Every AI run, with a receipt.
We build and operate AI workflows where every run has a spending ceiling set in code, a cost logged to four decimal places, a named person approving anything that leaves the system, and a one-click way back. We built it to run our own marketing agency. It now runs freight operations and AP/AR.
- Spend ceilings in code
- Human-approved
- One-click rollback
- A national freight brokerage
- Tradeloop
- Zehntek
- Pantek
- CampusWell
“One client has brought Bobos with him across two companies. We think that’s the only retention metric that matters.”
Watch what it costs, while it runs.
Start the conversation
Try: “https://example.com — we sell B2B SaaS to mid-market HR teams”
Type into the box above and a real agent answers you. Underneath its reply is what that turn cost us — which agent, which model, tokens in and out, dollars to four decimal places.
We’re not estimating that. Every agent call on this platform writes a row before anyone asks it to, and this is the same ledger our own delivery runs on. When your strategy finishes, you get the full receipt: every agent that touched it, what each one cost, and how close each came to its spending ceiling.
Every agent carries a hard dollar ceiling and a timeout, set in code. The cost is checked the moment the call returns, and an overage hard-stops everything downstream — it does not quietly fall back to a cheaper model and keep going.
The four questions we get asked before anyone signs anything.
“What will my CFO ask?”
What does it cost to run, per month, per workflow. We answer that from the ledger, not from an estimate — every invocation is attributed to a workflow and a client, so the monthly number is a query, not a reconstruction.
The cost ledger →“What will security ask?”
Whether one customer’s data can reach another’s. Our answer is enforced at build time: custom lint rules fail CI on any database query against a tenant-scoped table that isn’t correctly filtered, and a cross-tenant test suite runs 20 probes that must all come back empty before anything ships.
Tenant isolation →“What happens when it’s wrong?”
Every change to a live system is previewed, confirmed by a person, and reversible to the prior version in one click. We do not delete. 7 providers are wired this way, including GTM, WordPress and GA4 Admin.
Rollback →“Who owns it after you leave?”
You do. The runbook, the ledger and the workflow definition are deliverables, not retention hostages. We’d rather you stay because the next workflow is worth building.
Who’s on the other end →Four ways in.
Find out what your automations actually cost
You have Zapier flows, Make scenarios, n8n canvases and scripts nobody owns. We inventory them, price them, and tell you which are worth keeping.
$7,500 · 2 weeks · fixed
Put one workflow into production, properly
One business-critical workflow, live, with the cost ledger, the approval queue, the audit trail, the rollback and the runbook.
$65,000 · 8 weeks · fixed
Let us run them
We own the evals, the drift checks, the cost review and the incident response. You get a monthly Run Ledger.
From $12,500/month · month-to-month
Marketing, run this way
The function we proved it on. Strategy free, first campaign from $750, full team $4,500/month.
The agency we run on the platform
We didn’t build this to sell it.
We run a marketing agency on this platform. 32 recurring workflows, 12 production agents, every run cost-capped in code.
The caps exist because of arithmetic, not principle. When you publish a fixed price for a deliverable, an agent that quietly spends $40 producing a $12 asset doesn’t dent the margin — it ends the business. So we built the ledger, the ceilings, the approval queue and the rollback before a single client asked us for any of them.
Then a freight brokerage needed something else entirely: reconciling RFP submissions against carrier data, scoring which shippers were losing carriers, tracking trailer inventory across yards. Same platform. No changes to the foundations — four new agents on the substrate that was already there.
Then accounts payable and receivable, through QuickBooks. The agents draft the invoice or the bill. A person posts it. The tools are built so that they cannot post it themselves.
Every agent and every specialist fleet is off until its identifier is provisioned, and every scheduled workflow that can reach a customer is off until someone explicitly enables it — enforced in one place, with a visible record whenever a disabled workflow skips a tick. Exactly one workflow can’t be switched off by configuration: the job that redacts personal data after 30 days.
It is one platform, multi-tenant, not a bespoke build per client. Which means every hardening one client pays for, the next one inherits. Your fifth workflow is cheaper than your first because 32 governed workflows already found the edges.
- tenant-scoped tables
- 113
- governed recurring workflows
- 32
- production agents
- 12
- operational runbooks
- 68
“So why hasn’t your rate gone down?”
Fair. It has, per unit of work — which is why our prices are fixed and published rather than hourly. We don’t bill you for the time we saved ourselves. We bill for the workflow, and it costs the same whether it takes us three weeks or eight.
Proof.
A national freight brokerage — freight operations
RFP reconciliation against carrier data, carrier-churn scoring, trailer tracking across yards. Nobody timed the manual version before us, so we won’t invent a before/after — the workload and what it catches are documented in the case study.
Read the freight chapter →Bobos — the agency we run on it
33 completed campaign reviews · $0.0664 average cost per review, from the ledger
The weekly campaign review used to be analyst work; it now runs as a governed fleet with a cost per completed review read straight from the ledger. The analyst-hours baseline was never instrumented — we say that rather than reconstruct it.
Read the origin chapter →Finance — AP/AR, propose-only
Invoices and bills drafted by agents through QuickBooks; a person posts every one. The tools cannot post — propose-only is the design, not a policy promise.
How propose-only works →Including the runs we rolled back. 2 client-property writes applied to date, 0 rolled back.
$7,500 / 2 weeks
Agent Debt Audit.
Credited in full against a build within 60 days.
$65,000 / 8 weeks
First Governed Workflow.
Fixed price, fixed scope, runbook yours.
From $12,500 / month
Governed Operations.
Month-to-month, as everything here has always been.
Two builds run at a time. That’s not scarcity marketing, it’s the actual constraint — the people you meet are the people who ship.
Full prices and timelines →The failure modes worth knowing about.
Workflows with no clean before-state.
If nobody timed the manual version, we can’t prove we improved it. We’ll tell you that at the audit rather than reconstructing a baseline you’d be right not to believe.
Judgment-dense work.
Where the value is in one hard call rather than a hundred small ones, an agent adds cost and a review step and not much else. We’ve declined this work and will again.
Data that isn’t ready.
This is the most common reason a build slips. If the source systems disagree with each other, the first weeks go into reconciliation, not agents — and we’d rather scope that honestly than discover it in week five.
The second year.
Automations rot. Models change, APIs move, the business changes shape underneath. A workflow nobody owns in month fourteen is worth less than no workflow at all. This is why we sell an operations retainer and why we hand over runbooks either way.
Who this is for.
A fit if:
You’re 50–1,000 people. Someone owns the P&L on the workflow in question and can approve a build on it. You have real volume in a process that’s currently done by hand or held together by automations nobody maintains. You’d rather see one workflow work than receive a roadmap.
Not a fit if:
You want an AI strategy document. You want a headcount-reduction business case — we won’t write one, the evidence on those is bad and the rehiring is expensive. You need this in three weeks. Or you’re looking for a marketing agency, in which case: → /marketing, where we’re very good.
Questions buyers actually ask
What does it actually cost to run an AI workflow per month?
Every invocation is attributed to a workflow and a client in the cost ledger, so the monthly number is a query, not an estimate. The figure depends entirely on volume, which is why the audit prices the workflows you actually run before anything gets built.
How do you stop an agent from running up a bill?
Every agent carries a hard per-invocation dollar ceiling and a timeout, set in code. The cost is checked the moment the call returns, and an overage hard-stops everything downstream — it does not quietly fall back to a cheaper model and keep going.
Can an AI agent change something in our systems without a person approving it?
No. Every tool that reaches outside our platform proposes the change and returns a confirm link for a person to act on, and 7 providers are wired behind preview, confirm, and one-click rollback.
How do you keep one client’s data away from another’s?
It’s enforced at build time: custom lint rules fail CI on any database query against a tenant-scoped table that isn’t correctly filtered, and a cross-tenant suite runs 20 probes that must all come back empty before anything ships.
You’re a marketing agency. Why would we hire you for operations?
Marketing has the highest event volume and the most fragmented tool surface of any business function, so it forced the governance early. The ledger, the ceilings and the rollback were built to protect our own fixed-price margins. The same substrate now runs freight operations and AP/AR.
What happens to our data?
5 named subprocessors, each with a stated purpose and region. Your data is never used for model training, and deletion propagates within 30 days.
The full record is on our trust page →What do we own at the end?
The workflow, the runbook, and the ledger. They’re deliverables, not retention hostages — we’d rather you stay because the next workflow is worth building.
Start with what you already have running.
Most companies don’t need a new AI project. They need to know what the automations they already own are costing them, which ones are broken, and which one is worth doing properly. That’s a two-week, $7,500 answer.
Book the auditor write to notifications@bobos.ai