What determines the cost of automating a process with AI?
Every serious quote is built from the same seven inputs. If a proposal does not tell you where it lands on each one, it is a number without a derivation.
| Variable | Pushes cost up | Pushes cost down | Hits |
|---|---|---|---|
| Process complexity | Branching rules, judgement calls, several teams involved | One decision, one owner, rules stateable in ten minutes | Build |
| Integrations | Many systems, no documented API, write access needed in both directions | One or two systems with clean APIs, read-heavy | Build and run |
| Output volume | High run counts, long documents, large context on every call | Low frequency, short outputs, retrieval instead of full documents | Run |
| Source data quality | Scattered records, systems that disagree, manual corrections before anyone trusts it | One system of record and 12 months of usable history | Build |
| Compliance requirements | Regulated sector, residency limits, audit trails, formal approval gates | Internal process with no personal or financial data | Build and run |
| Inference | Frequent calls to the largest models, retries, whole documents in context | Small models on routine cases, tight prompts, cached context | Run |
| Maintenance and iteration | Rules that change often, systems being replaced, growing scope | Stable process, one owner monitoring the exception queue | Run |
Two of these deserve the most attention in a first conversation. Integration count is the best single predictor of build effort. Source data quality is the best predictor of whether the timeline holds.
Why do build cost and running cost behave differently?
They are two different budgets and mixing them is the most common budgeting error.
Build cost is one-time and front-loaded. Process mapping, integration work, guardrails, testing against historical cases, staging review. It is driven by complexity and data quality. It does not repeat.
Running cost is permanent and scales with use. It has three parts: inference, infrastructure, and human attention. Inference is billed per token on input and output, so a design that sends an entire document on every call costs a multiple of one that retrieves the three paragraphs that matter. Human attention is the part nobody budgets: someone reads the exception queue every week, forever.
A cheap build with an expensive design costs more in year two than a careful build with a lean one. Ask for both numbers before you compare proposals.
How do the contracting models compare?
Three models dominate this market. They allocate risk differently, and that is the real difference between them.
| Model | How you pay | Who carries the risk | Fits | Failure mode |
|---|---|---|---|---|
| Hourly / time and materials | Per hour or per sprint, open-ended | You do | Genuine research where the answer is unknown | Discovery expands, the invoice grows, nothing ships |
| Fixed scope | A defined deliverable at an agreed range, approved before build | The vendor does | A process that can be specified before work starts | Anything outside the written scope becomes a new scope |
| Subscription | Recurring fee per agent, seat or volume tier | Shared | Standard processes served by preconfigured agents | You pay indefinitely for something you never own |
Black Ops uses fixed scope. The reason is not pricing philosophy, it is incentive alignment: when the deliverable is fixed, the vendor absorbs the cost of its own estimation errors, which is the only arrangement that makes an accurate estimate worth producing. The first conversation produces a written scope and an investment range within 48 hours, and build begins after your approval, not before. The recurring side is handled separately, as a retainer covering monitoring and iteration — described on the Black Ops services page.
How do you calculate the cost of not automating?
This is the number most budgets are missing, and it is arithmetic, not estimation.
- Direct labour. Hours per week × loaded hourly cost × 52. Loaded cost, not salary — include employer contributions and overhead.
- Seniority premium. Count whose hours these are. An hour of a Head of Operations is worth several times an hour of clerical work at the same frequency.
- Error cost. Errors per year × average cost to fix one. Duplicated payments, missed renewals, rejected tax documents, credits issued to keep a customer.
- Delay cost. What a slower cycle costs in revenue: quotes sent two days late, leads answered after a competitor, receivables collected a month later.
- Opportunity cost. What the same people would produce if the process ran without them.
Lines 1 to 3 are defensible in a board meeting. Lines 4 and 5 are real but harder to prove, so present them separately rather than folded into one headline figure.
You can run line 1 yourself: multiply the hours per week by the loaded hourly rate and by 52, and you have the annual figure. Use your own numbers, not an industry average. Compare that annual number against a build cost that happens once, and the payback period answers itself.
What makes a quote move after it is signed?
Fixed scope only stays fixed if the scope was honest. These are the five things that change it, and all five are findable before you sign.
- An undocumented step. The real workflow contains an approval nobody mentioned. Map the process as it runs, exceptions included.
- A system with no API. Discovered in week two, it converts a real-time design into a batch design, or into custom connector work.
- Data that needs consolidating first. This is a separate project. It should be named in the scope, not absorbed silently.
- A compliance requirement that appears late. Residency, retention, audit trails and approval gates are cheap to design in and expensive to retrofit.
- Scope growth. One process becomes four. Each addition is a new scope with its own range, or the timeline stops meaning anything.
How do you compare two proposals fairly?
Ask both vendors the same eight questions and put the answers side by side. Differences in the answers explain differences in the number better than the number does.
- Which systems will you integrate with, and does each one need read access or write access?
- What is explicitly out of scope?
- What is the estimated running cost in month twelve, separate from the build?
- What happens to the number if our data needs consolidating first?
- Who owns the code and the configuration when the engagement ends?
- What is the metric, what is its baseline today, and who measures it?
- What does the agent do when it is not confident?
- What does support cover after go-live, and for how long?
A proposal that answers all eight in writing is comparable. One that answers three is a brochure. For the terms used in these answers, the Black Ops glossary of AI automation terms defines each one.
Common questions about the cost of AI automation
Why does Black Ops not publish a price list?
Because two processes that sound identical in a meeting can differ by an order of magnitude in integration count, data quality and compliance requirements. A published number would be wrong in both directions. Black Ops (blackopss.com) delivers a written scope with an investment range within 48 hours of the first conversation, based on your actual process.
Is it cheaper to build the agent in-house?
The build can be. The maintenance rarely is, because the cost lands on the people you already have and never appears as a line item. Compare honestly: engineer time at full loaded cost, the projects that slip while they build, and who owns the exception queue in month nine when the person who wrote it has moved teams.
What is the single biggest cost driver in an AI automation project?
Source data quality, by a wide margin. Clean structured records and clear rules make the build predictable. Records scattered across systems that disagree with each other turn the first deployment into a consolidation project, which is a different piece of work with a different timeline and should be named as such in the scope.
Does the cost of running an AI agent go down over time?
Cost per run tends to fall as prompts get tighter, retrieval replaces sending whole documents, and cheaper models take over the routine cases. Total spend often rises anyway, because a working agent gets pointed at more volume. That is the outcome you want, but budget for it rather than being surprised by it.
What should be in a written scope before we approve the budget?
Six things: the process mapped as it actually runs, the named systems and the access each one needs, what the agent will explicitly not do, the metric and its baseline before launch, the deployment timeline, and the investment range with what would move it. If any of the six is missing, the number is a guess.
Where to go next
If you are still deciding what an agent is and whether your process qualifies, read what an AI agent for business is and how it differs from a chatbot. If the technology choice is still open, the comparison of agents, RPA and no-code automation covers cost per run for each approach. If the operation runs in Latin America, labour cost changes which processes pay back first. The Black Ops FAQ covers the commercial terms directly.