.
Will it ever act without asking?+
Only if you decide it should, and only for actions that have earned it. Anything that sends, books, charges, or changes something starts with your approval on every action. When an action type has been approved unchanged at a high rate for long enough, we propose giving it more room — you confirm in writing, you still see everything it did in a daily digest, and it steps back down if errors rise.
What if it gets something wrong?+
Every workflow that writes or sends has an audit log and a kill switch, and the pilot runs with your approval on every action — so mistakes get caught as drafts, not discovered after the fact. Your first workflow is live in 4 weeks, doing the work with your approval on every action; if it isn’t working as defined, you don’t pay setup until it is.
Do we need an IT person or tech team?+
No. The parts you own are built in tools a non-developer can open and change, and handoff includes training until your owner can do that unaided. Anything too complex to maintain in-house, we run for you.
What happens if we stop working with Operably?+
Your data already lives in your own accounts, and the automations, plugins, and skills you own stay with you. Every engagement has an offboarding plan from the start: every access we hold is revoked, the services we run are handed over or shut down, and any of your data we hold is returned or deleted.
Stop executing. Start governing.
The worst case: you do the mapping session and leave with a clearer picture of what's costing you — before spending anything on a build.
Start with an operations audit →