AI automation - cost & ROI
How much does AI automation cost?
Enterprise AI automation is priced by four variables: integration count, data condition, approval requirements and regulatory load. A multi-agent platform across several departments costs a multiple of one well-scoped agent, and it keeps costing every year it runs.
Enterprise buyers ask three questions before anything else: what does one agent cost, what does a full program cost, and when does it pay back. The first two depend on the variables below. The third is arithmetic you can run yourself.
What actually drives the price
Two systems that look identical in a demo can cost four times apart to build. The gap comes from four variables, and the model itself is rarely one of them, because API pricing from Anthropic or OpenAI is a small fraction of the engineering time next to it.
- Integration count: an agent that reads one system and replies in a fixed format is a few weeks of work. One that reads a CRM, checks stock in an ERP, applies a discount rule, and writes back a confirmation is a different project.
- Data condition: retrieval over clean, structured documents builds fast. Retrieval over a decade of scanned PDFs, inconsistent naming, and three versions of the same policy needs real cleanup time before the AI part starts.
- Approval requirements: a fully automated workflow costs less to build than one with a human checkpoint on every output, because the checkpoint has to be designed, logged, and tested as a feature in its own right.
- Regulatory load: a claims or credit workflow in banking or insurance needs audit trails, model version pinning, and a documented human override. A marketing workflow needs none of that.
What each type of build involves
Build cost follows engineering weeks, and running cost follows volume. The timeline is the part you can compare across proposals.
| Automation type | Typical timeline | What the running cost follows |
|---|---|---|
| Single-task agent, one workflow, one system | 3 to 5 weeks | Decision volume on that one workflow |
| Multi-step agent, several systems, one approval gate | 6 to 10 weeks | Volume, plus the calls each extra system adds |
| RAG knowledge system over your own documents | 6 to 12 weeks | Query volume and the size of the document set |
| Multi-agent platform across several departments | 3 to 6 months | Several workflows sharing one infrastructure bill |
The monthly figure is not overhead to cut once the build ships. It covers model calls, hosting, monitoring, and the fixes a production system needs after real users hit the edge cases a demo never saw.
Build once, or run a retainer
Two commercial models exist, and agencies rarely explain which one they are selling. A fixed-price build hands over the system and the keys, and the client owns the model bill, the monitoring, and the next bug. A retainer keeps the agency operating it, usually priced on top of the model and hosting cost rather than instead of it.
Regulated sectors lean toward the retainer almost every time. A bank or insurer building AI for financial services carries audit obligations that do not end at go-live, so somebody has to own the logs, the model version history, and the override record for as long as the system runs. That ownership is the retainer, whatever the invoice calls it.
The agencies still standing after year two are usually the ones built to operate as an enterprise AI agency rather than a project shop that moves on after delivery. Year two is where the real cost shows up: a model gets deprecated, an API changes its schema, or volume triples and the architecture that worked at pilot scale needs rework.
The ROI math that holds up in a board meeting
Vague productivity language does not survive a finance review. What does: the number of decisions the system takes off a person's desk each month, multiplied by the fully loaded cost of the time that decision used to take, minus the running cost.
Parci, the real estate platform Digiton built and operates, prices its automated zoning and valuation checks at €29, €149, and €499 a month across three tiers, because the cost of running one query dropped once retrieval replaced a manual desk search that used to take an analyst most of an afternoon. That is the unit economics question worth putting to any vendor: what does one transaction cost to run once the system is live, separate from the number on the build invoice.
A WhatsApp AI assistant Digiton runs for a client answers customer questions in real time at a volume that would otherwise need two additional support staff. A closed-RAG university tutor built for another client answers student questions only from the institution's own course material, because a general model guessing at a specific syllabus produces confident, wrong answers. Neither project impresses on novelty. Both pay for themselves in hours a person no longer spends on repetitive replies.
What separates a real quote from an inflated one
Four signs a proposal is priced honestly rather than padded:
- It prices per workflow or per outcome, not per seat. Per-seat pricing on an AI product is a signal the vendor is repackaging software economics onto something with a genuinely variable cost. Model tokens scale with usage the way a software license never did.
- It names the model and discloses that model cost passes through separately from the build fee.
- It includes a monitoring and maintenance line, because a system with no maintenance plan degrades quietly until somebody notices a wrong answer went out for three weeks.
- It scopes one workflow first rather than an entire department, so cost and payback can both be measured before the next one gets funded.
A RAG system for business quoted as a flat annual license, with no disclosure of the underlying model or the retrieval architecture, is not a quote anyone can compare against another. Ask what happens when the document set doubles. That is where flat-fee proposals usually reveal their real margin.
Where to start
Digiton is an AI operations agency based in Lisbon, founded by Brandon Da Costa (background on the about page), working with enterprise teams across the UK, Ireland, the US, Canada, and Australia as well as Portugal. Start with a free 25-minute scoping call. If the system fits, the production review is the first paid step, and the build is quoted from its costed fix list against your actual workflows.
Frequently asked questions
How much does a single AI agent cost to build?
It depends on how many systems the agent reads and what it does on an exception, so treat a figure given before anyone has read your systems as a guess. Digiton quotes after a free 25-minute scoping call and a paid production review. A quote far below its peers is usually a demo dressed as a product, and one far above usually includes work that belongs in a separate project.
Is a retainer or a one-off build cheaper for AI automation?
A one-off build looks cheaper on the invoice, but the client then owns model costs, monitoring, and every fix a production system needs after real users hit it. A retainer folds that ownership into a monthly fee. Regulated sectors such as banking and insurance almost always end up on a retainer once audit and logging duties are counted properly.
How long does AI automation take to pay for itself?
Most single-workflow agents pay back within three to six months once decision volume and hours saved are measured honestly. Multi-agent platforms take longer, often nine to eighteen months, because the payback depends on several workflows going live, each carrying its share of the shared infrastructure cost.
Why do AI automation quotes vary so much between agencies?
Because the same phrase covers work that ranges from a single scripted agent to a full multi-agent platform. Ask what system it reads, what happens on an exception, who owns the model bill, and whether monitoring is included. Two quotes that answer those four questions the same way will usually land in the same price range.
Related
Building something that has to hold up in production?
Start with a free 25-minute scoping call. If it fits, the first paid step is a production review: five working days on your real repository, scored against ten layers, with a costed fix list.
Book a 25-minute scoping call