Business process automation: pipelines, integrations, data plumbing and agents inside the workflow
We automate the work that currently moves between systems by hand. A single workflow runs $5,000 to $15,000, a connected set of workflows $15,000 to $50,000, and a whole business function $50,000 and up. Ongoing operation runs $3,000 to $20,000 per month depending on volume and criticality.
What you get
A written map of the process before anything is built — Every step, every system, every human decision, every exception, and an honest count of how often the exception actually happens. Most of the value of an automation engagement is created in this document.
Workflows built on n8n, self-hosted — Self-hosting keeps your margin intact as complexity grows. Per-task pricing on hosted platforms is fine at ten runs a day and punitive at ten thousand — and complexity, not volume, is what usually crosses that line.
Scheduled jobs that run without supervision — Cron-driven work at the database level where it belongs. One client operation we run carries 40 scheduled jobs, each with a defined failure behaviour rather than a hope.
Serverless functions for the real logic — When a step needs actual code rather than a node in a canvas, it gets a function with tests. That same client operation runs 54 edge functions alongside its scheduled jobs.
Idempotency and retry behaviour — Every step is designed to be safe to run twice, because it will be. Retries with backoff, dead-letter handling for the ones that keep failing, and no silent double-charging or duplicate records.
Monitoring and alerting on the pipeline itself — Alerts when a run fails, when a run stops happening, and when a run succeeds but produces an implausible result. The third one is the failure mode nobody instruments and everybody eventually meets.
A human escape hatch on every automated path — A queue for the exceptions, with a person on it. Automation that has no exception path either blocks the whole business or quietly discards the awkward cases.
AI steps where judgement is genuinely required — Classification, extraction, summarisation and drafting inside the workflow — with a confidence threshold that routes uncertain items to a human rather than guessing on your behalf.
Documentation and a runbook — What each workflow does, what it touches, how to pause it, how to replay a failed run, and who to call. Written for whoever is on duty at 6am, not for us.
Published pricing
Single workflow — $5,000–$15,000 then Optional retainer. One process, two to four systems, monitoring, exception queue and runbook. Timeline: 2–4 weeks.
Multi-workflow programme — $15,000–$50,000 then $3,000–$8,000 / month. A connected set of processes sharing data and identifiers, with a shared exception queue and unified reporting. Timeline: 6–12 weeks.
Whole function automated — $50,000+ then $8,000–$20,000 / month. An entire operational function — intake, dispatch, billing, reporting — rebuilt as instrumented pipelines with agents in the loop. Timeline: 3–6 months.
Operate and improve retainer — — then $3,000–$20,000 / month. Monitoring, incident response, adapting to third-party API changes, and a monthly allocation for new and changed workflows. Timeline: Month to month.
How the work actually runs
01. Map the process as it really is (Week 1) — We sit with the people doing the work and record what actually happens, including the workarounds. Then we count the exceptions honestly, because a process that is 80% clean and 20% exception is a very different build to one that is 99% clean. Output: A process map with volumes, systems, decision points, exception rates and an estimate of hours currently spent.
02. Pick the cut line (Week 1–2) — Decide exactly which part gets automated and which part stays human. The goal is not full automation; it is removing the mechanical majority and routing the rest to a person with better context than they have today. Output: A written scope with the automated path, the exception path, and the success measure stated in hours or errors.
03. Build and instrument (Weeks 2–6) — Workflows built on self-hosted n8n, real logic in tested functions, scheduled jobs where work must happen on a clock. Idempotency and retries designed in from the start. Monitoring and the exception queue built at the same time as the happy path, not afterwards. Output: Working pipelines in a staging environment, with a failure you can trigger on purpose and watch alert correctly.
04. Run in parallel (1–3 weeks) — The automation runs alongside the manual process and the two outputs are compared. Discrepancies are investigated before anything is switched off. This stage is where you find the undocumented rule that lives in one person's head. Output: A reconciliation report and a signed-off decision to switch over.
05. Operate (Ongoing, monthly) — We watch the pipelines, respond to failures, and absorb the third-party API changes that would otherwise break things quietly. Each month a portion of the retainer goes to changing workflows as the business changes. Output: A monthly report of runs, failures, exceptions handled and hours saved, plus a changelog.
Who this is for
Businesses where someone spends hours a week moving data between two systems that will never integrate natively.
Operations with a process that is correct when people follow it and expensive when they do not.
Teams already paying for six SaaS tools that each hold a third of the truth.
Who this is not for
Processes nobody has written down yet. Automating an undefined process encodes the confusion and makes it faster.
Workflows that change fundamentally every month. Those need a decision, not a pipeline.
One-off data migrations. That is a script, we will tell you it is a script, and it should not cost $10,000.
The receipts behind those claims
35 scheduled jobs in one client operation — 40 pg_cron jobs running on a schedule inside a single client engine we build and operate, each with defined failure behaviour rather than a hope that someone notices.
56 edge functions in the same operation — 54 Supabase edge functions carrying the logic the workflow canvas should not. Automation at this scale is code plus orchestration, not a canvas alone.
25 CI workflows guarding changes — 25 GitHub Actions workflows with 157 test guards, so a change to shared automation code is caught before it reaches a production pipeline.
Self-hosted n8n, so margin survives complexity — Per-execution pricing on hosted automation platforms turns a successful automation into a growing bill. Self-hosting keeps the cost curve flat as the workflow count rises.
Which process should you automate first?
The one that is high volume, low judgement, well understood and currently done by a person who can describe it in one sitting. That combination is rarer than it sounds, and it is worth far more than the process everyone complains about most, which is usually high judgement and badly defined. A useful test: ask the person doing the work to explain the rule they follow. If the explanation contains more than three 'unless' clauses, the process is not ready to automate. It is ready to be decided about — which is a management conversation, not an engineering one, and we will say so rather than take the budget.
Why build on self-hosted n8n rather than a hosted automation platform?
Because of what happens at complexity. Hosted platforms price per task or per execution, which is comfortable when a workflow runs a hundred times a month and painful when a connected set of workflows runs tens of thousands of steps. The bill grows with your success, and it grows fastest exactly when the automation is working. Self-hosting moves the cost to infrastructure, which is close to flat, and removes the ceiling on what a workflow is allowed to do. It also means your credentials and your customer data stay in your own environment. The trade-off is that somebody has to operate it — which is what the retainer is for, and why we say so plainly rather than pretending self-hosting is free.
Where do AI agents belong inside a workflow?
In the steps that require judgement over unstructured input: classifying an inbound message, extracting fields from a document that arrives in nine different layouts, summarising a thread for a human who has to act on it, or drafting a reply that a person approves. These are steps where deterministic rules have historically failed and people have had to do the work. They do not belong in steps that are already deterministic. Using a language model to move a value from one field to another is slower, more expensive and less reliable than the integration that already does it. The discipline is to keep the model at the edges where ambiguity lives, and to attach a confidence threshold so that uncertain items go to a person rather than through the pipeline with a guess attached.
We already have Zapier and it mostly works.
Then keep it. Zapier is genuinely good for simple two-system workflows with low volume and no complex branching. The point where it stops being the right tool is recognisable: the per-task bill has become a line item people mention, workflows have grown branches nobody can follow, and there is no real monitoring when something fails silently. If none of those are true, you do not need us yet.
$3,000 a month to run some workflows seems steep.
It buys monitoring, incident response and absorbing third-party API changes before they break your operation. The alternative is not zero — it is a pipeline that fails quietly in month seven and is discovered by a customer. If your workflows are simple and low-stakes, take the build without the retainer and we will hand you the runbook. We would rather scope the retainer honestly than sell you one you cannot justify.
How much does business process automation cost?
A single workflow runs $5,000 to $15,000 to build, a connected multi-workflow programme $15,000 to $50,000, and a full operational function $50,000 and up. Ongoing operation — monitoring, incident response and continued development — runs $3,000 to $20,000 per month depending on volume and how critical the pipelines are.
How long does it take to automate a workflow?
Two to four weeks for a single workflow, six to twelve weeks for a connected set, and three to six months to rebuild a whole operational function. That includes a parallel-running period where the automation and the manual process are compared before anything is switched off.
What is the difference between automation and integration?
Integration connects two systems so data moves between them. Automation adds the decisions, the schedule, the exception handling and the monitoring around that movement — what happens when a record is malformed, when an API is down, when a human needs to approve something, and who finds out when a run fails.
Do you use n8n, Zapier or Make?
We build on self-hosted n8n, because per-execution pricing on hosted platforms erodes the value of an automation exactly as it grows and self-hosting keeps credentials and customer data in your own environment. Zapier and Make are reasonable for simple, low-volume workflows and we will say so if that is what you have.