# The Automation Department, in full

Every answer below is written to stand on its own. Prices are the published ones and are fixed before work starts.
Last updated: 2026-10-07
Source: https://theautomationdepartment.com/llms.txt

## How much does business process automation cost for a small business?
https://theautomationdepartment.com/answers/how-much-does-business-process-automation-cost/

Business process automation is priced in two parts: a fixed setup cost once, then maintenance, either pay as you go or a monthly plan. The cheapest way in is the starter, one trigger and one action in a tool you already use, at $499 one-off, live in ten business days and credited in full against a full system if you upgrade within ninety days. Full systems run $1,500 to $12,000 to set up and go live in ten business days or less: instant lead reply and missed-call text-back are $1,500, a booking engine or customer re-engagement $3,000, invoice and document processing $3,500, an AI voice receptionist $5,000, a connected back office $12,000. Three systems bundled by outcome come as a package at $6,000 setup, twenty per cent off buying them separately, with the Growth plan included. After launch: pay as you go at $175 a change with nothing monthly, Growth at $700 a month with six changes, or Scale at $1,800 with twenty. Platform subscriptions add roughly $30 to $150 a month, paid directly to the vendor. Scoping and the written proposal are free.

## What can a 20-person company actually automate?
https://theautomationdepartment.com/answers/what-can-a-20-person-company-automate/

At twenty people, the automations that reliably pay for themselves are lead intake and routing, quote and proposal generation, client or customer onboarding, invoice capture and matching, recurring reports, and inbound email triage. These share a common shape: they run often, they follow rules, and a person is currently moving data between systems that do not talk to each other. What does not pay for itself at this size is anything that runs a handful of times a month, anything still changing shape, and anything requiring genuine judgment. A useful test: if nobody could write down the rules, it is not ready to automate yet.

## How long does an automation project take?
https://theautomationdepartment.com/answers/how-long-does-an-automation-project-take/

A single automation from the price list takes 10 business days from go-ahead to running in production. That is the build itself. Around it sit a thirty-minute scoping call, a written proposal within two business days, one to two weeks of parallel running, and a couple of days for cutover and handover, so end to end a first automation is typically three to five weeks from first conversation to fully live. Several automations built together as a package, or a build with genuine two-way sync, run longer. The things that actually cause overruns are almost never the engineering: they are waiting for API credentials, discovering undocumented exceptions, and not having one person who can sign off.

## Do I need an AI agent or just a workflow?
https://theautomationdepartment.com/answers/ai-agent-or-workflow/

If the task follows rules, you need a workflow. It will be cheaper to build, cheaper to run, faster, and it will behave identically every time. You need an AI agent only when the task requires reading and interpreting language that arrives in an unpredictable shape, such as sorting messy inbound email, summarizing a long document, drafting a reply, or pulling structured facts out of free text. Most businesses need mostly workflows with a small amount of AI at the edges. The single most common and most expensive mistake in this market is putting a language model where an if statement belongs.

## What happens when an automation breaks?
https://theautomationdepartment.com/answers/what-happens-when-an-automation-breaks/

A well-built automation stops, alerts a named person, and leaves the underlying data untouched. A badly built one keeps running with wrong data and nobody notices until a customer complains. The difference is entirely down to whether error handling, retry logic and failure alerting were designed in from the start. Expect something in a busy automation stack to need attention every few months: APIs get versioned, OAuth credentials expire, vendors deprecate endpoints, and someone renames a column in a spreadsheet. The question is never whether it breaks, but whether you find out from an alert or from an angry email.

## Should I hire an ops person or automate?
https://theautomationdepartment.com/answers/hire-an-ops-person-or-automate/

Hire when the work needs judgment, relationships or physical presence, or when volume is unpredictable. Automate when the work is high-volume, rule-driven and stable. A US operations hire costs roughly $55,000 to $85,000 a year fully loaded; a build that removes 15 hours a month of the same work costs $499 to $4,500 to set up, then $175 a change or a plan. That comparison is only fair if the automation genuinely replaces the whole role, and it almost never does. For most businesses the realistic answer is both: automate the rule-driven core so the person you hire spends their time on the judgment work you actually needed them for.

## Is n8n or Make better for a small business?
https://theautomationdepartment.com/answers/is-n8n-or-make-better-for-a-small-business/

Make is the better default for most small businesses. There is no infrastructure to run, the connector library is larger, and the visual editor is followable by non-developers. It is priced per operation, so cost rises with volume. n8n wins in three specific situations: high execution volume where per-operation pricing becomes painful, a requirement to self-host for data residency or compliance, and logic complex enough that you want real code inside nodes. The crossover is driven almost entirely by how many operations your workflows consume per month, not by how many workflows you have. Under roughly 10,000 operations a month Make is almost always cheaper and easier; above roughly 100,000, self-hosted n8n is usually decisively cheaper.

## What is business process automation?
https://theautomationdepartment.com/answers/what-is-business-process-automation/

Business process automation is software doing the routine steps between your existing systems (the copying, checking, chasing and filing) the same way every time, without a person in the middle. It is not a new system you migrate into; it is connective tissue between the systems you already run. It replaces four kinds of work: re-keying the same information into a second system, chasing people for approvals and payments, checking one record against another, and filing documents and updating statuses. It is distinct from robotic process automation, which drives the user interface and breaks when the screen changes, and from AI agents, which handle judgment rather than rules. A process is a candidate if it happens often, follows rules someone can state, runs on systems with APIs, has reasonably clean data, and has one person who can make decisions about it. The starter is $499 one-off; full systems run $1,500 to $12,000 to set up, with nothing monthly unless you take a plan.

## How much does an AI agent cost for a business?
https://theautomationdepartment.com/answers/how-much-does-an-ai-agent-cost/

An AI chat agent costs $3,500 to set up and an AI voice agent $5,000, both fixed before work starts, with most live within 10 business days. Both include training on your own material, an evaluation set built from your real cases, confidence thresholds with a handoff path, structured output, logging and documentation. After launch there is nothing monthly unless you choose a plan: pay as you go bills each change at $175, Growth is $700 a month for six changes with monitoring and fixes, Scale is $1,800 for twenty. Model usage and telephony are paid to the providers rather than to us and we take no margin; for a chat agent handling a few hundred conversations a day, model usage is usually tens of dollars a month. Voice costs more than chat because latency is a product requirement, speech recognition adds a second failure surface, interruption handling is necessary, and telephony has to be managed.

## Is our data safe if we automate?
https://theautomationdepartment.com/answers/is-our-data-safe-with-automation/

Automations we build run inside your own accounts, on subscriptions you own, with credentials held in your vault rather than ours, so revoking our access stops us working on them without stopping them running. Your data does pass through the automation platform in transit and through a model provider wherever AI is involved, and the platform stores execution logs for a retention period you configure, which should be set deliberately rather than left at the default. Business API tiers at the major AI providers do not train on your inputs by default, unlike consumer chat products, and zero-retention arrangements are available from some providers. Self-hosting removes the platform vendor from the path entirely where data residency matters. Health data requires a business associate agreement from every component, which rules out several mainstream platforms at standard tiers, and card data should be kept out of automations entirely by working with payment processor tokens. In practice incidents come from over-broad API permissions, credentials tied to an individual who leaves, shared logins, and unconfigured log retention rather than from platforms being breached.

## What happens if we stop working with you?
https://theautomationdepartment.com/answers/what-happens-if-we-stop-working-with-you/

Everything keeps running. The automations live in your accounts, on your subscriptions, authenticating with your credentials, so they do not call anything of ours and removing our access changes nothing operationally. You keep the workflows themselves, the credentials which never left your vault, the plain-language documentation and recorded walkthrough handed over at go-live, the evaluation sets behind any AI component, any custom code, and your execution history. There is no proprietary runtime to be locked out of and any plan is month to month with no exit fee. Lock-in, where it exists elsewhere, takes four forms worth checking for: the workflows sitting on the firm's platform subscription rather than yours, a proprietary middleware layer the automation calls, credentials wired through the firm's identity so revoking access breaks production, and documentation offered as a paid extra. The test is simple: if you revoked their access this afternoon, what breaks? The answer should be nothing except having nobody to call.

## How do I choose an automation partner?
https://theautomationdepartment.com/answers/how-do-i-choose-an-automation-partner/

Judge an automation firm on how it estimates rather than on what it has built, because anyone can demonstrate a finished system. Four questions separate them: how did you arrive at this estimate and who did you watch, whose accounts will this run in, what happens when it breaks at 2am, and what would you tell us not to automate. An estimate produced without anyone shadowing the work has priced the process as a manager believes it works, which is reliably not how it works. Warning signs include quoting before observing, percentage-of-savings pricing, reselling the platform they recommend, promising 100% automation, no mention of a parallel run, and documentation offered as a line item. Good signs include asking what happens today when something goes wrong, discussing exceptions early and specifically, stating running costs as well as build costs, and willingness to lose the sale by recommending an off-the-shelf product. Compare quotes by normalizing six things first: exception handling, the review queue, testing against real data, parallel running, documentation and handover, and who owns the accounts.

## Can you automate an old system with no API?
https://theautomationdepartment.com/answers/can-you-automate-an-old-system-with-no-api/

Usually yes, but a system with no API is the one factor that genuinely changes both the price and the risk, so it belongs on the first call rather than being discovered mid-build. There are five routes in, and the rule is to take the highest one the system supports rather than the one that always works. First, a documented API, which more systems have than their users realize. Second, scheduled export and import, which is stable, cheap and survives vendor updates but inherits the export's cadence. Third, direct database access, where reading is safe and writing around the application bypasses its validation. Fourth, email or file drop as a trigger, combined with extraction. Fifth and last, driving the user interface, which works but breaks whenever a screen changes and carries a permanent maintenance obligation. A frequently better option is to automate everything around the legacy system and leave one short manual step inside it.

## Do we need to fix our processes before we automate?
https://theautomationdepartment.com/answers/do-we-need-to-fix-our-processes-first/

No. You need to be able to describe a process, which is not the same as having tidied it. A messy process, meaning one with many undocumented exceptions and workarounds, is fine to automate because the exceptions are the scope rather than a blocker. An unsettled process, meaning one whose shape has genuinely changed twice in the last quarter, is not, because you would be encoding something you are about to abandon. Fix-it-first advice stalls projects for three reasons: nobody has time to improve a process for its own sake, you cannot see the real process from inside it, and much of the mess turns out not to be worth fixing. Four things genuinely should be sorted first: duplicate records that will defeat matching, the absence of one named person who can make decisions, access nobody can authorize, and a process whose underlying software is being replaced within the year. Writing a process down precisely enough to automate is usually what forces it to settle.

## How many hours will automation actually save us?
https://theautomationdepartment.com/answers/how-many-hours-will-automation-save/

Take runs per month multiplied by minutes per run to get gross hours, then multiply by the realistic straight-through rate to get net hours, then multiply by the fully loaded hourly cost. The second step is the one most proposals skip. A document process running 300 times a month at 9 minutes each is 45 gross hours; at a realistic 85% straight-through rate the automation removes about 38 of them and the rest move to a review queue. Realistic straight-through rates after tuning are 75 to 90% for document work, 90 to 98% for rule-based routing with structured data, and near 100% for a simple two-system handoff. Get the inputs right by counting volume from the system rather than estimating, timing the task rather than recalling it, including the interruption cost of work that arrives unpredictably, and using a fully loaded hourly cost of roughly $35 to $55 for a US administrator. Saved hours are rarely saved salary at small-business scale: automation more often removes the need for the next hire. For collections chasing, lead routing and invoice matching the second-order benefits, such as working capital, conversion and avoided duplicate payments, frequently exceed the value of the hours.

## What should not be automated?
https://theautomationdepartment.com/answers/what-should-not-be-automated/

Do not automate anything that runs rarely, anything still changing shape, anything requiring genuine judgment, or anything where being wrong is expensive and the automation cannot tell when it is wrong. Also do not automate a process that should simply be deleted, because automating waste makes it permanent and faster. The most reliable warning sign is that nobody can write down the rules: if the person who does the work cannot explain when they do X instead of Y, the process is not ready. Document it first, and automate it later if at all.