AI automation
How much does AI automation cost, and how do you measure payback honestly?
AI automation is bought in four layers: an audit that picks what to automate, a one-workflow pilot, a build if the pilot earns it, and monthly running costs. 2026 price guides start builds at about $1,500; the VITON13 audit starts at $73. Payback is one-off cost divided by monthly net benefit, against a baseline measured before the pilot.
How much does AI automation cost in 2026, from the audit to monthly running costs?
Price it as four separate purchases: an audit, a pilot, a build and monthly running costs. Many quotes blend them, which is one reason two proposals for the same idea can be several times apart.
| Source | Build (one-off) | Running cost |
|---|---|---|
| Parix, 5 Jun 2026 | Simple $1,500–$4,000; medium $7,000–$12,000; advanced $15,000+ | $50–$400 a month, or 5–15% of setup a year |
| Syntora, 13 Feb 2026 | DIY $0; hybrid $5,000–$15,000; custom code $10,000–$40,000 | Tools $50–$500 (DIY) or $200–$800 (hybrid) a month; DIY adds 5–15 hours a month of your own upkeep |
| Codewave, updated 10 Aug 2026 | 1–3 workflows by a freelancer or small firm: $1,500–$8,000 | Maintenance 15–20% of implementation a year |
Why do price guides start at $1,500, and what can a smaller first step buy?
Those ranges price a working system: integrations, testing and launch. The step before it is often free: Parix offers a 30-minute scoping call ending in a fixed-price proposal, Syntora a discovery call. Such a call prices the build that vendor sells. If you have not chosen which process to automate, or need that choice in writing for any supplier to quote from, the first money should buy a decision, not software.
| Step | What it answers | Starting price | Typical timeline |
|---|---|---|---|
| AI automation audit | Which process to automate first, what data it needs, where a person must approve | $73 AI-assisted or $93 human-led | 5–8 working days |
| One-workflow pilot | Does one automated bottleneck hold up on real data, with approvals and a fallback | $73 AI-assisted or $93 human-led | 5–8 working days |
| Knowledge assistant | Can staff or clients get cited answers from your approved documents | $113 AI-assisted or $173 human-led | 10–18 working days |
| Product build | Several features, roles or systems, once the pilot has earned it | $213 AI-assisted or $313 human-led | 10–15 working days |
An agent wired into five systems is not a lower-rung job: it is quoted after the audit, as a product build.
How do you measure payback without guessing?
Hours saved times an hourly rate is where many estimates stop, and on its own it overstates payback. Use five inputs, written down before the pilot starts.
- Capture a 2–4 week baselineLog weekly volume, minutes per item, error rate and time to first response from the inbox, CRM or spreadsheet, not memory; the pilot will be measured the same way.
- Use a loaded hourly rateIn June 2026 US private-industry employers paid $32.82 per hour worked in wages and salaries and $46.89 in total compensation, so benefits were 30.0% of the total (BLS). Without your own figure, multiply the wage by about 1.4.
- Price the errorsRework minutes, refunds, missed orders, penalties. Halving handling time while doubling errors can lose money.
- Value response time only if it convertsCount a faster reply as money only when it changes something measurable, such as orders from after-hours enquiries.
- Subtract what remainsDeduct monthly running costs and the review time that remains.
In one line: monthly net benefit = (baseline hours − remaining review hours) × loaded rate + (baseline errors − new errors) × cost per error − monthly running cost. Payback in months = one-off cost ÷ monthly net benefit. At zero or below there is no payback period, however cheap the build was.
What does a payback calculation look like, step by step?
| Input | 600 enquiries a month | 60 enquiries a month |
|---|---|---|
| Handling time today, 4 min each | 40 h | 4 h |
| Loaded rate (placeholder: $21 wage × 1.43) | $30 an hour | $30 an hour |
| Errors today, 3% at $25 each | 18 → $450 | 1.8 → $45 |
| Monthly cost today | $1,650 | $165 |
| Review after automation, 1 min each | 10 h → $300 | 1 h → $30 |
| Errors after, 1% at $25 each | 6 → $150 | 0.6 → $15 |
| Running cost: tokens and platform plan | $60 | $60 |
| Monthly cost after | $510 | $105 |
| Monthly net benefit | $1,140 | $60 |
| One-off cost: studio fees, setup, staff hours | $1,200 | $1,200 |
| Payback | about 1 month | 20 months |
Same automation, same costs. At 600 enquiries a month the placeholders pay back in about a month; at 60 they need 20 months, long enough for the process or model prices to change first. Volume and the running-cost floor matter more than the build price.
Then stress-test it: double the review time, since early pilots get checked closely, and raise running costs to match 50% more volume. If payback still holds under both, the pilot is worth testing.
When does AI automation not pay back, and when should you stop a pilot?
- Low volume: a task done a few times a week rarely covers its running cost.
- A changing process: every change to steps, forms or prices means a rebuild. Syntora's bar for even exploring automation is at least 5 hours a week saved and a process stable for 3 months.
- Judgment-heavy work: when every output needs a full re-check, review time eats the saving.
- Lopsided mistakes: when one wrong refund, price or legal answer costs more than a month of savings, keep the decision human and automate only the preparation.
- No owner: if nobody owns the process today, nobody will accept, monitor or fix the automation.
Write the kill rule before the pilot, from baseline numbers. For example: stop or redesign if, after four weeks, review time per item exceeds half the baseline, errors exceed the baseline, or the running cost passes a ceiling set in advance.
Stopping is normal: in June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. A rule agreed in advance makes that call early and cheap.
What should an AI automation audit deliver before you pay for a build?
Test any audit, ours included, with one question: could a second supplier quote the build from its output without starting over?
- A process and data inventory: steps, systems, owners and where the data can be exported.
- A risk-value matrix ranking opportunities by value, data readiness, risk and the review each needs.
- One real task traced from source data to the human decision, with the point where automation must stop written down.
- A pilot recommendation, or a written recommendation not to automate yet.
- Acceptance on representative data with at least one failure case, such as a duplicate record or an API that does not answer.
- Success measures and a kill rule tied to your baseline.
- A monthly running-cost estimate by provider.
The VITON13 audit covers the first five; ask any supplier, us included, for the last two before signing a build.
AI-assisted or human-led: what changes in price, speed and risk?
Every VITON13 rung comes in two modes with the same deliverables. AI-assisted: AI drafts the inventory, matrix and documents from your exports and interviews; a person directs the work and checks every conclusion. Human-led: a specialist runs the interviews and mapping, with AI helping on drafts. The audit takes 5–8 working days either way.
| AI-assisted | Human-led | |
|---|---|---|
| Starting price | $73 | $93 |
| Fits when | The process is documented and exports from the inbox, CRM or spreadsheets exist | Know-how sits in people's heads, several teams share the process, or the data is sensitive |
| Main risk | Gaps in your documents become gaps in the map | Costs more, and interviews take your team's time |
How VITON13 does it
We run a fixed-scope AI automation audit for small businesses deciding which process to automate first, if any. It ends in a written recommendation, not a system, so you can take it to any builder, us included. Deliverables, inputs, revisions and exclusions are confirmed in writing before work starts.
AI automation audit
- Price
- $73 AI-assisted or $93 human-led
- Timeline
- 5–8 working days
AI automation services
- Price
- $73 AI-assisted or $93 human-led
- Timeline
- 5–8 working days
AI-assisted: AI drafts inside defined steps, a person directs and checks every result. Human-led: a specialist does the work and AI assists.
What is included
- 1 process and data inventory for the workflows agreed in the written scope
- 1 risk-value priority matrix ranking the automation opportunities
- 1 real task followed from your records to the person who decides, marking where automation stops
- 1 pilot recommendation, or a written recommendation not to automate yet
- Acceptance on representative data with at least one failure case
- 1 written completion note
- Delivery window: 5–8 working days once the brief, materials and access are in
Not included
- Building or integrating the automation, quoted separately as a pilot
- Model, automation-platform and hosting fees, paid by you to each provider
- Data cleanup and migration
- Legal or compliance review of the data you process
- Monitoring after handoff
Revisions: The number of review rounds is confirmed in the written scope
Get an estimate in 1 working dayExample
INTERNALOur own product
VIT LIVE: an assistant that answers from a catalogue before it calls a model
In the signed-in VITON ID cabinet, VITON13's own assistant answers from a catalogue of 84 items per language (32 products, 52 studio packages) in 5 languages, with at most 6 cards per answer. Catalogue matches and fixed replies need no model call. Only the remaining questions reach a self-hosted model, with replies capped at 180 tokens by default and a limit of 8 messages per 5 minutes per member. For running costs, that order is the point: the model sees only what cheaper steps cannot answer.
What this does not prove
- It is our own product, not client work, and it has no usage, conversion or cost figures, so it says nothing about payback.
- The case does not show the model running in production or judge the quality of its answers.
- Matching is by keywords, not meaning, and the word lists have gaps between languages.
Read the case: A VITON ID assistant that answers from its own catalogue first
Questions and answers
Are there monthly fees after an AI automation is built?
Usually yes: tokens, a platform plan, hosting for anything self-hosted, and someone's time to check failed runs. Before signing, ask for a monthly estimate by provider at today's volume and at double it. Optional monthly support from VITON13 is quoted separately from the build.
Who owns the workflow, the prompts and the accounts?
You should. At handoff, ask for the workflow export or code, the prompts, test cases and a runbook, so another supplier or your team can maintain it. If a supplier keeps everything in its own accounts, price the move out: run history may be lost and every connection set up again.
What happens if a model or API price changes?
Your running cost changes, not the build fee. Prices move both ways: Anthropic cancelled a rise for Claude Sonnet 5 scheduled for 1 September 2026. Set a spending limit where the provider offers one, keep the model swappable, and rerun the payback sum whenever a price moves.
Will VITON13 promise a saving or a payback period?
No. Payback depends on your volumes, error costs and running costs, which we do not control. We commit in writing to the scope, deliverables, timeline and acceptance test. The audit helps you check, with your numbers, whether a pilot is worth paying for; sometimes the honest answer is no.
What if the audit says not to automate?
Then you spent the audit fee rather than a build budget, and you keep the inventory and the ranked opportunities. The recommendation says what would have to change, such as volume or data, for automation to make sense later.
What do you need from me to start the audit?
A short brief, one normal example of the task and one that went wrong, the tools you use, any access limits, and the person who will accept the result. Exports from the inbox, CRM or spreadsheets help most, because the inventory is built from real records.
Sources
- Parix — How much does AI automation cost —
- Syntora — What AI automation actually costs a small business —
- Codewave — AI automation pricing: small businesses vs enterprises —
- Anthropic — Claude API pricing —
- OpenAI — API pricing —
- Zapier — Plans and pricing —
- n8n — Plans and pricing —
- U.S. Bureau of Labor Statistics — Employer Costs for Employee Compensation, June 2026 —
- Gartner — Over 40% of agentic AI projects will be canceled by end of 2027 —