AI automation services: scope, pricing, red flags

The category is crowded and the label means very little on its own. Two firms can both sell "AI automation services" and mean completely different work: one wires your intake form to your CRM and adds a model to classify the message, the other rebuilds how a claims team handles exceptions.

That ambiguity makes the buyer's job specific. You are not evaluating whether AI automation works. You are evaluating what one provider will build, what it costs to run, and who owns it when the engagement ends.

What is actually in the scope

Strip the positioning and a typical engagement contains five things. Only one is distinctively "AI".

If a proposal is heavy on the third row and thin on the first, fourth and fifth, you are being sold a demo. The demo will work. The system will not.

Which parts genuinely need a model

This is the question that separates a real scope from an expensive one, because model-based steps cost more to build, more to run, and considerably more to verify.

Use ordinary automation when the logic can be written as rules. Moving a record between systems, sending a reminder at a fixed interval, applying a routing rule, calculating a total. Deterministic, testable, cheap, and they do not surprise you at 2am. Anything expressible as "if this, then that" should be built that way — see workflow automation.

Use a model when the input is unstructured or the step requires reading. Classifying an inbound message by intent. Pulling fields from a document that arrives in forty different layouts. Drafting a reply a person will review. Summarising a call. Rules handle these badly and models handle them well.

Be cautious when the model is asked to act rather than interpret. An agent that reads a request and then executes a refund or updates a customer record is a different risk class from one that reads and suggests. That is where Gartner's three named causes of cancellation tend to bite. Autonomy is a dial, not a switch, and most production systems settle lower on it than the sales conversation implied.

How the work gets priced

Four structures dominate, and each one shifts risk somewhere different.

Two costs are routinely absent from proposals and always present in reality. Model usage is metered and scales with volume, so a system that costs little in pilot can cost meaningfully more at production traffic — get the per-unit cost in writing and model it at ten times pilot volume. Maintenance is not optional: APIs change, models get deprecated, formats drift. A build with no maintenance line decays from the day it ships.

DoubleTime publishes its prices — custom AI systems are typically $25,000 on a fixed bid — because quoted pricing rewards whoever guesses your budget best.

Questions that separate real from sold

Ask these before signing. The answers are more informative than any case study.

1. What have you built that is still running, and who operates it? Not a demo — a system in production with someone responsible for it.
2. Which steps call a model, and what is the fallback when output is low-confidence? If there is no fallback, there is no system.
3. What does this cost to run at ten times the pilot volume? Usage-metered costs are the most common budget surprise.
4. Who owns the code, prompts, integrations and credentials at the end? Get it in the contract. "We host it for you" and "you own it" are different arrangements.
5. What are you doing about my data quality? Automation over duplicated or stale records propagates the mess — see data enrichment for what that layer involves.
6. What would make you tell me not to do this? A provider who has never declined work has no standard.

When AI automation is the wrong purchase

Worth saying plainly, because it saves money:

• The process isn't stable. If how the work gets done changes every quarter, you'll rebuild the automation faster than it pays back. Map the process before anyone quotes on it.
• The volume is low. A task done four times a month rarely justifies a build, however annoying it is.
• The bottleneck is a decision, not a task. If work stalls waiting for an approval, faster processing either side of it changes nothing.
• Nobody internally will own it. Systems no one owns quietly stop being used — adoption is a project of its own.
• The process is undocumented and disputed. Automating something two teams describe differently makes the disagreement run faster.

Frequently asked questions

What is the difference between AI automation and regular workflow automation?

Regular workflow automation executes rules you define: when a form is submitted, create a record; when an invoice arrives, route it for approval. It is deterministic, meaning identical input produces identical output every time. AI automation adds steps a model performs — classifying an unstructured message, extracting fields from a document with an unpredictable layout, drafting text, summarising a conversation. Those steps are probabilistic, so they need confidence thresholds and a human review path. Most working systems combine both: rules for the flow, a model only where reading or judgement is genuinely required.

How much do AI automation services cost?

Pricing varies by scope and structure — fixed-bid projects, time and materials, monthly retainers, and per-outcome pricing are all common, and each shifts risk differently. Two costs are frequently missing from the proposal and always present later. Model usage is metered and grows with volume, so ask for the per-unit cost and calculate it at ten times pilot traffic. Maintenance is unavoidable, because APIs change and models are deprecated on the vendor's schedule rather than yours. DoubleTime publishes fixed prices for all four of its services rather than quoting per client.

Why do so many AI automation projects fail?

Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, attributing it to escalating costs, unclear business value and inadequate risk controls. In practice those show up as recognisable patterns: a pilot built on clean sample data that meets real inputs and degrades; a usage bill that scales past the savings; a system with no defined handling for low-confidence output that fails without anyone noticing; and internal ownership that was never assigned, so the system quietly stops being used. Each is preventable at the scoping stage and expensive afterwards.

What is "agent washing"?

Agent washing is the practice of rebranding existing products — AI assistants, robotic process automation, rule-based chatbots — as agentic AI without adding substantial agentic capability. Gartner uses the term explicitly, estimating that only about 130 of the thousands of vendors marketing agentic AI are genuinely delivering it. For a buyer, the defence is to test claims against behaviour rather than description: ask which decisions the system makes without a person, what it is permitted to execute, what happens when it is uncertain, and to see it running on data resembling yours rather than a curated demo.

Should I hire an agency or build in-house?

It depends on whether this is a one-time build or a continuing capability. A single system with a clear boundary — document processing, call handling, lead routing — is usually faster and cheaper to buy, because the provider has built it before and you are paying for pattern recognition rather than learning. A continuing programme where automation becomes part of how the company operates eventually wants internal ownership, because response time and institutional knowledge matter more than build speed. A common middle path is an agency build with an explicit handover: code, prompts, credentials and documentation transfer at the end, specified in the contract.

How long does an AI automation project take?

A single well-scoped workflow with systems that already have usable APIs is typically weeks rather than months, and most of that time goes to discovery and exception handling rather than the model work. Timelines stretch for identifiable reasons: legacy systems without APIs, data that must be cleaned before it can be used, processes that turn out to be undocumented or disputed between teams, and approval cycles on the client side. Beware of anyone quoting a timeline before watching the process being performed — the estimate is describing the happy path, and the exceptions are where the hours live.